@slatesvideo/shared 0.6.5 → 0.6.6

package/dist/skills/content.js

@@ -3,37 +3,37 @@
 export const SKILLS = {
     "slates-blocking-to-prompt": "---\nname: slates-blocking-to-prompt\ndescription: Write the generation prompt that matches a blocking clip second by second, so the reference video and the text agree instead of fighting. Use after rendering a previs blocking pass, when a generated shot ignores the reference video, when timings drift, or when the model invents shots and camera angles that are not in the blocking.\n---\n\n# Blocking → prompt\n\nYou have a blocking clip. This is how you write the prompt that goes with it.\n\n## The one idea\n\nThe clip already contains the camera, the cuts and the timing. **The prompt's job is to say what everything looks like — and, where the grey boxes are ambiguous, to disambiguate them.** It is not a second, competing description of the motion.\n\nState that contract inside the prompt, because the model needs it as much as you do:\n\n> Where a timeline line below names a camera position or move, it is a restatement of what the reference video already does at that timestamp — a disambiguation, never a new instruction.\n\nAnd give it a tie-break, because ambiguity is guaranteed:\n\n> If any text in this prompt appears to disagree with the reference video about camera, framing, direction, motion, timing or object placement, the reference video wins.\n\nThose two sentences do more work than any other part of the prompt.\n\n## Get the real numbers first\n\n```\nslates_blender_scene\n```\n\nWrite timings from `cutSeconds`, never from the shot list you intended to build. It resolves to the marker frames on a multi-camera edit and to the camera's own keyframes otherwise, so it is the one field that is never empty on a rig that has cuts. At 24fps cuts land on frame boundaries and the honest values are not round — `7.79s`, `9.33s`, `19.875s`. **Use the exact ones.** Rounding to `7.8` is a tenth of drift you are handing the model for free.\n\n## Structure\n\nOrder matters — contract, then globals, then timeline, then the re-assertion.\n\n```\nTITLE — one line: what this is, how long, that it is video-to-video\n\nLOGLINE — 2-4 sentences. The whole piece in plain language.\n\nACTIVE REFERENCES\n  <one entry per reference: what it defines, and what is NOT inherited>\n\nTECHNICAL BLOCK    (format, grade, lens, and the blanket negatives — see below)\n\nSTYLE / LOOK\nLIGHTING\nCOLOR\nCAMERA\nPHYSICS            (only if things move, collide or deform)\n\nRULES              (numbered — the invariants, see below)\n\nACTION TIMING      (beat by beat, against the clip's real timestamps)\n\nAUDIO\n  [Sound design]   [Timed accents]   [Dialogue]   [Music]\n\nENDING LOCK        (one line — where the film stops)\n\nHOLD FOR THE FULL TIMELINE\n  <the 5-6 constraints most likely to drift, compressed>\n```\n\n## ACTIVE REFERENCES — every entry has an exclusion\n\nThe single highest-leverage format in this whole workflow. Each reference is a **positive claim plus an exclusion list**, because a reference the model over-reads is as damaging as one it ignores.\n\nLabel each by the badge code Slates echoes back (`IMG-A8`, `VID-C2`) or by an unmistakable role name, and use that same label everywhere below.\n\n**The blocking clip:**\n\n> VID-C2 = the blocking previz (30s, 720 frames, 24fps) — the MASTER for everything that moves and everything that stands. It defines the full edit one-to-one: every cut point, every camera position, angle, move and framing, all action timing, screen direction, and the geometry of the world. Its untextured grey surfaces, flat colours and viewport grid are NOT inherited — every grey proxy is dressed into a real object in the exact position the previz puts it. Proxies give position, angle, scale and motion only; never surface, shape detail or design.\n\nThat last sentence is the **placement-only clause** and it is not optional. Without it the model renders grey boxes.\n\n**A character sheet:**\n\n> IMG-A8 = the driver — defines his face, hair and wardrobe. Identity 100% consistent at every distance and through every motion blur. Background, lighting and pose are NOT inherited.\n\n**A location/style reference:**\n\n> IMG-B3 = the tunnel — defines location geometry, look and grade. Camera angle, framing and any people in it are NOT inherited; the camera comes exclusively from VID-C2.\n\n**An atmosphere or style master — a reference that is never a shot:**\n\n> IMG-D9 = ATMOSPHERE MASTER — NOT a keyframe, NOT a location to reproduce, NOT a frame that ever appears in the film: its own subject, framing and composition are never seen in any shot. It defines ONLY the weather, light, colour and grade: deep clean night just after rain, wet asphalt as a dark mirror, cool white-cyan lamps as the ambient key, teal-and-amber grade, deep clean blacks. Every shot is lit and graded in this regime for all 30 seconds.\n\nWithout those three NOTs the model reproduces the reference's composition as an actual shot — you get its street corner in your film. The same wording covers a rendering-style master; see `slates-restyle-from-blocking`.\n\n**References can be scheduled.** If something is only true for part of the timeline, say so: *the hooded panel applies only to 0–7.0s and 27.5–30s*, or per-reference: *Active for 00:03.3–00:06.7 only.* On a piece that travels through several locations, every location still carries its own window and the model stops blending two sets into one shot.\n\n## Translate the blocking's artifacts\n\nYour grey-box render contains things that are *notation*, not content. Every one needs an explicit reinterpretation or it gets rendered literally:\n\n| In the blocking | Say in the prompt |\n|---|---|\n| Colour-coded bodies | `red = the boss, green = the kid, blue = the driver` |\n| A marked face on a proxy | `RED face = the direction he faces, BLACK = his back` |\n| A checkered floor or wall | `the checkerboard is a scale reference, not a surface — it becomes <the real material>` |\n| Flat black background | `a PLACEHOLDER — replace with the location assigned below` |\n| A floor grid | `a motion-tracking aid — render as light on the surface, never as wireframe or tiles` |\n| A deliberate black gap | `CUT 7 (14.5-17.0, black gap in the reference) — <what fills it>` |\n| Frame goes dark mid-move | `the camera is passing through the ground — a doorway to the NEXT location, never back to a previous one` |\n| The source hard-resets mid-move | `each reset begins a NEW, completely different room — never a replay of one already seen` |\n| A proxy that is a PROP or VEHICLE | `the low-poly flying model in SHOT 18 is THE HELICOPTER · blocks on the rear bench are the luggage · the small dark block in his hand IS the pistol` |\n| A blocky proxy limb in a tight insert | `the blocky low-poly leg is a stand-in and must NOT be replicated — generate complete human anatomy: a real boot, a real trouser leg, a correct ankle at this exact camera angle` |\n| A stray object at the frame edge | `ignore it completely — never blend two sets into one shot` |\n\n## TECHNICAL BLOCK — format, lens, and the blanket negatives\n\nOne paragraph, before the timeline. It carries the things that are true of every frame and that no beat should have to repeat:\n\n> Cinematic, photoreal. 21:9. 30s. SFX only, no music. Kodak 500T film look, natural 35mm grain, organic colour, soft highlight roll-off, anamorphic lens character with oval bokeh and gentle barrel distortion at the edges, chromatic aberration creeping in at the frame edges, natural motion blur on every fast move, faint bloom on hot speculars. Every location well exposed — night interiors bright and readable, open shadows, no crushed blacks, no murk. NO CGI. NON-IP, no brand badges or logos anywhere, no text, no watermark.\n\nThree parts worth naming:\n\n- **Lens realism is a list, not an adjective.** Aberration, motion blur, depth of field, barrel distortion, bloom, grain. \"Cinematic\" buys you nothing; these buy you the look.\n- **Exposure needs saying on dark work.** Models crush night scenes into murk. *Night interiors bright and readable, open shadows, no crushed blacks* is what keeps a scene legible.\n- **The blanket negatives go here once** — `NON-IP`, no logos, no on-screen text, no subtitles, no watermark — rather than being scattered through the beats.\n\n## RULES — the invariants\n\nNumbered, short, absolute. These are the things that must hold in every frame, and they are where you put anything that has already gone wrong once.\n\nTwo patterns worth stealing outright:\n\n**Countable state.** Give the model arithmetic it can check itself against:\n\n> At every second: standing + fallen + on the lintel = 6. Never a seventh figure — no extras, no duplicates, no distant silhouettes, no half-bodies at frame edges.\n\n> Bodies on the ground count exactly: 0 before 12s → 1 → 2 → 3 → 4 at 12/13/14/16s → 5 at 20s → 6 at 26s. Never more.\n\n**Every mass is dressed, and nothing is invented.** The blocking is authority over what EXISTS, not just what moves — otherwise the model deletes the masses it finds boring and adds architecture you never blocked:\n\n> Every lamppost, guardrail, road and terrain mass visible in the reference exists in the output in the same place, at the same scale, in the same position in frame — the opening blocks are dark-brick warehouse facades, the roadside masses are the waterfront skyline, the finale rocks are the city's tower walls. Nothing is deleted, and no structure is invented where the reference shows none.\n\n**Anatomy is never inherited from a proxy.** Tight inserts on hands and feet are where blocking leaks straight into the render:\n\n> The reference shows only WHERE hands and feet are. In the output they are always complete human anatomy — a five-fingered gloved hand with natural knuckles, a real leg in wool trousers, a real foot in a leather shoe — never the proxy's blocky shape.\n\n**A ledger for anything that happens a countable number of times.** A ritual, a reload, a set of falls: state it as a linear sequence, each step exactly once, and close with the tally:\n\n> Strictly linear, six steps in fixed order, each happening EXACTLY ONCE and never repeating; once a step is done it is done for good, and the sequence only ever moves FORWARD, never backward. Count of weapon events in the entire video: one draw, one magazine insertion, one slide rack, one shot.\n\nWithout the ledger the model loops the most cinematic beat — it will rack the slide four times because racking looks good.\n\n**Named misreads.** When a generation gets something specifically wrong, do not rewrite the description — **name the wrong reading and kill it**:\n\n> The lamp is a man-made steel structure — NOT an animal, NOT a snake, NOT any living or organic shape.\n\n> The rear of the car and its tail lights are NOT visible in this shot.\n\nThis is the highest-value edit available after a failed roll, and it is why the prompt grows rather than changes between takes.\n\n## ACTION TIMING — the beats\n\nOne block per shot or beat. Two notations; pick one and hold it.\n\n**For a continuous take**, ranges with a camera note and a closing state audit:\n\n```\n8-12s — THE SWEEP (per VID-C2: elevated rear push, swinging to profile by 12s):\n<what happens, in prose, with sub-beats on tenths and → chaining cause to effect>\nEND 12s: bodies 1 (behind him as he steps past) · standing — four ahead, holding.\n```\n\n**For a cut edit**, numbered shots ending on their cut:\n\n```\n9.33-10.33s — SHOT 10 — Interior over the centre console as in VID-C2: <what the\nframe contains>. Hard cut at 10.33s.\n```\n\nThree habits that separate a beat that works from one that does not:\n\n- **Declare the frame's contents as a closed set** when the shot is tight: *the frame holds exactly the console, the lever, his hand, and the edges of both seats.* An open description invites additions.\n- **Chain cause to effect inside one sentence** with `→`. `he overcommits a lunge → the Hero drops low and sweeps his standing leg → he hits the earth at 12s`.\n- **Put events on tenths.** `11.7s`, `19.5s`, `22.5s`. Vague beats generate vague timing.\n\nDensity: roughly 60–130 words per second of screen time is what these prompts actually run at. That is much denser than a normal video prompt, and it is the point.\n\n## AUDIO\n\n`[Timed accents]` uses the same timestamps as the beats:\n\n> 3.1s tyres light up into the burnout squeal · 7.0s drift-entry screech · 10.1s hard mechanical shifter clack · 20.3s full-speed pass-by whoosh\n\n`[Dialogue]` is a closed list — count the lines, give each a window, quote it verbatim, and forbid everything else:\n\n> Exactly TWO vocal events in the entire 30 seconds, both screamed, in English, VERBATIM: 1. 17.3-18.6s \"STOOOOOP!!\" 2. 23.4-24.0s \"You crazy!\" Nothing else is ever spoken.\n\n**Then forbid the lines it will invent anyway.** A closed list is a rule; an enumerated blacklist is enforcement, and the phrases to list are the clichés the scene invites:\n\n> FORBIDDEN — she never says any of these and no one else says anything: \"they're behind us\", \"cops\", \"go go go\", \"are you crazy\", \"you're insane\", or ANY other invented phrase. All other human voice is wordless screaming or laughing.\n\n**When lines are lip-synced, give each one a timestamp** in the same list, and say that they change nothing else:\n\n> Timing: \"You wind up for this one?\" ~18.2s · \"Three full turns.\" ~19.0s · \"Company.\" ~24.3s. Every line lip-synced; the lines never change the camera.\n\nTwo rules that stop dialogue from breaking the edit:\n\n> DIALOGUE NEVER CREATES SHOTS: spoken lines happen inside the reference's takes exactly as blocked — no cutaways to a speaker, no reverse shots, no added close-ups. If a line plays while the camera is elsewhere, the line stays off-screen audio.\n\n> A line marked off-screen must STAY off-screen — never show the speaker, never move him into frame, never route the camera to him because he spoke.\n\nModel note: dialogue direction as separate layers is minimax-h3's seat; native synced audio is Veo's. Route per `slates-model-selection` and read the model's own prompting skill before writing the audio block.\n\n## ENDING LOCK\n\nOne line, and it is the cheapest fix in the document. Models drift at the end — they hold a frame too long, add a beat after the last one, or fade somewhere the reference does not:\n\n> The film ends exactly where the reference ends: the final take runs unbroken to its last frame, and that source frame IS the final frame of the film. Nothing follows. The last five seconds follow the source exactly as strictly as the first five.\n\n## HOLD FOR THE FULL TIMELINE\n\nClose with a terminal re-assertion of only the constraints most prone to drift — five or six lines, compressed, no new information:\n\n```\nHOLD FOR THE FULL TIMELINE\n- VID-C2 camera path 1:1 — any deviation = failure.\n- Six and only six figures; the count above holds at every second.\n- IMG-A8 identity constant at every distance and through motion blur.\n- The IMG-B3 location in every frame; no subtitles, no watermarks.\n```\n\nRestating is not redundancy here. It is the last thing the model reads.\n\n## Checklist before you generate\n\n- [ ] Timings taken from `slates_blender_scene`'s `cutSeconds`, frame-exact, not rounded\n- [ ] Every reference has an explicit \"NOT inherited\"\n- [ ] The placement-only clause is present\n- [ ] The tie-break clause is present\n- [ ] The disambiguation clause is present\n- [ ] Every blocking artifact is translated — colours, marked faces, checkers, grid, black background, gaps, dark dips, resets, prop proxies, proxy limbs, strays\n- [ ] Any style/weather reference is declared NOT a keyframe and never a shot\n- [ ] The every-mass-is-dressed / nothing-invented rule is present\n- [ ] Tight inserts on hands or feet demand complete anatomy\n- [ ] Counts are stated where anything is countable, and repeatable actions carry a ledger\n- [ ] A TECHNICAL BLOCK carries format, lens realism, exposure and the blanket negatives\n- [ ] `[Dialogue]` is a closed list with a FORBIDDEN blacklist\n- [ ] An ENDING LOCK says where the film stops\n- [ ] `videoReferenceSecondsEach` matches the clip's real duration\n- [ ] A HOLD block closes it\n\n## Related\n\n`slates-previs-blocking` (producing the clip) · `slates-camera-language` (the moves being described) · `slates-dialogue-blocking` (multi-character continuity) · `slates-restyle-from-blocking` (reusing this prompt across styles) · `slates-prompting-seedance-2-5` / `slates-prompting-minimax-h3` (model-specific rules)\n",
     "slates-camera-language": "---\nname: slates-camera-language\ndescription: Turn director vocabulary into real Blender camera rigs — orbits, floor rises, robo-arm whips, handheld, speed ramps, over-the-shoulder cuts — as bpy code. Use when building or refining the camera on a previs blocking pass, when a move needs to accelerate/hold/snap, or when someone asks for a \"cinematic\" camera and you need to convert that into an actual shot list.\n---\n\n# Camera language — from a shot list to a rig\n\nCompanion to `slates-previs-blocking`. That skill owns the workflow; this one owns the camera.\n\n## The first rule\n\n🚨 **Never build \"a cinematic camera move.\" Brief the camera the way you would brief an operator:** rails, target, height, lens, and the frame each move starts and ends on. \"Cinematic\" is not a specification, and asking for one produces the drifting slop the whole blocking workflow exists to avoid.\n\nBad: *a dynamic cinematic orbit around the subject.*\nGood: *a 3/4 orbit starting rear-left at 1.6m, ending front-right at 0.9m, frames 1–96, 35mm, easing out of the start and holding hard on the last 8 frames.*\n\nIf the user gives you the first, convert it to the second and say what you assumed.\n\n## Look it up, don't recall it\n\nBefore any constraint or operator you are not certain of, call `slates_blender_docs` (e.g. `bpy.types.FollowPathConstraint`) or `slates_blender_search_docs`. Invented enum values are the most common failure here and they often fail *quietly* — the constraint gets added, the axis is wrong, and the camera points at nothing.\n\n## The two primitives everything is built from\n\n### Target-based aiming\n\nAlmost every move in this skill is *position on a path* plus *aim at a target*. Separating them is what lets framing vary while the subject stays in frame.\n\n```python\nimport bpy\n\ntarget = bpy.data.objects.new(\"CAM_TARGET\", None)   # an Empty\ntarget.empty_display_type = 'PLAIN_AXES'\nbpy.context.collection.objects.link(target)\n\ncon = cam.constraints.new('TRACK_TO')\ncon.target = target\ncon.track_axis = 'TRACK_NEGATIVE_Z'   # cameras look down -Z\ncon.up_axis = 'UP_Y'\n```\n\n**Aim at a separate target, not at the subject's head.** A camera locked to the head produces dead, centred framing. Offset the target beside or ahead of the subject and the shot breathes — that small offset is most of what reads as \"real operator.\"\n\nFor a subject that should notice the camera, keyframe the target's follow with a **few frames of lag** behind each camera move. Heads catch up; they don't teleport.\n\n### Path-based movement\n\n```python\ncurve = bpy.data.curves.new(\"CAM_PATH\", 'CURVE')\ncurve.dimensions = '3D'\nspline = curve.splines.new('BEZIER')\nspline.bezier_points.add(len(points) - 1)\nfor bp, co in zip(spline.bezier_points, points):\n    bp.co = co\n    bp.handle_left_type = bp.handle_right_type = 'AUTO'\n\npath = bpy.data.objects.new(\"CAM_PATH\", curve)\nbpy.context.collection.objects.link(path)\n\ncon = cam.constraints.new('FOLLOW_PATH')\ncon.target = path\ncon.use_curve_follow = False   # aiming is the Track To constraint's job\n\n# Animate progress explicitly rather than relying on the default path animation.\ncurve.path_duration = 96\ncurve.eval_time = 0\ncurve.keyframe_insert(\"eval_time\", frame=1)\ncurve.eval_time = 96\ncurve.keyframe_insert(\"eval_time\", frame=96)\n```\n\n**Why a path and not raw location keys:** the user can drag a control point to retime or reshape the move without you regenerating anything. That is the difference between \"re-prompt and hope\" and \"nudge it.\"\n\n## Speed — the part that reads as production value\n\nMovement at one constant speed is the tell of a machine. Real moves accelerate, hold, and snap.\n\nSpeed lives in the **f-curve handles** of `eval_time` (or of location, if you keyed it directly):\n\n- **Long, near-horizontal handle** at a key → slow near that key.\n- **Short, steep handle** → fast.\n- `interpolation = 'CONSTANT'` → no movement at all until the next key. This is how you get an absolute dead stop.\n\n```python\nfc = curve.animation_data.action.fcurves.find(\"eval_time\")\nfor kp in fc.keyframe_points:\n    kp.interpolation = 'BEZIER'\n    kp.handle_left_type = kp.handle_right_type = 'FREE'\n\na, b, c = fc.keyframe_points          # start, middle, end\n# Speed ramp: fast in, sag in the middle, accelerate out.\na.handle_right = (a.co.x + 2,  a.co.y + 18)    # steep = launches fast\nb.handle_left  = (b.co.x - 14, b.co.y)         # flat  = holds\nb.handle_right = (b.co.x + 14, b.co.y)\nc.handle_left  = (c.co.x - 2,  c.co.y - 18)    # steep = arrives fast\n```\n\n**Zero drift at a stop.** If a move is supposed to be locked off, it must be *actually* locked — a slow crawl at a \"stop\" reads as a mistake. Hold with `CONSTANT` interpolation, or duplicate the key so the segment is genuinely flat.\n\n## The moves\n\n### Orbit\n\nCircle the subject on a path, target at subject height. Vary radius and height across the move so it does not read as a turntable. Half-orbits and 3/4 orbits look more intentional than full ones.\n\nFor a multi-scene continuous orbit, keep one unbroken `eval_time` curve and move the *world* under it — the camera never cuts, the set changes.\n\n### Floor rise\n\nPure vertical translation, **no rotation**, smooth acceleration with a slow middle. Each \"floor\" is a different set stacked on Z at a fixed interval; the subject sits centre-frame at each pass.\n\n```python\nFLOOR_H = 4.0\nfor i in range(4):\n    cam.location = (0, -6, i * FLOOR_H)\n    cam.keyframe_insert(\"location\", frame=1 + i * 48)\n```\n\nRotation during a rise destroys the effect. Leave it out.\n\n### Robo-arm\n\nThe whip-and-lock commercial move: a fast flight along a curved arc, an **absolute** dead stop at a completely different angle, repeat. Each relocation is roughly a third of a second; each stop is a distinct, readable frame.\n\nBuild it as a path with a control point per stop, then make the stops real:\n\n```python\nHOLD_FRAMES = 10\nfor kp in fc.keyframe_points:\n    kp.interpolation = 'CONSTANT'      # hold dead still between flights\n```\n\nKeep the target separate and slightly offset per stop, so each lock-off is a different composition of the same subject rather than six centred portraits.\n\n### Handheld\n\nApplied **last**, on top of a finished move. Slow organic sway, not jitter: long waves plus a barely-perceptible tremor.\n\n```python\nfor path in (\"location\", \"rotation_euler\"):\n    for i in range(3):\n        fc = cam.animation_data.action.fcurves.find(path, index=i)\n        if fc is None:\n            continue\n        n = fc.modifiers.new('NOISE')\n        n.scale = 120        # large scale = long lazy waves (4-6s at 24fps)\n        n.strength = 0.035   # small; raise for rotation, lower for location\n        n.phase = i * 7.3    # decorrelate the axes or it reads as a slide\n```\n\n**Never fast jitter, wobble or snap corrections.** Wrong-flavour handheld is more damaging than none.\n\n### Over-the-shoulder cuts\n\nPer cut: a camera position below shoulder height, a near-foreground body mass, and a target on the far face. Move barely — a slow sideways crawl. See `slates-dialogue-blocking` for who may occupy the foreground and why it matters.\n\n### Lens\n\nAnimate focal length like any other channel; a slow lens breath under a move adds a lot for nothing.\n\n```python\ncam.data.lens = 35\ncam.data.keyframe_insert(\"lens\", frame=1)\ncam.data.lens = 50\ncam.data.keyframe_insert(\"lens\", frame=96)\n```\n\n**But the lens must not drift across a cut.** Within a cut it can animate; on the cut frame it changes instantly with everything else.\n\n## Multiple cameras and cuts\n\nTwo ways, and only one of them survives contact with a 19-shot edit:\n\n- **One camera, jump-cut it.** Keyframe location/rotation/lens with `CONSTANT` interpolation on the cut frames. Fine up to a handful of cuts.\n- **A camera per setup, bound to timeline markers.** Correct for anything bigger, because each setup stays independently editable.\n\n```python\nmarker = bpy.context.scene.timeline_markers.new(\"SHOT_04\", frame=188)\nmarker.camera = cam_shot_04\n```\n\nEither way the invariant is the same: **camera, target and lens all change on the cut frame, and no frame between two setups is interpolated.** One transition frame reads as a whip-pan, and the model will reproduce it faithfully.\n\n## Verify before you render\n\n```\nslates_blender_scene\n```\n\n`cutSeconds` is your actual cut list in seconds. Check it against the shot list you were given — mismatches here become mismatched prompt timings, and the prompt is what you write next.\n\n⚠️ **On a marker-bound rig, read `cutSeconds` or `markers`, never `camera.keyframeSeconds`.** The per-setup cameras are usually static (a Track To constraint does the aiming), so the active camera's action is empty and that field reads `[]` on a perfectly good three-cut edit. `cutSeconds` resolves to whichever the scene actually used.\n\n## Related\n\n`slates-previs-blocking` (the workflow) · `slates-blocking-to-prompt` (turning these moves into prompt text) · `slates-dialogue-blocking` (OTS and eyelines)\n",
-    "slates-character-identity": "---\nname: slates-character-identity\ndescription: Build a Slates character from a reference image — generate one identity sheet and bind it to the character so the card updates live. Use when the user wants to create a character, build a character from an image, or starts a storyboard flow that needs consistent character references.\n---\n\n# Character identity sheet — Slates workflow\n\nA character's identity sheet is attached to **every** downstream generation that mentions it, so a flaw in the sheet becomes a flaw in every shot made from it. Building it well is the highest-leverage thing you can do for a project.\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n## The shape: ONE sheet, three panels\n\nSlates generates **one identity sheet per character**, bound as the character's canonical reference:\n\n| Panel | What it carries |\n|---|---|\n| **Chest-up portrait, three-quarter angle, largest panel (~25–30% of the sheet)** | The face. **This is the only place the model reads facial identity from** — every detail it will ever know comes from those pixels, so it gets the resolution. Off-frontal, never dead-on: an angled head reads its volume instantly. |\n| **Full-body front, relaxed A-pose — cropped at the collarbone, just the face cropped out** | Build, proportion, wardrobe. The face is cropped off on purpose: a front-facing body panel renders a ~40px face that can't match the portrait's, so the sheet would carry two competing identities and the model averages them. **Only the face** — neck, arms and hands render as skin. |\n| **Full-body back, head and hair visible** | Hair fall and the back of the outfit — the only panel where either reads. Keeps its head because there's no face to compete with. |\n\nThe rule is **kill every competing rendering of the FACE, not every head** — which is why exactly one body panel is headless.\n\nOn a deep neutral-grey plate (hex `3a3a3c`, emitted without the `#` — see the sigil warning in Don'ts), flat and shadowless, with catchlights in the eyes, irises never crushed to black, surface texture at the medium's own natural level of detail, broken symmetry, and no over-clean 3D-game-model look. Expression is **a slight natural smile with the teeth just visible** — a closed mouth carries no dental information, so every downstream smiling shot invents teeth, and teeth are person-specific.\n\n**Two carve-outs, scoped differently on purpose.** Non-human characters get a natural neutral expression instead of a smile — that one is scoped by *having a human mouth*, so a bipedal robot or humanoid alien is covered. Quadrupeds and non-bipedal characters get a natural standing stance with the head shown on both body panels — that one is *anatomical*. **Both are conditionals the image model evaluates against your reference; neither is a code branch, because the op has no character-kind input.**\n\n**The sheet inherits the source's medium** — photo, anime, illustration, painterly, 3D render — unless the user explicitly asks for a transform. None of the craft clauses above override that: they ask for *readable* eyes and *material-looking* surfaces within whatever medium the character is in, not for photorealism.\n\n**Why one sheet.** Every `@character` mention attaches that character's canonical identity image, so each character costs one reference slot. It also reduces competing facial renderings to **one** — with the front panel headless and the back panel turned away, the portrait is the only face on the sheet, so there is nothing left to average.\n\n## Workflow\n\n### Get the reference\nThe user has either:\n- Pasted/uploaded an image of the character (real person, drawing, AI render).\n- Described the character in text only.\n\nIf image: upload it as a reference<!-- slates-only --> (`slates_upload_reference_image`)<!-- /slates-only -->.\nIf text only: generate from prompt-only — less consistent, so warn the user.\n\n<!-- slates-only -->\n### Create the character record\n`slates_create_character` with:\n- `name` (ask if not given)\n- `description` — 1-2 sentences, *visual* only (\"tall, dark hair, scar over left eye\"), not personality.\n- `style` — leave as the source's own medium by default. Only name a transform if the user wants one (e.g. anime → realistic).\n<!-- /slates-only -->\n\n### Generate the sheet\n<!-- slates-only -->\n`slates_generate_character_identity` with `characterId`, `projectId`, and `baseAssetId` (the source portrait).\n\n**Do not hand-write the sheet prompt.** Slates builds it from the canonical template in `@slatesvideo/shared/prompts` (`buildCharacterIdentityPrompt`) — panels, plate, lighting and craft clauses included — and appends your `userNotes`. Use `userNotes` for what the template can't know: *\"use the woman on the left\"*, *\"keep the scar on the right cheek\"*. A hand-written prompt is a fork of the template and will drift from it.\n\n- Estimate cost first with `slates_estimate_generation_cost` and announce in **credits** — never quote a price from memory.\n<!-- /slates-only -->\n\n- Default to Nano Banana 2 at 2K. **Never 4K** — no identity gain at sheet scale, wasted spend.\n- When the result returns inline, **evaluate it before binding**:\n  - Is the portrait clearly the largest panel, and is it off-frontal?\n  - **Is the front body panel cleanly headless** — an empty collar above a normally rendered body, no partial face, no floating jaw, no smeared neck stump? A botched crop is worse than no crop.\n  - **Is the body still there?** Neck, forearms and hands rendered as skin, not an empty outfit floating on nothing. A hollow garment means the invisible-mannequin genre ran unbounded.\n  - Do the body panels read as the same build, wardrobe and hair as the portrait?\n  - Catchlights present, irises readable rather than black holes?\n  - Is it in the source's medium, and does it read as *that* medium done well — or has it drifted toward the over-clean game-model look?\n  - Plate a flat deep grey, not white and not black?\n- If off: one focused refinement, then regenerate. The sheet is upstream of everything — it is worth a re-roll that a scene frame is not.\n<!-- slates-only -->\n- The op binds the result as the canonical identity automatically.\n\n### Hand back\n> \"Character {name} ready — identity sheet bound. Use `@{name}` in any prompt and Slates attaches it and names it inline, so the face stays consistent.\"\n<!-- /slates-only -->\n\n## How the reference gets used at scene time\n\nSlates cites the sheet inline under the character's name — `{name} (image N)` — in the exact order it sends references. That **name** is the anti-averaging lever, and it is each model's own official mechanism (NB2: \"assign a distinct name\"; Seedance: `Reference <Subject_N> in <Image_N>`; Kling: reuse a fixed label verbatim).\n\nCritically, the app injects **no** wardrobe, expression, or lighting directive. The user's scene prompt owns all of that — which is why `@{name}` dropped into a movie-still injection keeps the still's own clothing and lighting instead of dragging the sheet's.\n\n## Anti-patterns\n\n- **Don't** studio-light, white-background, or black-background the sheet. White bleeds into the video and washes out the location; black eats edge detail. Flat, even, shadowless light on a deep neutral grey.\n- **Don't** hand-write the sheet prompt when the op will build it — that is how the template and the shipped prompt fork.\n- **Don't** create a second character image. One canonical identity is what the storyboard pipeline reads.\n- **Don't** skip binding. An unbound asset doesn't help downstream.\n- **Don't** invent character details. Stick to what's in the reference image and the user's description.\n- **Don't** describe the front panel's crop as an absent head — in `userNotes` or any hand-written variant. The template asks for it as *framing*: **\"cropped at the collarbone, an invisible-mannequin presentation with just the face cropped out\"**, a standard e-commerce genre with deep training data. **\"the head not shown\" is a hard 422 on gpt-image-2** — fal returns `content_policy_violation` with `loc: [\"body\",\"prompt\"]`, so the text is rejected before any image is read, because an anatomical absence reads as gore to OpenAI's classifier. It passed NB2, which is why the original receipt looked safe: **it was model-scoped.** State an exclusion as a framing choice, never as a missing body part.\n- **Don't** invoke the invisible-mannequin genre without bounding it to the face. **\"an invisible-mannequin presentation where the clothing holds its own shape\" removed all the skin** — no neck, no hands, no forearms, a garment floating on nothing — because that *is* the e-commerce genre in full: an empty outfit. **\"with just the face cropped out\"** keeps the anchor and bounds it. Generalises: a genre anchor imports the whole genre, so name what STAYS, not only what goes.\n- **Don't** put `#` or `@` anywhere in prompt text. Both are reference-token sigils in the desktop prompt composer and an unresolved one is **silently deleted** — no error, no log, just missing words. `#3a3a3c` reached fal as `background ()` on a real 2026-07-30 request, meaning the plate value had never been delivered to any model since the composer shipped. Write hex values bare.\n- **Don't** use 4K — wastes credits, no quality gain at sheet scale.\n- **Don't** feed a multi-view sheet into a Seedance shot that has **several characters in frame** without binding each character to its image and appending the anti-twin constraint — ByteDance documents multi-view assets as a cause of duplicate characters. See `slates-prompting-seedance`.\n",
+    "slates-character-identity": "---\nname: slates-character-identity\ndescription: Build a Slates character from a reference image — generate one identity sheet and bind it to the character so the card updates live. Use when the user wants to create a character, build a character from an image, or starts a storyboard flow that needs consistent character references.\n---\n\n# Character identity sheet — Slates workflow\n\nA character's identity sheet is attached to **every** downstream generation that mentions it, so a flaw in the sheet becomes a flaw in every shot made from it. Building it well is the highest-leverage thing you can do for a project.\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n## The shape: ONE sheet, three panels\n\nSlates generates **one identity sheet per character**, bound as the character's canonical reference:\n\n| Panel | What it carries |\n|---|---|\n| **Chest-up portrait, three-quarter angle, largest panel (~25–30% of the sheet)** | The face. **This is the only place the model reads facial identity from** — every detail it will ever know comes from those pixels, so it gets the resolution. Off-frontal, never dead-on: an angled head reads its volume instantly. |\n| **Full-body front, relaxed A-pose — cropped at the collarbone, just the face cropped out** | Build, proportion, wardrobe. The face is cropped off on purpose: a front-facing body panel renders a ~40px face that can't match the portrait's, so the sheet would carry two competing identities and the model averages them. **Only the face** — neck, arms and hands render as skin. |\n| **Full-body back, head and hair visible** | Hair fall and the back of the outfit — the only panel where either reads. Keeps its head because there's no face to compete with. |\n\nThe rule is **kill every competing rendering of the FACE, not every head** — which is why exactly one body panel is headless.\n\nOn a deep neutral-grey plate (hex `3a3a3c`, emitted without the `#` — see the sigil warning in Don'ts), flat and shadowless, with catchlights in the eyes, irises never crushed to black, surface texture at the medium's own natural level of detail, broken symmetry, and no over-clean 3D-game-model look. Expression is **a slight natural smile with the teeth just visible** — a closed mouth carries no dental information, so every downstream smiling shot invents teeth, and teeth are person-specific.\n\n**Two carve-outs, scoped differently on purpose.** Non-human characters get a natural neutral expression instead of a smile — that one is scoped by *having a human mouth*, so a bipedal robot or humanoid alien is covered. Quadrupeds and non-bipedal characters get a natural standing stance with the head shown on both body panels — that one is *anatomical*. **Both are conditionals the image model evaluates against your reference; neither is a code branch, because the op has no character-kind input.**\n\n**The sheet inherits the source's medium** — photo, anime, illustration, painterly, 3D render — unless the user explicitly asks for a transform. None of the craft clauses above override that: they ask for *readable* eyes and *material-looking* surfaces within whatever medium the character is in, not for photorealism.\n\n**Why one sheet.** Every `@character` mention attaches that character's canonical identity image, so each character costs one reference slot. It also reduces competing facial renderings to **one** — with the front panel headless and the back panel turned away, the portrait is the only face on the sheet, so there is nothing left to average.\n\n## Workflow\n\n### Get the reference\nThe user has either:\n- Pasted/uploaded an image of the character (real person, drawing, AI render).\n- Described the character in text only.\n\nIf image: upload it as a reference<!-- slates-only --> (`slates_upload_reference_image`)<!-- /slates-only -->.\nIf text only: generate from prompt-only — less consistent, so warn the user.\n\n<!-- slates-only -->\n### Create the character record\n`slates_create_character` with:\n- `name` (ask if not given)\n- `description` — 1-2 sentences, *visual* only (\"tall, dark hair, scar over left eye\"), not personality.\n- `style` — leave as the source's own medium by default. Only name a transform if the user wants one (e.g. anime → realistic).\n<!-- /slates-only -->\n\n### Generate the sheet\n<!-- slates-only -->\n`slates_generate_character_identity` with `characterId`, `projectId`, and `baseAssetId` (the source portrait).\n\n**Do not hand-write the sheet prompt.** Slates builds it from the canonical template in `@slatesvideo/shared/prompts` (`buildCharacterIdentityPrompt`) — panels, plate, lighting and craft clauses included — and appends your `userNotes`. Use `userNotes` for what the template can't know: *\"use the woman on the left\"*, *\"keep the scar on the right cheek\"*. A hand-written prompt is a fork of the template and will drift from it.\n\n- Estimate cost first with `slates_estimate_generation_cost` and announce in **credits** — never quote a price from memory.\n<!-- /slates-only -->\n\n- Default to Nano Banana 2 at 2K. **Never 4K** — no identity gain at sheet scale, wasted spend.\n- When the result returns inline, **evaluate it before binding**:\n  - Is the portrait clearly the largest panel, and is it off-frontal?\n  - **Is the front body panel cleanly headless** — an empty collar above a normally rendered body, no partial face, no floating jaw, no smeared neck stump? A botched crop is worse than no crop.\n  - **Is the body still there?** Neck, forearms and hands rendered as skin, not an empty outfit floating on nothing. A hollow garment means the invisible-mannequin genre ran unbounded.\n  - Do the body panels read as the same build, wardrobe and hair as the portrait?\n  - Catchlights present, irises readable rather than black holes?\n  - Is it in the source's medium, and does it read as *that* medium done well — or has it drifted toward the over-clean game-model look?\n  - Plate a flat deep grey, not white and not black?\n- If off: one focused refinement, then regenerate. The sheet is upstream of everything — it is worth a re-roll that a scene frame is not.\n<!-- slates-only -->\n- The op binds the result as the canonical identity automatically.\n\n### Hand back\n> \"Character {name} ready — identity sheet bound. Use `@{name}` in any prompt and Slates attaches it and names it inline, so the face stays consistent.\"\n<!-- /slates-only -->\n\n## How the reference gets used at scene time\n\nSlates cites the sheet inline under the character's name — `{name} (image N)` — in the exact order it sends references. That **name** is the anti-averaging lever, and it is each model's own official mechanism (NB2: \"assign a distinct name\"; Seedance: `Reference <Subject_N> in <Image_N>`; Kling: reuse a fixed label verbatim).\n\nCritically, the app injects **no** wardrobe, expression, or lighting directive. The user's scene prompt owns all of that — which is why `@{name}` dropped into a movie-still injection keeps the still's own clothing and lighting instead of dragging the sheet's.\n\n## Anti-patterns\n\n- **Don't** studio-light, white-background, or black-background the sheet. White bleeds into the video and washes out the location; black eats edge detail. Flat, even, shadowless light on a deep neutral grey.\n- **Don't** hand-write the sheet prompt when the op will build it — that is how the template and the shipped prompt fork.\n- **Don't** create a second character image. One canonical identity is what the storyboard pipeline reads.\n- **Don't** skip binding. An unbound asset doesn't help downstream.\n- **Don't** invent character details. Stick to what's in the reference image and the user's description.\n- **Don't** describe the front panel's crop as an absent head — in `userNotes` or any hand-written variant. The template asks for it as *framing*: **\"cropped at the collarbone, an invisible-mannequin presentation with just the face cropped out\"**, a standard e-commerce genre with deep training data. **\"the head not shown\" is a hard 422 on GPT Image** (measured on `gpt-image-2`, the model 2.5 replaced; the classifier is OpenAI's, not the version's, so the rule carries — but nobody has re-run it on Flare or Sunburst) — fal returns `content_policy_violation` with `loc: [\"body\",\"prompt\"]`, so the text is rejected before any image is read, because an anatomical absence reads as gore to OpenAI's classifier. It passed NB2, which is why the original receipt looked safe: **it was model-scoped.** State an exclusion as a framing choice, never as a missing body part.\n- **Don't** invoke the invisible-mannequin genre without bounding it to the face. **\"an invisible-mannequin presentation where the clothing holds its own shape\" removed all the skin** — no neck, no hands, no forearms, a garment floating on nothing — because that *is* the e-commerce genre in full: an empty outfit. **\"with just the face cropped out\"** keeps the anchor and bounds it. Generalises: a genre anchor imports the whole genre, so name what STAYS, not only what goes.\n- **Don't** put `#` or `@` anywhere in prompt text. Both are reference-token sigils in the desktop prompt composer and an unresolved one is **silently deleted** — no error, no log, just missing words. `#3a3a3c` reached fal as `background ()` on a real 2026-07-30 request, meaning the plate value had never been delivered to any model since the composer shipped. Write hex values bare.\n- **Don't** use 4K — wastes credits, no quality gain at sheet scale.\n- **Don't** feed a multi-view sheet into a Seedance shot that has **several characters in frame** without binding each character to its image and appending the anti-twin constraint — ByteDance documents multi-view assets as a cause of duplicate characters. See `slates-prompting-seedance`.\n",
     "slates-content-policy": "---\nname: slates-content-policy\ndescription: Read before writing any prompt involving conflict, creatures, crowds, destruction, weapons, or young characters. Builds the scene safe from the first word, so it hits full cinematic impact without depicting prohibited content — and without being silently rejected or degraded by the model's filter. Mirror of @slatesvideo/shared/prompts content-policy fragment — SSOT: second-brain business/projects/slates/product/prompting-ssot.md.\n---\n\n# Content-policy-safe construction — read before any risk-surface prompt\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** — each one is a filter tripwire with a substitution in the table above:\n- `civilians in panic`, `crowds fleeing`, `blood`, `gore`, `corpse`\n- `ignite`, `catch fire`, `on fire` applied to a person — frame body-contact effects as magical or harmless VFX\n- `candle-like`, `flame-like` and any real object used as a metaphor for an effect\n- a real, named public figure\n<!-- @banned:end -->\n\nDon't depict the harm — depict the energy, the aftermath, the threat, or the scale. Build the scene safe from the first word.\n\nWrite scenes that hit full cinematic impact without ever *needing* to depict prohibited content. This is a craft move, not a compromise — the substitutions below usually read as more cinematic, not less, and they keep your generation from getting silently rejected or degraded by the model's filter. A standoff is more tense than a massacre; an evacuated city is eerier than a crowd in panic; a roar lands harder than a kill. Load this whenever a prompt involves conflict, creatures, crowds, destruction, weapons, or young characters.\n\n## Substitution table\n\n| Avoid | Use instead |\n|---|---|\n| Civilians in panic, crowds fleeing under debris | An evacuated / empty city; abandoned streets; a lone figure for scale |\n| Weapons firing into buildings or at people | Energy-discharge standoffs, searchlights sweeping, charged auras, shockwaves with no muzzle fire |\n| Creatures tearing into each other, gore | A grapple / standoff — roars, near-misses, circling, an energy clash; combat that stays contained (in/on the water, never lifting into the air) |\n| Destruction with people in harm's way | Destruction in uninhabited terrain — glaciers, deserts, ruins, open sea, evacuated zones |\n| Realistic guns as the focus | Stylized / fantasy implements, weapons slung-not-fired, the weapon as silhouette or prop only |\n| Blood, wounds, death | Impact light, dust, debris, buckling and collapse, a silhouette dropping out of frame |\n| Real, named public figures | Original / anonymous characters |\n| Real brand logos | Original or generalized branding — except the user's own product, which is the whole point of a brand film |\n\n## The safe benchmark\n\nWhen a scene starts drifting risky, pull it back toward this shape: **one original creature, in a generalized monument or amphitheatre, in daylight, no weapons present, performing expressive action** (rising, roaring, spreading wings). Original design, generalized location, daylight, expressive rather than violent. That's the confirmed-safe envelope — most epic ideas re-stage into it without losing the punch.\n\n## Containment rule — it doubles as a physics win\n\nGive any creature or combat scene a **containment rule** that grounds the physics at the same time:\n- \"the fight STAYS at the sea surface — they breach, dive, grapple, submerge, but never fly or get carried into the air\"\n- \"boss scale locked ~2.5 human-heights, NOT kaiju-giant\"\n- \"destruction stays in the evacuated valley\"\n\nThis improves coherence (the model isn't inventing absurd escalation) AND keeps the scene inside policy — same clause buys both.\n\n## Scale and stakes without harm\n\nEpic stakes come from environmental danger and reaction, not depicted victims: tiny figures diving clear of *collapsing* terrain (not being crushed), a war-horn over an *empty* field, an army *scrambling* across a frozen valley as a titan tears free of a glacier. The danger is the environment; the figures are reacting, not dying. Snow plumes, glowing runes, splintering ice, shockwaves, and dust carry the chaos.\n\n## Minors — hard rule\n\nNever write romantic, sexual, or suggestive content involving or directed at minors, and never anything that sexualizes a young-presenting character. Any scene with children stays wholesome and age-appropriate. Non-negotiable — it overrides every stylistic goal.\n\n## Pre-flight (run before delivering any risk-surface prompt)\n\n- [ ] No civilians depicted in panic/harm; crowds are evacuated or absent.\n- [ ] No weapons firing at people/buildings; threat is energy / searchlight / silhouette.\n- [ ] No creature-on-creature or creature-on-person gore; combat is grapple / standoff / roar, contained.\n- [ ] Destruction is in uninhabited / evacuated terrain.\n- [ ] Creatures are original (\"not based on any franchise\"); no real public figures; no real brand logos except the user's own product.\n- [ ] Anything with children is wholesome and age-appropriate.\n\nIf a box fails, apply the substitution table before writing the prompt.\n\n## Editing real footage (Kling O3 edit / Omni Flash edit) — real people in the SOURCE\n\nVideo edit takes the user's own footage, which often contains real people. Rules:\n\n- The user must hold rights/consent for any real person's likeness in footage they edit — ask once when it's clearly someone other than the user, then proceed.\n- Kling's video-to-video filter behavior on real faces is **not yet verified** (unlike Seedance, where the consent-gated real-face route is confirmed). If an edit of real-person footage is rejected by the provider, do NOT retry-spam variations — tell the user the filter blocked it and offer a no-face crop/segment or an AI-character swap instead.\n- **Omni Flash: own-footage editing of the uploader's own face PASSED live 2026-07-09** (real talking-head clip, edited on our fal route) despite Google's documented \"recognizable people\" restriction — treat that restriction as aimed at third-party/public figures, but expect probabilistic refusals and never promise passage.\n- Never use edit to put a real, named public figure into a scene, or to make someone appear to say/do something they didn't. Faceless b-roll (hands, products, landscapes, crowds-from-behind) edits freely.\n\n## Gemini / Omni Flash filter regime (video gen + edit) — receipts 2026-07-09\n\nGoogle's filter is its own regime (stricter than fal-hosted Kling about harm-to-a-person, looser than BytePlus about faces). Live receipts:\n\n| Blocked (`content_policy_violation`) | Passed |\n|---|---|\n| \"his fingertips **ignite** with a small real flame\" (fire ON a body part = harm) | \"small **magical** flames appear on his fingertips … vanish when he blows on them\" |\n\n- **Harm-to-person framing is the tripwire**, not the effect itself. Reframe body-contact effects as magical / supernatural / harmless VFX: \"magical flames\", \"a glowing aura\", \"sparks of light dance on\". Avoid ignite / burn / on fire / catch fire applied to a person.\n- **Never use a real object as a metaphor** — \"candle-like flame\" rendered a literal candle in the subject's hand. Describe the effect, not an object that resembles it.\n- The block is a 422 refund (no credits lost) and arrives mid-generation — one reframe per the substitution mindset above, don't retry-spam.\n",
     "slates-cost-discipline": "---\nname: slates-cost-discipline\ndescription: Mandatory pre-flight discipline before ANY generation call (image or video) — estimate cost, announce in credits, get confirmation, aggregate batches. Read this every time before calling slates_generate_image or any future slates_generate_* op. Skipping this risks burning the user's credits on guesses.\n---\n\n# Slates cost discipline — read before every generation\n\nGeneration costs real money. Every call is on the user's credits. The user can't see what you're about to spend until you tell them. **Tell them first, generate second.**\n\n## The 4 rules\n\n### 1. Pre-flight estimate — never call generate without one\n\nBefore ANY `slates_generate_*` call, run `slates_estimate_generation_cost` first. Inputs you must lock before estimating:\n\n- **Model** — the id you are about to pass, whatever it is. `slates_estimate_generation_cost` takes the same base ids the generate ops take and resolves the billing key itself; do not build one by hand.\n- **Resolution** — never let the op default. Pick deliberately. Drafts → 1k. Hero → 2k. Print → 4k.\n- **Aspect ratio** — never let the op default to 1:1. Pick from the use case (cinematic → 16:9, mobile vertical → 9:16, square feed → 1:1).\n- **Count** — explicit. Don't generate 4 when 1 will tell you if the prompt works.\n\nIf aspect ratio or resolution isn't obvious from the user's request, **ask before estimating**. Don't guess.\n\n### 2. Announce in credits, plainly, before spending\n\nSlates bills abstract **credits** (they never expire). Announce the credit total the estimate returns — never dollars.\n\nFormat: `About to spend N credits on M image(s) at [resolution] [aspect ratio]. Proceed?`\n\nExamples:\n- `About to spend 4 credits on 1 image at 1k 16:9. Proceed?`\n- `About to spend 24 credits on 4 images at 2k 9:16 (variants). Proceed?`\n\nAnnounce the cost once, then proceed for anything small. For anything the user would notice on their balance, wait for an explicit yes. Where the LINE is, exactly:\n\n<!-- @inject:thresholds -->\n<!-- GENERATED from @slatesvideo/shared — do not edit between the markers.\n     Source: CONFIRM_CREDITS, DEVIATION_FACTOR and the audio bounds in\n     packages/shared/src/operations/index.ts. Every number here is REFUSED by an\n     op when a prompt gets it wrong, which is why none of them is typed by hand\n     any more: this block replaced four claims that contradicted the code. -->\n\n**The thresholds, from the code that enforces them:**\n\n- **Confirm gate:** above **17 credits** an op returns `requires_confirm` and will not\n  proceed until you re-call with `confirm: true`. Below it, announce the cost once and go.\n- **Deviation pause:** the desktop Studio Agent stops and re-asks when projected generation spend\n  exceeds the approved plan by more than **20%**. You do not trigger this; the app does.\n- **Seed Audio duration:** **3–120 seconds.** There is no duration\n  parameter on the model — the number you pass is written into the prompt AND is what the user is\n  billed. Outside that range the op refuses rather than clamping.\n- **Sound Effects duration:** **1–22 seconds**, billed per second, never left for the\n  model to pick.\n\nNever quote a credit figure from memory: `slates_estimate_generation_cost` returns the real one.\n<!-- @end:thresholds -->\n\n### 3. Aggregate batches into ONE upfront announcement\n\nIf you're planning a multi-call workflow (5 storyboard frames, 3 character variants, a grid of options), **announce the total before the first call**, not five small announcements after the fact.\n\nFormat: `Plan: N generations totaling C credits. [Brief description of the sequence.] Proceed with the batch?`\n\nExample: `Plan: 6 frame generations at 1k 16:9 totaling 24 credits — establishing wide, push-in, two-shot, reverse, OTS, insert. Proceed?`\n\n### 3b. Batch authorization — one approval covers the enumerated batch\n\nWhen the user approves a batch plan with one aggregated cost total up front (\"8 scenes, ~$X total — go\"), that single approval authorizes `confirm=true` on **each enumerated call in that batch** — and nothing beyond it. You do not need to re-ask per call; that's the point of the upfront announcement. Hands-off multi-scene runs depend on this.\n\nBoundaries that re-trigger confirmation:\n\n- Any call's actual estimate exceeds what the announced plan implied → stop, surface the delta, get a fresh OK. The app enforces its own ceiling on top of this (see the thresholds above); do not wait for it to catch you.\n- New calls are added that weren't in the enumerated plan (extra variants, retries beyond the plan, a new scene) → those are NOT covered. Announce and confirm separately.\n- The batch scope changes (different model, resolution, or duration than announced) → re-announce, re-confirm.\n\nOne approval = that plan, as enumerated, at those prices. Nothing else.\n\n### 4. Track the running total\n\nAfter each generation completes, the response includes `cost_credits` (when available). Keep a running tally in your context. Surface it every 3 generations or whenever the user asks \"how much have we spent?\"\n\n## Resolution decision rules\n\n| Use case | Resolution |\n|---|---|\n| First draft of a new prompt | 1k |\n| Storyboard frame (will likely regenerate) | 1k |\n| Hero shot, locked composition | 2k |\n| Print, marketing asset, final delivery | 4k |\n| Iterating to refine | match the previous resolution |\n\nResolution is a price lever, not a free choice: on Nano Banana 2 and FLUX.2 Max, 4k costs roughly 2x 1k (Seedream 5 Lite is flat-priced regardless of resolution). Prices change — call `slates_estimate_generation_cost` or `slates_list_available_models` for current numbers instead of assuming. Pick the cheapest resolution that serves the use case.\n\n**4K VIDEO is Pro-only (2026-07-07).** The ladder above is for IMAGES (open at every tier). For VIDEO — Kling, Seedance, Veo — 4K requires a Slates Pro account; a base-tier 4K video gen is rejected server-side with `PRO_REQUIRED`. Default video to 1080p or lower and only reach for 4K when the user is on Pro and explicitly asks. 4K *images* are never gated.\n\n## Aspect ratio decision rules\n\nAsk the user when ambiguous. Otherwise:\n\n| Context cue | Aspect ratio |\n|---|---|\n| \"cinematic\", \"film\", \"movie\", \"wide\" | 16:9 |\n| \"TikTok\", \"Reels\", \"Story\", \"mobile vertical\", \"phone\" | 9:16 |\n| \"square\", \"Instagram feed\", \"thumbnail\" | 1:1 |\n| \"ultra-wide\", \"anamorphic\", \"cinemascope\" | 21:9 |\n| \"portrait\", \"magazine cover\", \"vertical\" | 4:5 or 2:3 |\n| \"landscape photo\", \"horizontal\" | 3:2 or 4:3 |\n\nIf the user prompt mixes signals (e.g. \"cinematic Instagram post\"), ask. Don't guess.\n\n## When the gate fires\n\nThe server returns `requires_clarification` when aspect ratio or resolution is missing, and `requires_confirm` when total spend crosses the gate above. In both cases:\n\n1. Surface the gate response to the user\n2. Get a clean answer\n3. Re-call with the explicit values + `confirm: true` if applicable\n\nDon't bypass the gate by silently filling in defaults. The gates exist because defaults waste money.\n\n## Video is slow + async — a timeout is NOT a failure\n\nVideo gens take minutes (Seedance 4K can run far longer). A client/CLI timeout or a slow, empty-looking response is **not** a failed generation — the job is still running on the provider.\n\n- **Never re-submit a video gen because it \"timed out.\"** That double-charges the user for one video. Re-rolling a slow gen is the single most expensive mistake here.\n- **Poll, don't re-roll.** Use `background: true` on `slates_generate_video`, then poll `slates_get_generation_status` (free, read-only) until it reports `completed` or `failed`. In-flight jobs survive app restarts and are recovered.\n- A gen has only failed when the status comes back `failed` — and a provider *rejection* **refunds** the credits, so failed isolation tests are ~free. Until you see a terminal status, the job is in flight. Wait.\n\n## 🔴 The still-gate — the most expensive mistake in the pipeline\n\n<!-- @inject:still-gate -->\n**A visible defect in the still is already a STOP.** Do not animate it. Fix the frame first, then move to motion — and go to motion only when the crop passes the still scan and you genuinely need movement to confirm an uncertain edge, reflection, or object.\n\nThis is a **cost** rule as much as a craft rule: a 1080p/10s premium video generation costs many multiples of an image re-roll, and video is where a defect stops being fixable. Anything wrong in the still gets worse in motion — soft geometry mushes, broken-but-plausible objects fall apart, oily textures start crawling. **Animating a known-bad frame is the single most expensive mistake in the pipeline.** Re-rolling the image is the cheap move; re-rolling the video is not.\n<!-- @end:still-gate -->\n\nThe check itself lives in `slates-vision-feedback-loop` (the four slop tells and the per-model accents). The **stop** is a cost rule and belongs here: before every image→video call, confirm the source frame passed the still scan. If it didn't, spending video credits on it is not iteration — it is buying a more expensive copy of a defect you already found.\n\n## The 3-strike rule\n\nStop after 3 iterations on the same prompt. Hand back to the user with what you tried and what's not working. The slot machine doesn't converge — if it's not landing, the prompt structure is wrong, not the seed.\n",
     "slates-dialogue-blocking": "---\nname: slates-dialogue-blocking\ndescription: Keep multiple characters spatially consistent across cuts — seating, screen direction, eyelines, the 180-degree rule — by blocking the scene in 3D first. Use for any multi-character dialogue scene, conversations around a table or in a car, or when generated characters swap seats, change sides, or look the wrong way between shots.\n---\n\n# Dialogue blocking — six people who stay where you put them\n\nThe hardest thing to generate, and the case where previs beats raw prompting by the widest margin.\n\n## Why this is hard\n\nEvery cut is an independent guess unless something forces agreement. Prompt a six-person conversation four times and you get four different seating charts: characters swap places, the 180-degree line breaks, and nothing cuts together. The failure is not aesthetic — the shots are simply unusable as an edit, and you find out only after paying for all four.\n\n**The blocking fixes it structurally.** Positions exist in 3D, so every camera sees the same arrangement, and consistency stops being something the model has to remember.\n\n## Build\n\nFollow `slates-previs-blocking` and add these.\n\n### Seated proxies, colour-coded\n\nSimple seated shapes. **Do not animate the heads** — a proxy head turning the wrong way is worse than one that never turns.\n\nGive each character a distinct viewport colour and write the mapping down. This is the identity channel:\n\n> red = the boss · green = the kid · blue = the driver · yellow = the fixer · purple = the cousin · cyan = the nephew\n\nThat mapping goes verbatim into the generation prompt. It is what lets the model bind a grey body to a character sheet across four cuts.\n\nGive every proxy a material and set `mat.diffuse_color` to its identity colour — the blocking render pins Workbench to `MATERIAL` shading, so **the material's `diffuse_color` is what reaches the clip**. Set `object.color` to the same value too, so the user's viewport matches what renders. See `slates-previs-blocking` for the snippet.\n\n### Fix the geography, then never move it\n\nPlace people once. Write down who sits where relative to the camera's opening position, in words, because that sentence is going into the prompt:\n\n> Across the table, facing camera: yellow dead centre, purple far left, blue and green to the right.\n\n### The camera plan\n\nPer `slates-camera-language`, with two things specific to dialogue:\n\n- **Below shoulder height, slow rail glides.** Eye-level-and-above reads as surveillance.\n- **Decide who owns the near foreground in each cut and honour it.** A shoulder in frame is a spatial anchor; a different shoulder in the next cut relocates the whole room.\n\nThe move that earns the most: **a gaze handoff without a cut** — the camera keeps gliding while the target hands off across the table, face to face, slowing on each but never stopping. Build it by keyframing the Track To target's position between subjects.\n\n### Crossing behind someone\n\nA head wiping frame during a move is a strong depth cue. It is also a spatial claim, so pick who gets crossed and say so — *the camera crosses directly behind cyan's back mid-shot and his head wipes the frame once.*\n\n## The prompt\n\nEverything in `slates-blocking-to-prompt`, plus these blocks.\n\n### Geography — restate it as a rule\n\n> TABLE GEOGRAPHY — do not deviate: the camera is never parked behind red. Only in the opening seconds does his dark shoulder hang at the near frame RIGHT edge, and it slides out as the camera travels LEFT. The true near-foreground of this shot is cyan: the camera crosses directly behind him mid-shot. After the opening seconds red is gone from the foreground, and the camera never travels behind anyone except cyan.\n\n### Screen direction, and the mirror that is not a swap\n\nThe 180-degree rule survives on its own in the blocking. What breaks is the model **\"correcting\" a legitimate mirror** — when the camera faces back through a scene, sides invert, and that inversion is correct. Say so explicitly or it gets flipped:\n\n> The driver's seat is on the LEFT for the entire timeline; this layout never mirrors or flips. When a camera faces BACKWARD into the car, screen sides mirror naturally: the driver reads on the RIGHT of frame, the passenger on the LEFT — that is correct left-hand drive, not a swap. They never swap seats or roles anywhere in the timeline.\n\nThen compress it into the HOLD block: *(backward camera mirrors them: he right of frame, she left)*.\n\n### Presence\n\n> A seated person stays drawn even when partially occluded — in every interior frame some part of each seat's owner is visible: a hand, an arm, a shoulder, a head above the bolster. Every occupied seat visibly holds its person.\n\n### Keep everyone alive\n\nThree orthogonal layers. Without them, whoever is not speaking freezes:\n\n- **ONGOING BUSINESS** — a small continuous physical action per character, running whether or not they are speaking. *turns his glass a quarter every few seconds · thumbs a lighter without lighting it.*\n- **BACKGROUND LIFE** — soft-focus, low contrast, never pulls attention, never crosses in front of a speaking face.\n- **SCENE EVENT** — the unnamed thing everyone is playing but nobody says. One line, repeated verbatim in every character's direction: *keep tomorrow sounding like a fishing trip.*\n\n### Acting, per character\n\nSame six slots each. Terse:\n\n```\nACTING TASK — <character>\n  SCENE DIRECTION (shared, unspoken): <the same line for everyone>\n  MOTIVE (his fuel): <what he wants underneath>\n  GOAL: <what he wants in this scene>\n  OBSTACLE: <what is in the way>\n  TACTIC: <how he goes about it>\n  Moment to moment: <2-3 beats keyed to timestamps>\n  (Safety: gaze always engaged in the task — never a frozen, glassy,\n   unfocused stare; natural blink cadence.)\n```\n\nThat safety line is not filler. Dead eyes are the characteristic failure of generated faces in dialogue, and naming it is what prevents it.\n\n### Split any strong emotion into phases\n\nThe other characteristic failure is a face that strikes one extreme expression and holds it for the whole shot — a mouth stuck open for three seconds. Give the beat two phases with a hinge, and name the failure you are excluding:\n\n> PHASE 1 (17.3-18.6s) — she SCREAMS at him, mouth wide, eyes huge, hand clamped on the grab handle. PHASE 2 (18.6-19.9s) — the scream breaks off: she shuts her eyes tight and CLOSES her mouth, both hands now on the handle, head ducked, braced. Scream, then brace — never one frozen open mouth held through the whole shot.\n\nThe hinge timestamp is what makes it a performance instead of a pose.\n\n### Dialogue must not restructure the edit\n\nBoth of these, verbatim, every time:\n\n> DIALOGUE NEVER CREATES SHOTS: spoken lines happen inside the reference's takes exactly as blocked — no cutaways to a speaker, no reverse shots, no added close-ups. If a line plays while the camera is elsewhere, the line stays off-screen audio.\n\n> OFF-SCREEN VOICES RULE: a line marked off-screen must STAY off-screen — never show the speaker, never move him into frame, never route the camera behind him because he spoke.\n\nA sentence may cross a cut. Say so where it does: *the sentence does not pause for the edit.*\n\n## Model routing\n\nDialogue directed as separate layers (voices, scene sound, score) is **minimax-h3**'s seat; it also takes declared reference relationships, which suits a colour-coded cast. Native synced audio is **Veo**'s niche. seedance-2.5 carries the reference-video capacity. Route per `slates-model-selection` and read the chosen model's prompting skill before writing the audio block.\n\n## Checklist\n\n- [ ] Colour→character mapping written down and pasted into the prompt\n- [ ] Heads not animated in the blocking\n- [ ] Seating stated as a geography rule\n- [ ] Foreground owner named per cut\n- [ ] Mirror-is-not-a-swap clause present if any camera faces back through the scene\n- [ ] Presence rule present\n- [ ] Ongoing business, background life and scene event all specified\n- [ ] Acting task per character, safety line included\n- [ ] Any strong emotion split into phases with a hinge timestamp\n- [ ] Dialogue-never-creates-shots and off-screen-voices rules present\n\n## Related\n\n`slates-previs-blocking` · `slates-camera-language` · `slates-blocking-to-prompt` · `slates-character-identity` (the sheets) · `slates-prompting-minimax-h3`\n",
     "slates-direct-response-ad": "---\nname: slates-direct-response-ad\ndescription: Build a 30-second hyper-motion direct-response ad in Slates from a product image and brief. Composes upload → storyboard → frame gen → motion gen → timeline → export. Use when the user drops a product image and asks for \"an ad\", \"a promo video\", \"a TikTok ad\", \"an Instagram ad\", a launch video, or any short-form direct-response video built around a product.\n---\n\n# Direct-response ad — Slates workflow\n\n🚨 **Wrong skill if a PERSON talks to camera.** This file builds a product-led hyper-motion spot — product hero frames, punchy cuts, no presenter. **A creator-style ad where a synthetic person speaks to the lens is a different discipline with an inverted rulebook (ugly on purpose, one shot per generation, plate-before-video): read `slates-ugc-influencer-ad`.** Route on the presence of a talking person, not on the platform.\n\nYou are building a 30-second hyper-motion direct-response ad. The user has handed you a product image (or product URL) and a short brief. Slates desktop is open on the second monitor; the user watches it populate as you work.\n\n**Hard rules**\n\n- Always estimate cost before generating. Use `slates_estimate_generation_cost` and surface the total.\n- All Slates generation routes through Slates Credits, period (BYOK is retired) — don't suggest \"use your own keys\" workarounds.\n- Default model: `nano-banana-2-2k`. For close-up product hero frames step up to `4k` only if the user asks.\n- Hyper-motion = punchy cuts, 4 frames in 30 seconds, ~7s each. Don't over-storyboard.\n\n## Workflow\n\n### 1. Set up the project\n- Create a project named for the product (`slates_create_project`).\n- If the user gave a product image as a file path, upload it (`slates_upload_reference_image`).\n- If they pasted base64 / a data URL, use the same op with `dataUrl`.\n\n### 2. Generate the storyboard frames\nBuild exactly 4 frames in this order:\n\n| # | Beat | Visual goal |\n|---|------|-------------|\n| 1 | Hook | Hyper-close-up of the product, dramatic light, motion blur edge |\n| 2 | Lifestyle | Real person using/wearing/holding the product, eye contact |\n| 3 | Problem→solution | The before/after moment that justifies the buy |\n| 4 | CTA | Clean product hero with mental room for an overlaid CTA in editing |\n\nFor each frame:\n1. Draft a tight 1-2 sentence prompt (visual only — no copy text in the image).\n2. Reference the product upload's URL or asset ID for visual fidelity.\n3. Call `slates_generate_image` with that prompt + reference. **You see the result inline — evaluate it.**\n4. If it's wrong: refine prompt, regenerate. If it's right: bind it as a frame in the storyboard (`slates_add_frame`).\n\n### 3. Build the storyboard\n- `slates_create_storyboard` named \"30s ad — v1\".\n- Default scene already exists. Add 3 more scenes (\"Hook\", \"Lifestyle\", \"Problem-Solution\", \"CTA\") via `slates_add_scene`, or just add all 4 frames to the default scene.\n- For each generated image, add a frame referencing the asset id (`slates_add_frame`).\n\n### 4. Hand back to the user\n- Surface estimated total credits spent.\n- Tell the user the storyboard is ready and they can either:\n  - **In Slates desktop:** click each frame to generate motion (the existing UI handles motion generation).\n  - **Continue here:** ask you to keep going.\n\n### 5. If they say keep going — motion, assembly, export\n- Generate motion per frame with `slates_generate_video` (`firstFrameAssetId` = the frame's asset, `background: true`), routed per `slates-model-selection` (Kling 3.0 std 8s by default; Seedance 2 for any physics-heavy beat like the hook). Submit all four, then poll `slates_get_generation_status` until each completes (1-5 min).\n- Assemble: `slates_add_clip_to_timeline` for each completed clip in beat order (Hook → Lifestyle → Problem-Solution → CTA). Verify with `slates_get_timeline`; fix order with `slates_reorder_clips`.\n- Export: `slates_export_video` to an absolute `.mp4` path (default `<slates_get_project_directory>/exports/<product>-ad.mp4`), then `slates_reveal_file` so the user sees the file.\n- Full pipeline doctrine (batch cost authorization, model mixing, multi-take selection): `slates-one-prompt-film`.\n\n## Anti-patterns\n\n- **Don't** generate text overlays in the image. Slates renders captions/CTAs at the editor stage.\n- **Don't** burn credits on slot-machine prompting. If the first generation is off, refine the prompt; don't just regenerate.\n- **Don't** skip the cost estimate. Confirm with the user above ~17 credits.\n- **Don't** invent visual specifics about the product (colors, textures, angles) that aren't in the reference image. Reference-anchored prompts only.\n\n## Voice\n\nThe ad lives or dies on the hook frame. Tight, sensory, no fluff. Match the user's brand. Default tone is \"scroll-stopping\" not \"informative.\"\n",
     "slates-edit-and-iterate": "---\nname: slates-edit-and-iterate\ndescription: Iterate on an existing Slates asset — re-evaluate, refine prompt, regenerate or edit. Use when the user has an existing generated image in Slates and wants to \"tweak it\", \"change one thing\", \"make it warmer\", \"remove the second person\", or any other surgical refinement instead of full regeneration.\n---\n\n# Edit and iterate — Slates workflow\n\nThe user already has a generated image in Slates and wants to refine it. The vision-feedback-loop skill defines the general pattern; this skill is the specific recipe for \"I have asset X, here's what's wrong with it.\"\n\n## 🔴 The master rule — an edit is a LEAF, not a node\n\n**Never re-edit an edit. Always go back and re-edit the master.**\n\nEvery edit model silently re-renders the **whole frame**, not just the region you named. So the parts you didn't ask to change come back slightly different every pass — softer texture, drifted colour, mushier fine detail. It is barely visible after one edit and obvious by the second. Chaining edits compounds the damage and there is no way to undo it, because each generation *is* the new source.\n\nThe fix is structural, not a matter of care:\n\n- **Want two changes?** Make them in ONE edit off the master, or make them as two separate edits **both taken from the master**, then keep whichever you prefer.\n- **An edit came back wrong?** Do NOT edit the result to fix it. Discard it and re-edit the master with a better instruction.\n- **Only the changed region is worth keeping?** That is a compositing job — the edit supplies the new region, the untouched master supplies everything else.\n\nSlates records this: an edit result carries `sourceAssetIds` pointing at the asset it was made from, so **you can tell whether the thing you are about to edit is itself an edit.** Check before you edit — `[Edit]`-prefixed prompts and a populated source lineage both say \"this is a leaf; go back to its parent.\"\n\n## Workflow\n\n### 1. Pull the current asset back into context\n- The user references an asset by id, frame number, or \"the latest one.\"\n- Resolve to an asset id (`slates_list_assets` if needed).\n- `slates_get_asset_image` with that id to load it inline. **You see the image.**\n\n### 2. Identify the delta\nThe user's request is one of:\n- **Surgical** — \"remove the second figure\", \"make the sword red\", \"swap the background to a bamboo forest\".\n- **Aesthetic** — \"warmer light\", \"more dramatic\", \"softer focus\".\n- **Compositional** — \"wider shot\", \"lower angle\", \"centered subject\".\n- **Wholesale** — \"actually let's try a totally different look.\"\n\n### 3. Pick the right tool\n| Delta type | Approach |\n|---|---|\n| Surgical | `slates_edit_image` — `sourceAssetId` = the original, `prompt` = the change only (\"remove the second figure\"), not a re-description of the whole image. |\n| Aesthetic / compositional | `slates_generate_image` with the original in `referenceAssetIds` + a refined prompt. Don't re-roll from scratch. |\n| Wholesale | New prompt, no reference, fresh generation. Treat as a new brief. |\n\n**`slates_edit_image` shape:** `projectId` + `sourceAssetId` + `prompt` (the edit instruction). Default model `nano-banana-2` — the only edit model that also takes extra `referenceAssetIds`; `flux-2-max` / `seedream-5-lite` use their own edit endpoints and ignore references. The result lands as a NEW asset (prompt prefixed `[Edit]`); the source is untouched. Cost above ~17 credits gates on `confirm=true`.\n\n### 4. Generate, evaluate, decide\n- Estimate cost first.\n- After generation, the result is inline. Compare side-by-side with the original (`slates_get_asset_image` again).\n- If the delta is correct: bind to the same role (frame, character identity, etc.) the original was bound to.\n- If the delta missed: one focused refinement, then regenerate. Cap at 3 tries.\n\n### 5. Hand back\n- \"Asset updated. Frame 3 now uses {new_asset_id}.\"\n- Always note what changed and what didn't, so the user can see the surgery worked: \"Lighting shifted to warmer, composition unchanged.\"\n\n## Anti-patterns\n\n- **Don't** delete the original asset until the user confirms the new one. Slates keeps both; the user picks.\n- **Don't** mix surgical and wholesale changes in one regeneration. The user said \"make it warmer\" — don't also reframe the shot.\n- **Don't** re-generate when `slates_edit_image` would work. Edits preserve composition and identity; full regen rolls the dice.\n- **Don't** edit an edit — ever. Not once, not \"just a small one.\" Go back to the master (see the master rule above). Every attempt re-renders the full frame and the degradation is cumulative and permanent.\n- **Don't** keep re-rolling the same failed edit. If three tries off the master didn't land, the brief is wrong, not the model — check in with the user.\n",
-    "slates-model-selection": "---\nname: slates-model-selection\ndescription: Which model to pick for a given job — the routing doctrine. Read BEFORE choosing any video or image model, before quoting a plan, and before defaulting anywhere. Kling 3.0 is the general-purpose video default; Seedance 2.0 is the premium tier for anything where physics, effects, or scale remotely matter; Seedance 2.5 is a SECOND SEAT beside 2.0 (30s takes, 30 references and timestamp control, but no 4K and dearer at every shared resolution — never an upgrade); MiniMax H3 is the AUTHORED-AUDIO seat (three directable sound layers in one pass, declared reference relationships, 480p-4K) with MiniMax H3 Max beside it as a faster 768p-capped, reference-free premium; Veo 3.1 is a narrow niche (native synced audio in one gen, 16:9 or 9:16, 4/6/8s) and never the default.\n---\n\n# Model selection — the routing doctrine\n\nPick the model FIRST, deliberately, before writing a prompt or quoting a plan. Model routing is a core part of the intelligence users are paying for: the agent knows what each model is good at and which ones underperform for a job — defaulting to the wrong model burns the user's credits on a weaker result.\n\n## 🔑 The meta-rule — above the table\n\nThe tables below are a snapshot. This roster churns constantly (NB2 Lite, Omni Flash, Seedream 5 Lite, GPT Image 2 all landed recently) — **a table rots; a rule doesn't.** When the tables and this rule disagree, or when a model appears that the tables don't cover, run the rule:\n\n> **Name ONE must-preserve requirement for the shot.** Not a vibe — the single thing that, if it breaks, makes the shot unusable: this face stays this face · the fluid behaves like fluid · the text stays legible · the take stays one unbroken move.\n>\n> **Inspect the output at its intended crop.** A frame that holds up as a thumbnail can fall apart at the size it will actually be watched. For a location, look at atmosphere, material texture, and anchor objects; for a character, identity, skin, pose, and gradients.\n>\n> **Choose the model that PROVES that requirement** and leaves only failures you can afford to rerun or mask.\n>\n> **When the roster changes, repeat the evidence test.** Do not carry today's ranking forward on reputation.\n\n## Video routing\n\n| Job | Model | Why |\n|---|---|---|\n| **General-purpose — the default for most shots** | **Kling 3.0 std** | Cost-effective workhorse. Strong image-to-video: preserves identity, layout, and text from the start frame. 16:9 / 9:16 / 1:1, 3–15s. |\n| Higher visual polish, no physics demands | Kling 3.0 pro | Mid-price fidelity bump on the same strengths. |\n| Multi-character dialogue / audio co-generation | Kling 3.0 omni | Dialogue syntax, voice direction, language codes, `@element` refs. |\n| **Anything with remotely important physics** — effects, destruction, water/fire/smoke/cloth, creature motion, scale, complex simultaneous action | **Seedance 2.0** | The premium tier. Physics and effects are its whole edge; up to 9 ingredient refs, first+last frame, native 4K (4K video is Pro-only). |\n| The premium hero shot a piece hangs on | Seedance 2.0 | Spend where it shows. |\n| **One take longer than 15 seconds**, or a shot needing more than 9 image references, or an AUDIO-ONLY reference, or **beats that have to land at a named second** | **Seedance 2.5** | A SECOND SEAT beside 2.0, never an upgrade: 4–30s in one take, 30 image + 10 video + 10 audio references, audio-only refs, and the only Seedance seat that **acts on timestamps** (rules in `slates-prompting-seedance-2-5` § Timestamps) — 480p / 720p / 1080p, **no 4K**, and **dearer than 2.0 at every shared resolution** (720p $0.231/s vs $0.15/s, +54%). If you want 4K, or the same resolution cheaper, stay on 2.0. 🚨 Two live hazards: (a) with references attached, the words *add / remove / replace / change / extend / continue* make it reclassify the request as a video EDIT and fail AFTER the job queues — describe the finished frame, or use `seedance-2.5-edit`; (b) LENGTH is the price dial, not resolution — a 30s 720p face gen is 489 credits and a 30s 1080p faceless gen is 614, against a 1,000-credit welcome grant. Quote before any take over ~10s. |\n| **The SOUND has to be directed, not just present** — a specific line delivered a specific way, scene sound that has to sit under it, and score that must stay out of the characters' world | **MiniMax H3** | The only seat where audio is authored in three separate layers in ONE pass (synchronised events in the body, ambience in a soundscape section, audience-only score in its own) rather than toggled on. 5–15s, 480p / 768p / 2K / 4K, 24fps, 32kHz stereo, 11 languages. Rules in `slates-prompting-minimax-h3`. |\n| **A reference has to keep a DECLARED amount of itself** — especially moving one subject's characteristic onto a *different* subject | **MiniMax H3** | The only seat that understands a stated retention relationship (kept whole / kept in part / transferred onto another subject / loose echo). 9 images + 3 video + 3 audio, 12 files total. 🚨 The first 5 reference images are free and every one after that costs 4 credits — pass `referenceImages` to `slates_estimate_generation_cost` before a reference-heavy job. |\n| **Turnaround is the requirement** on a text-to-video or start-frame shot at 480p/768p | **MiniMax H3 Max** | fal's self-hosted post-train of H3. **Measured 2026-08-27: a 5s 768p clip finished in 4.8s against 57s on base H3 — about 12x faster**, same prompt, queue to file. When turnaround is the requirement this is not a marginal win. 🚨 It is the PREMIUM seat, not a cheap H3 — $0.080/s at 768p against base H3's $0.060/s, 33% more, and it tops out at 768p. It still animates a start frame and an end frame — image-to-video is one of the two things it is for — but it has no reference-to-video endpoint, so the omni-reference set (9 images + video + audio) is base-H3 only. Never the default; never reach for it to save money. |\n| Native synchronized audio (dialogue + SFX generated WITH the video in one gen), 16:9, ≤8s | Veo 3.1 | Narrow, and now narrower: if the sound needs DIRECTING rather than merely existing, MiniMax H3 is the better seat. |\n\n### Named Seedance escalation triggers\n\n\"Physics matter\" is an abstract category and it under-fires. These are the beats Seedance is **observably** good at — if the shot contains one, escalate without deliberating:\n\n- **Real-time → slow-motion contrast.** The signature beat; nearly every strong clip rides it.\n- **The camera moving while debris, meteors, sparks or particles crash around the subject.** Distinctly a feature of this model, not just a thing it survives.\n- **Massive scale that has to read as genuinely huge** — not \"a big thing\", a thing whose size is the point of the shot.\n- **One continuous unbroken take.**\n\nConcrete beats route better than an abstract category. Cost stays a tiebreaker, never the router (see below).\n\n## Video EDIT routing (changing an existing clip)\n\n| Job | Tool | Why |\n|---|---|---|\n| **Footage-synced VFX on real footage** — add/remove an effect, prop, or lighting change while the take stays the take (incl. talking heads) | **Omni Flash Edit** (`slates_edit_video`, `omni-flash-edit`) | **The edit-fidelity winner** (head-to-head receipt 2026-07-09, WITH a short prompt): lip movement held perfectly, audio near-identical, effect landed and released on cue — where Kling missed an action beat and drifted lips. Prompt-only, 3–10s clips, 720p out, ~6.4 cr/s (cheapest). Quirk: occasional tail jitter / doubled final speech beat — trim the tail on the timeline. Fidelity is EARNED by prompt discipline: one short line + \"Keep everything else the same\"; long prompts destroy it (see below). |\n| **Identity swap needing reference images** — put @marcus into the clip, lock a style from refs | **Kling O3 Edit** (`slates_edit_video`) | The only edit engine that takes element/style reference images (frontal + angles lock identity). ~19¢/s. |\n| **Spoken words must be bit-exact** (VO, legal copy, music) | **Kling O3 Edit** with `keepAudio` (default true) — or segment-splice | Kling keeps the ORIGINAL audio track verbatim — but re-synthesizes the video, so lips can drift slightly against it (7/09 receipt). Omni Flash regenerates audio (voice editing unsupported): on the 7/09 receipt it came back near-identical with perfect lips, but \"near-identical\" is not a guarantee. Zero-risk path for critical audio: segment-splice — edit only the non-talking seconds and keep the original track under the cut. |\n| Style-transfer-heavy re-imagining, full relocate of the scene, or edit quality worth a premium at 1080p+ | Seedance edit/relocate (`videoReferenceAssetId` on `slates_generate_video`) | Seedance's strength is transfer intensity; it re-generates rather than surgically edits. Head-to-head receipt 2026-07-09 (photoreal-insert job, same clip): at 720p it LOST to Omni Flash edit on result while costing ~3× (vref bills input+output seconds; face-lane rates when people are in frame). Route here for its strengths or at 1080p/4K where its ceiling is higher — never as the cheap default. (2.5's relocate lane reaches 1080p too as of 2026-08-24, at $0.2457/s of combined input+output.) Takes long descriptive prompts fine (no Omni-style hard-fail on timing phrasing). |\n| **A clip LONGER THAN 15 SECONDS** | **Seedance 2.5 Edit** (`slates_edit_video`, `seedance-2.5-edit`) | The only edit engine that takes a 4–30s clip — length is the whole reason to route here. 480p/720p/1080p out, native audio, prompt + clip only (no reference images). Output length AND aspect ratio follow the source, so the billed key is the ceiled source length; an edit bills roughly DOUBLE a plain 2.5 generation of the same length because every provider charges an edit on input + output seconds. Set `seedanceFace: true` when a face is visible — the faceless provider blocks faces outright. No consented-real-face route for editing. Inside 15s, choose on fidelity instead. |\n| AI-edit the user's OWN footage | Omni Flash Edit (3–10s), Kling O3 Edit (3–15s, 720–3840px) or Seedance 2.5 Edit (4–30s) | Both take any MP4/MOV — not just Slates gens. Phone footage MUST be rotation-normalized first (players honor the rotation flag; models don't — raw portrait phone clips come back SIDEWAYS). |\n\n- **Edit before re-roll.** A re-roll gambles away the parts the user already likes; an edit changes only what the prompt names. Quote the edit first when a clip is mostly right.\n- **Ship via segment-splice.** Every edit model re-synthesizes the whole clip, so fidelity risk scales with clip length. For real deliverables: trim out ONLY the seconds where the change happens, edit that segment, splice it back over the original on the timeline with the ORIGINAL audio underneath. Most of the final video stays the untouched original — that's how the polished split-screen demos going around actually work, plus gesture-only beats with voiceover laid over in post.\n- **One change per pass, short prompts.** On Omni Flash this is documented law (\"overly descriptive prompts can lead to unintended changes\" — long identity-lock preambles make drift WORSE, receipt 7/09); on Kling multi-beat instructions get dropped. Chain passes instead.\n- Edited clips are themselves editable clips — chain passes; lineage links each output to its parent.\n\n## Motion Transfer & Lip Sync routing (Kling-only tools)\n\nBoth tools are **Kling-only**. Every entry in them is a real Kling endpoint that bolts motion or lip movement onto a finished source as a dedicated post-process.\n\n| Job | Tool | Why |\n|---|---|---|\n| Motion retarget onto a still character | Kling MC std/pro (`slates_generate_motion_transfer`) | Structured skeleton/depth retarget, ~32–42 credits / 5s, takes up to 30s driving clips. |\n| Re-voice a clip, or animate a still portrait | Kling lip-sync / avatar (`slates_generate_lip_sync`) | ~4–29 credits / 5s blocks. |\n\n**Want the Seedance version of either?** It is not a switch on these tools — it is a normal `slates_generate_video` on `seedance-2` with the clip attached as a **video reference** and the motion or dialogue written into the prompt (\"the character from image 1 performs the exact motion from video 1\"). That routes to the same endpoint the tool would have called, with the prompt visible and editable instead of ghost-written. Single-pass conditioning genuinely beats post-hoc retargeting on fast choreography, contact, cloth and hair — and it carries native audio — so escalate there whenever fidelity matters.\n\n- Seedance video-reference gens bill COMBINED input+output seconds (`seedance-2*-vref-*` keys) — pass the clip duration and quote before confirming. Driving clips must be 2–15s on Seedance 2.0 and up to 30s on 2.5; past that it is Kling MC's lane.\n- Faces on that route go through the normal cascade: `seedanceFace` for a character, `[REAL_FACE_DETECTED]` → `seedanceRealFace` + `realFaceConsent` for a real person (premium realface pricing).\n\n**Rules:**\n\n- **Default video = Kling 3.0 std.** Escalate to Seedance the moment the shot has physics/effects weight or is the hero moment — and say why in the plan (\"physics-heavy, routing to Seedance\").\n- **Veo is never the default.** 16:9 or 9:16 only, 4/6/8s only (and 8s only at 1080p/4K, or with reference images), and it is not the quality pick — treat it as a single-purpose tool for native-synced-audio shots. If audio can be added after (Kling lip-sync, edit stage), prefer Kling or Seedance + audio in post.\n- **9:16 vertical → Kling or Seedance by preference**, not by necessity: Veo does take 9:16 on the route Slates uses. Route away from it because it is the niche seat, not because it can't.\n- **Ratios and durations are enforced before submit.** `slates_generate_video` validates the aspect ratio, resolution and duration against the model you picked and refuses out-of-set values with the legal list — it will not silently ignore or downgrade them. The authoritative per-model sets are in the op's own param descriptions, which are generated from the capability SSOT; prefer those over any list written in prose here.\n- **Image-to-video from an NB2 start frame** (the standard pipeline) → Kling by default, Seedance when the motion is physics-heavy. Not Veo.\n- **User names a model explicitly → use it.** But if it's a mismatch for the job (crazy physics on Kling std, a 30s take on anything but Seedance 2.5, 4K on Seedance 2.5 which has none), say so in one line and offer the right route before generating.\n\n## Image routing\n\n**Video models (Kling, Seedance, Veo) cannot generate standalone images — ever.** A \"premium hero reference image\" is still an image job: it routes to an image model below, never to Seedance.\n\n- **Default: Nano Banana 2** — best reference handling (14 refs), best legible text, the standard start-frame generator.\n- **NB2 Lite** — the fast/draft seat: ~half NB2's price, ~2.7× faster, 1K only. Route iteration volume and drafts here; finals go back to NB2 full (2K/4K).\n- **Nano Banana Pro** — the hero-frame/typography ceiling (~2× NB2). NB2 ≈ 95% of Pro; escalate only when spatial composition, cinematic lighting/skin, fine typography-in-scene, or deep multi-element frames must be perfect. Up to 14 refs — feed it a full subject library.\n- **GPT Image 2** — readable text / panels / UI king: character sheets, shot grids, diagrams, text-bearing panels. **Also the photoreal front-runner (Eric, 2026-08-24)** — at `quality: high` it beat both Nano Banana rails head-to-head on skin realism, which is why the AI-influencer ad lane generates every plate here. Medium is the value seat (half NB2's price at 1080p); **high is the seat for photoreal skin and for text precision**. 4K at both tiers is API-only — even paid ChatGPT can't render it.\n- **FLUX.2 Max** — photoreal texture, hex-color binding, typography, less censored.\n- **Seedream 5 Lite** — uncensored + any-resolution flat price; volume exploration when the Gemini filter is in the way.\n\n**Split rule of thumb:** readable text / panels / UI **and photoreal people** → GPT Image 2 (`high` for photoreal); edit-heavy work, or anything needing the 14-reference ceiling → the Banana line; drafts → NB2 Lite; uncensored or odd resolutions → Seedream/FLUX.\n\n⚠️ **This line said the opposite until 2026-08-24** — it sent photoreal *away* from GPT Image 2 on reputation, which is the exact failure § The meta-rule above warns about. Re-run the evidence test when the roster moves.\n\n## Audio routing\n\n**Image and video models cannot generate standalone audio, and neither audio model can generate images or video.** A shot that needs synced audio generated WITH the picture is still a video job (Kling omni / Veo / Omni Flash / Seedance all carry native audio); the models below produce audio *as its own asset*, to lay on the timeline.\n\n| Job | Model | Why |\n|---|---|---|\n| **Default — a whole audio scene in one pass**: room tone, ambience beds, crowds, nature, layered dialogue + effects, spoken lines inside a scene | **Seed Audio 1.0** (`seed-audio`) | One plain sentence in, a complete scene out. The continuity-bed workhorse; dialogue is performed inside the room, not cast. |\n| **One named voice saying one line** — a character's own voice, a narrator, a clean VO to lip-sync against | **Inworld TTS-2** (`inworld-tts-2`) | The prompt IS the words, spoken verbatim and billed per character. Voice = the character's clip (cloned for the take), a description, or a preset. No room tone — mix it on the timeline. |\n| **One effect that lands on a known frame**, or a seamless loop | **Sound Effects v2** (`eleven-sfx`) | The only surface with an exact duration control and a real loop mode. |\n\n**There is no music model.** A song is imported (Slates reads audio files and puts them on the timeline), not generated. A line that has to be spoken in a SPECIFIC voice is generated on Inworld TTS-2 and lip-synced against; a line that belongs to a scene is performed by Seed Audio inside it.\n\n### Named audio escalation triggers\n\n- **\"It needs to sound like a place\"** → Seed Audio. Three separate SFX generations layered on the timeline is the wrong shape and costs more.\n- **\"Read this line\"** → Seed Audio, with the line in quotes inside the scene sentence. Re-roll until the take is right, then lip-sync against it.\n- **\"That needs a thump right there\"** → Sound Effects, with the duration set to roughly the length of the event.\n- **\"Give it a track\"** → there is no music generation. Say so and offer to lay an imported track on an audio track.\n\n**Rules:**\n\n- **🚨 Seed Audio has NO duration parameter.** Length comes from the prompt text, so Slates writes the requested duration into the prompt and **bills what you asked for**. Choose the duration deliberately and never write a second, different length into the sentence. Full doctrine: `slates-prompting-seed-audio`.\n- **Kling's audio syntax does not transfer.** `SFX:` / `Ambient noise:` / `Background music:` prefixes are Kling 3.0 *video* prompt syntax. Seed Audio reads them as literal words and the result degrades.\n- **Beds outlast the cut.** Always ask for more seconds than the clip needs so the edit has fade handles — and remember those extra seconds are billed on both surfaces.\n- **Audio inside the video vs audio as an asset.** If the sound must be locked to what happens on screen, generate it with the video (Kling omni / Seedance / Omni Flash / Veo). If it needs to be moved, trimmed, re-used, or layered, generate it here and drop it on an audio track.\n- Per-model prompting: `slates-prompting-seed-audio`, `slates-prompting-elevenlabs`.\n\n## Cost is a tiebreaker, not the router\n\nRoute by capability first, then pick the cheapest tier that serves the job (per `slates-cost-discipline`). Never pick a model because its per-second price looked lowest — a cheap clip that has to be regenerated on the right model costs more than routing correctly once.\n",
+    "slates-model-selection": "---\nname: slates-model-selection\ndescription: Which model to pick for a given job — the routing doctrine. Read BEFORE choosing any video or image model, before quoting a plan, and before defaulting anywhere. Kling 3.0 is the general-purpose video default; Seedance 2.0 is the premium tier for anything where physics, effects, or scale remotely matter; Seedance 2.5 is a SECOND SEAT beside 2.0 (30s takes, 30 references and timestamp control, but no 4K and dearer at every shared resolution — never an upgrade); MiniMax H3 is the AUTHORED-AUDIO seat (three directable sound layers in one pass, declared reference relationships, 480p-4K) with MiniMax H3 Max beside it as a faster 768p-capped, reference-free premium; Veo 3.1 is a narrow niche (native synced audio in one gen, 16:9 or 9:16, 4/6/8s) and never the default.\n---\n\n# Model selection — the routing doctrine\n\nPick the model FIRST, deliberately, before writing a prompt or quoting a plan. Model routing is a core part of the intelligence users are paying for: the agent knows what each model is good at and which ones underperform for a job — defaulting to the wrong model burns the user's credits on a weaker result.\n\n## 🔑 The meta-rule — above the table\n\nThe tables below are a snapshot. This roster churns constantly (NB2 Lite, Omni Flash, Seedream 5 Lite, GPT Image 2.5 all landed recently) — **a table rots; a rule doesn't.** When the tables and this rule disagree, or when a model appears that the tables don't cover, run the rule:\n\n> **Name ONE must-preserve requirement for the shot.** Not a vibe — the single thing that, if it breaks, makes the shot unusable: this face stays this face · the fluid behaves like fluid · the text stays legible · the take stays one unbroken move.\n>\n> **Inspect the output at its intended crop.** A frame that holds up as a thumbnail can fall apart at the size it will actually be watched. For a location, look at atmosphere, material texture, and anchor objects; for a character, identity, skin, pose, and gradients.\n>\n> **Choose the model that PROVES that requirement** and leaves only failures you can afford to rerun or mask.\n>\n> **When the roster changes, repeat the evidence test.** Do not carry today's ranking forward on reputation.\n\n## Video routing\n\n| Job | Model | Why |\n|---|---|---|\n| **General-purpose — the default for most shots** | **Kling 3.0 std** | Cost-effective workhorse. Strong image-to-video: preserves identity, layout, and text from the start frame. 16:9 / 9:16 / 1:1, 3–15s. |\n| Higher visual polish, no physics demands | Kling 3.0 pro | Mid-price fidelity bump on the same strengths. |\n| Multi-character dialogue / audio co-generation | Kling 3.0 omni | Dialogue syntax, voice direction, language codes, `@element` refs. |\n| **Anything with remotely important physics** — effects, destruction, water/fire/smoke/cloth, creature motion, scale, complex simultaneous action | **Seedance 2.0** | The premium tier. Physics and effects are its whole edge; up to 9 ingredient refs, first+last frame, native 4K (4K video is Pro-only). |\n| The premium hero shot a piece hangs on | Seedance 2.0 | Spend where it shows. |\n| **One take longer than 15 seconds**, or a shot needing more than 9 image references, or an AUDIO-ONLY reference, or **beats that have to land at a named second** | **Seedance 2.5** | A SECOND SEAT beside 2.0, never an upgrade: 4–30s in one take, 30 image + 10 video + 10 audio references, audio-only refs, and the only Seedance seat that **acts on timestamps** (rules in `slates-prompting-seedance-2-5` § Timestamps) — 480p / 720p / 1080p, **no 4K**, and **dearer than 2.0 at every shared resolution** (720p $0.231/s vs $0.15/s, +54%). If you want 4K, or the same resolution cheaper, stay on 2.0. 🚨 Two live hazards: (a) with references attached, the words *add / remove / replace / change / extend / continue* make it reclassify the request as a video EDIT and fail AFTER the job queues — describe the finished frame, or use `seedance-2.5-edit`; (b) LENGTH is the price dial, not resolution — a 30s 720p face gen is 489 credits and a 30s 1080p faceless gen is 614, against a 1,000-credit welcome grant. Quote before any take over ~10s. |\n| **The SOUND has to be directed, not just present** — a specific line delivered a specific way, scene sound that has to sit under it, and score that must stay out of the characters' world | **MiniMax H3** | The only seat where audio is authored in three separate layers in ONE pass (synchronised events in the body, ambience in a soundscape section, audience-only score in its own) rather than toggled on. 5–15s, 480p / 768p / 2K / 4K, 24fps, 32kHz stereo, 11 languages. Rules in `slates-prompting-minimax-h3`. |\n| **A reference has to keep a DECLARED amount of itself** — especially moving one subject's characteristic onto a *different* subject | **MiniMax H3** | The only seat that understands a stated retention relationship (kept whole / kept in part / transferred onto another subject / loose echo). 9 images + 3 video + 3 audio, 12 files total. 🚨 The first 5 reference images are free and every one after that costs 4 credits — pass `referenceImages` to `slates_estimate_generation_cost` before a reference-heavy job. |\n| **Turnaround is the requirement** on a text-to-video or start-frame shot at 480p/768p | **MiniMax H3 Max** | fal's self-hosted post-train of H3. **Measured 2026-08-27: a 5s 768p clip finished in 4.8s against 57s on base H3 — about 12x faster**, same prompt, queue to file. When turnaround is the requirement this is not a marginal win. 🚨 It is the PREMIUM seat, not a cheap H3 — $0.080/s at 768p against base H3's $0.060/s, 33% more, and it tops out at 768p. It still animates a start frame and an end frame — image-to-video is one of the two things it is for — but it has no reference-to-video endpoint, so the omni-reference set (9 images + video + audio) is base-H3 only. Never the default; never reach for it to save money. |\n| Native synchronized audio (dialogue + SFX generated WITH the video in one gen), 16:9, ≤8s | Veo 3.1 | Narrow, and now narrower: if the sound needs DIRECTING rather than merely existing, MiniMax H3 is the better seat. |\n\n### Named Seedance escalation triggers\n\n\"Physics matter\" is an abstract category and it under-fires. These are the beats Seedance is **observably** good at — if the shot contains one, escalate without deliberating:\n\n- **Real-time → slow-motion contrast.** The signature beat; nearly every strong clip rides it.\n- **The camera moving while debris, meteors, sparks or particles crash around the subject.** Distinctly a feature of this model, not just a thing it survives.\n- **Massive scale that has to read as genuinely huge** — not \"a big thing\", a thing whose size is the point of the shot.\n- **One continuous unbroken take.**\n\nConcrete beats route better than an abstract category. Cost stays a tiebreaker, never the router (see below).\n\n## Video EDIT routing (changing an existing clip)\n\n| Job | Tool | Why |\n|---|---|---|\n| **Footage-synced VFX on real footage** — add/remove an effect, prop, or lighting change while the take stays the take (incl. talking heads) | **Omni Flash Edit** (`slates_edit_video`, `omni-flash-edit`) | **The edit-fidelity winner** (head-to-head receipt 2026-07-09, WITH a short prompt): lip movement held perfectly, audio near-identical, effect landed and released on cue — where Kling missed an action beat and drifted lips. Prompt-only, 3–10s clips, 720p out, ~6.4 cr/s (cheapest). Quirk: occasional tail jitter / doubled final speech beat — trim the tail on the timeline. Fidelity is EARNED by prompt discipline: one short line + \"Keep everything else the same\"; long prompts destroy it (see below). |\n| **Identity swap needing reference images** — put @marcus into the clip, lock a style from refs | **Kling O3 Edit** (`slates_edit_video`) | The only edit engine that takes element/style reference images (frontal + angles lock identity). ~19¢/s. |\n| **Spoken words must be bit-exact** (VO, legal copy, music) | **Kling O3 Edit** with `keepAudio` (default true) — or segment-splice | Kling keeps the ORIGINAL audio track verbatim — but re-synthesizes the video, so lips can drift slightly against it (7/09 receipt). Omni Flash regenerates audio (voice editing unsupported): on the 7/09 receipt it came back near-identical with perfect lips, but \"near-identical\" is not a guarantee. Zero-risk path for critical audio: segment-splice — edit only the non-talking seconds and keep the original track under the cut. |\n| Style-transfer-heavy re-imagining, full relocate of the scene, or edit quality worth a premium at 1080p+ | Seedance edit/relocate (`videoReferenceAssetId` on `slates_generate_video`) | Seedance's strength is transfer intensity; it re-generates rather than surgically edits. Head-to-head receipt 2026-07-09 (photoreal-insert job, same clip): at 720p it LOST to Omni Flash edit on result while costing ~3× (vref bills input+output seconds; face-lane rates when people are in frame). Route here for its strengths or at 1080p/4K where its ceiling is higher — never as the cheap default. (2.5's relocate lane reaches 1080p too as of 2026-08-24, at $0.2457/s of combined input+output.) Takes long descriptive prompts fine (no Omni-style hard-fail on timing phrasing). |\n| **A clip LONGER THAN 15 SECONDS** | **Seedance 2.5 Edit** (`slates_edit_video`, `seedance-2.5-edit`) | The only edit engine that takes a 4–30s clip — length is the whole reason to route here. 480p/720p/1080p out, native audio, prompt + clip only (no reference images). Output length AND aspect ratio follow the source, so the billed key is the ceiled source length; an edit bills roughly DOUBLE a plain 2.5 generation of the same length because every provider charges an edit on input + output seconds. Set `seedanceFace: true` when a face is visible — the faceless provider blocks faces outright. No consented-real-face route for editing. Inside 15s, choose on fidelity instead. |\n| AI-edit the user's OWN footage | Omni Flash Edit (3–10s), Kling O3 Edit (3–15s, 720–3840px) or Seedance 2.5 Edit (4–30s) | Both take any MP4/MOV — not just Slates gens. Phone footage MUST be rotation-normalized first (players honor the rotation flag; models don't — raw portrait phone clips come back SIDEWAYS). |\n\n- **Edit before re-roll.** A re-roll gambles away the parts the user already likes; an edit changes only what the prompt names. Quote the edit first when a clip is mostly right.\n- **Ship via segment-splice.** Every edit model re-synthesizes the whole clip, so fidelity risk scales with clip length. For real deliverables: trim out ONLY the seconds where the change happens, edit that segment, splice it back over the original on the timeline with the ORIGINAL audio underneath. Most of the final video stays the untouched original — that's how the polished split-screen demos going around actually work, plus gesture-only beats with voiceover laid over in post.\n- **One change per pass, short prompts.** On Omni Flash this is documented law (\"overly descriptive prompts can lead to unintended changes\" — long identity-lock preambles make drift WORSE, receipt 7/09); on Kling multi-beat instructions get dropped. Chain passes instead.\n- Edited clips are themselves editable clips — chain passes; lineage links each output to its parent.\n\n## Motion Transfer & Lip Sync routing (Kling-only tools)\n\nBoth tools are **Kling-only**. Every entry in them is a real Kling endpoint that bolts motion or lip movement onto a finished source as a dedicated post-process.\n\n| Job | Tool | Why |\n|---|---|---|\n| Motion retarget onto a still character | Kling MC std/pro (`slates_generate_motion_transfer`) | Structured skeleton/depth retarget, ~32–42 credits / 5s, takes up to 30s driving clips. |\n| Re-voice a clip, or animate a still portrait | Kling lip-sync / avatar (`slates_generate_lip_sync`) | ~4–29 credits / 5s blocks. |\n\n**Want the Seedance version of either?** It is not a switch on these tools — it is a normal `slates_generate_video` on `seedance-2` with the clip attached as a **video reference** and the motion or dialogue written into the prompt (\"the character from image 1 performs the exact motion from video 1\"). That routes to the same endpoint the tool would have called, with the prompt visible and editable instead of ghost-written. Single-pass conditioning genuinely beats post-hoc retargeting on fast choreography, contact, cloth and hair — and it carries native audio — so escalate there whenever fidelity matters.\n\n- Seedance video-reference gens bill COMBINED input+output seconds (`seedance-2*-vref-*` keys) — pass the clip duration and quote before confirming. Driving clips must be 2–15s on Seedance 2.0 and up to 30s on 2.5; past that it is Kling MC's lane.\n- Faces on that route go through the normal cascade: `seedanceFace` for a character, `[REAL_FACE_DETECTED]` → `seedanceRealFace` + `realFaceConsent` for a real person (premium realface pricing).\n\n**Rules:**\n\n- **Default video = Kling 3.0 std.** Escalate to Seedance the moment the shot has physics/effects weight or is the hero moment — and say why in the plan (\"physics-heavy, routing to Seedance\").\n- **Veo is never the default.** 16:9 or 9:16 only, 4/6/8s only (and 8s only at 1080p/4K, or with reference images), and it is not the quality pick — treat it as a single-purpose tool for native-synced-audio shots. If audio can be added after (Kling lip-sync, edit stage), prefer Kling or Seedance + audio in post.\n- **9:16 vertical → Kling or Seedance by preference**, not by necessity: Veo does take 9:16 on the route Slates uses. Route away from it because it is the niche seat, not because it can't.\n- **Ratios and durations are enforced before submit.** `slates_generate_video` validates the aspect ratio, resolution and duration against the model you picked and refuses out-of-set values with the legal list — it will not silently ignore or downgrade them. The authoritative per-model sets are in the op's own param descriptions, which are generated from the capability SSOT; prefer those over any list written in prose here.\n- **Image-to-video from an NB2 start frame** (the standard pipeline) → Kling by default, Seedance when the motion is physics-heavy. Not Veo.\n- **User names a model explicitly → use it.** But if it's a mismatch for the job (crazy physics on Kling std, a 30s take on anything but Seedance 2.5, 4K on Seedance 2.5 which has none), say so in one line and offer the right route before generating.\n\n## Image routing\n\n**Video models (Kling, Seedance, Veo) cannot generate standalone images — ever.** A \"premium hero reference image\" is still an image job: it routes to an image model below, never to Seedance.\n\n- **Default: Nano Banana 2** — strongest reference HANDLING (14 refs; GPT Image now takes more, at 16, but Banana is still the one that holds many subjects coherently), best legible text, the standard start-frame generator.\n- **NB2 Lite** — the fast/draft seat: ~half NB2's price, ~2.7× faster, 1K only. Route iteration volume and drafts here; finals go back to NB2 full (2K/4K).\n- **Nano Banana Pro** — the hero-frame/typography ceiling (~2× NB2). NB2 ≈ 95% of Pro; escalate only when spatial composition, cinematic lighting/skin, fine typography-in-scene, or deep multi-element frames must be perfect. Up to 14 refs — feed it a full subject library.\n- **GPT Image 2.5** — two seats, `gpt-image-2-5-flare` and `gpt-image-2-5-sunburst`, **same price**. Readable text / panels / UI king: character sheets, shot grids, diagrams, text-bearing panels. **Also the photoreal front-runner (Eric, 2026-08-24)** — it beat both Nano Banana rails head-to-head on skin realism, which is why the AI-influencer ad lane generates every plate on this line. **The seat split is SPEED vs QUALITY, not generate vs edit** (OpenAI's own rule): Flare is the small, fast model with quality *comparable to* GPT Image 2 — drafts, exploration, volume; Sunburst is OpenAI's *most capable* image model, higher quality than GPT Image 2, deliberately slower — finals, hero frames, photoreal, and multi-reference edits, where its lead is widest. **Explore on Flare, finish on Sunburst.** Five quality tiers, cheapest first — `low` (layout checks only), `medium` (drafts), **`high` (the default)**, `xhigh`, `max` (the top). Uneven: `max` is 4× `high`, `xhigh` only ~1.8× it. **16 reference images**, the schema ceiling. **Transparent backgrounds** via `backgroundMode` — free, and the only image family that offers them.\n\n  🚨 **The tier names moved when 2.5 replaced GPT Image 2, and the strings did not.** GPT Image 2's `medium` is 2.5's `high`; its `high` is 2.5's `max` — same money, one rung of renaming. The 2026-08-24 photoreal result was measured at GPT Image 2 `high`, so **the tier that reproduces it is `max`**. Nobody has re-run it on 2.5; the ranking is inherited, not re-measured.\n- **FLUX.2 Max** — photoreal texture, hex-color binding, typography, less censored.\n- **Seedream 5 Lite** — uncensored + any-resolution flat price; volume exploration when the Gemini filter is in the way.\n\n**Split rule of thumb:** readable text / panels / UI → GPT Image 2.5 (Flare to explore, Sunburst to finish); **photoreal people, finals and hero frames → Sunburst at `max`** — the 2026-08-24 result was measured at GPT Image 2's `high`, which is `max` here, and Flare only *matches* GPT Image 2 while Sunburst exceeds it; multi-reference edits where several references must all survive into one frame → Sunburst; edit-heavy work → the Banana line; drafts → GPT Image 2.5 Flare at `medium`, which now undercuts NB2 Lite on both price and resolution; uncensored or odd resolutions → Seedream/FLUX.\n\n⚠️ **This line said the opposite until 2026-08-24** — it sent photoreal *away* from GPT Image on reputation, which is the exact failure § The meta-rule above warns about. Re-run the evidence test when the roster moves. It moved again on 2026-09-09, and the ranking was carried across rather than re-measured — exactly what the meta-rule says not to trust. Treat it as a starting hypothesis for 2.5, not a receipt. **The seat choice above is likewise reasoned from OpenAI's positioning, not measured:** run Flare-`max` against Sunburst-`max` on one plate and write the answer into `slates-prompting-gpt-image-2-5`.\n\n## Audio routing\n\n**Image and video models cannot generate standalone audio, and neither audio model can generate images or video.** A shot that needs synced audio generated WITH the picture is still a video job (Kling omni / Veo / Omni Flash / Seedance all carry native audio); the models below produce audio *as its own asset*, to lay on the timeline.\n\n| Job | Model | Why |\n|---|---|---|\n| **Default — a whole audio scene in one pass**: room tone, ambience beds, crowds, nature, layered dialogue + effects, spoken lines inside a scene | **Seed Audio 1.0** (`seed-audio`) | One plain sentence in, a complete scene out. The continuity-bed workhorse; dialogue is performed inside the room, not cast. |\n| **One named voice saying one line** — a character's own voice, a narrator, a clean VO to lip-sync against | **Inworld TTS-2** (`inworld-tts-2`) | The prompt IS the words, spoken verbatim and billed per character. Voice = the character's clip (cloned for the take), a description, or a preset. No room tone — mix it on the timeline. |\n| **One effect that lands on a known frame**, or a seamless loop | **Sound Effects v2** (`eleven-sfx`) | The only surface with an exact duration control and a real loop mode. |\n\n**There is no music model.** A song is imported (Slates reads audio files and puts them on the timeline), not generated. A line that has to be spoken in a SPECIFIC voice is generated on Inworld TTS-2 and lip-synced against; a line that belongs to a scene is performed by Seed Audio inside it.\n\n### Named audio escalation triggers\n\n- **\"It needs to sound like a place\"** → Seed Audio. Three separate SFX generations layered on the timeline is the wrong shape and costs more.\n- **\"Read this line\"** → Seed Audio, with the line in quotes inside the scene sentence. Re-roll until the take is right, then lip-sync against it.\n- **\"That needs a thump right there\"** → Sound Effects, with the duration set to roughly the length of the event.\n- **\"Give it a track\"** → there is no music generation. Say so and offer to lay an imported track on an audio track.\n\n**Rules:**\n\n- **🚨 Seed Audio has NO duration parameter.** Length comes from the prompt text, so Slates writes the requested duration into the prompt and **bills what you asked for**. Choose the duration deliberately and never write a second, different length into the sentence. Full doctrine: `slates-prompting-seed-audio`.\n- **Kling's audio syntax does not transfer.** `SFX:` / `Ambient noise:` / `Background music:` prefixes are Kling 3.0 *video* prompt syntax. Seed Audio reads them as literal words and the result degrades.\n- **Beds outlast the cut.** Always ask for more seconds than the clip needs so the edit has fade handles — and remember those extra seconds are billed on both surfaces.\n- **Audio inside the video vs audio as an asset.** If the sound must be locked to what happens on screen, generate it with the video (Kling omni / Seedance / Omni Flash / Veo). If it needs to be moved, trimmed, re-used, or layered, generate it here and drop it on an audio track.\n- Per-model prompting: `slates-prompting-seed-audio`, `slates-prompting-elevenlabs`.\n\n## Cost is a tiebreaker, not the router\n\nRoute by capability first, then pick the cheapest tier that serves the job (per `slates-cost-discipline`). Never pick a model because its per-second price looked lowest — a cheap clip that has to be regenerated on the right model costs more than routing correctly once.\n",
     "slates-one-prompt-film": "---\nname: slates-one-prompt-film\ndescription: Use when the user gives ONE idea and wants a finished video out the other end — \"make me a video about X\", \"turn this idea into an ad\", \"make a short film from this\". The full pipeline: script, project, characters, storyboard, frame images, video generation, timeline assembly, MP4 export. This is the master recipe; the other Slates skills are its sub-steps.\n---\n\n# One prompt → finished film — Slates master pipeline\n\nThe user gives an idea. You hand back an MP4 on disk. Everything in between is yours, with exactly TWO mandatory user checkpoints: the creative plan, and ONE aggregated cost approval.\n\n## The pipeline\n\n### 1. Script the beats\nTurn the idea into a beat-level script: 4-10 shots, each with subject, action, setting, camera, and duration (4-8s per shot). Surface it as a tight table. Get the user's nod on the plan, format (aspect ratio — 16:9 vs 9:16 decides everything downstream), and rough budget appetite before touching any op.\n\n🚨 **Before you fire the set, read its variety counts.** `slates_list_shots` returns the distribution with every listing — shot sizes, camera moves, durations, and any bucket repeating three or more times in a row. Read the table as a COLUMN, not as rows: if push-in is the plurality or every row says wide, the batch is wrong before a credit is spent. The craft is `slates-shot-variety`.\n\n**Surface a decision log with the plan.**\n\n<!-- @inject:decision-log -->\nWhen you surface the plan, include a short **decision log** — one line per decision *you* made that the user did not specify **and that no row already records**:\n\n```\nsource phrase or declared default → what you wrote → what it resolves\n\"in a diner\"        → warm, and the light is the reason           → why the anchor was chosen, not what it is\n(no time of day)    → late afternoon, low warm key                 → default; say the word and it changes\n```\n\n🚨 **Keep it to what is NOT already data — and almost everything now IS.** A Shot holds the references and their roles, the model, every param, the shot size, the camera, the prop, the action and the spoken line, and `slates_list_shots` reads the whole board back in order with its variety counts. Narrating any of those is retelling a row the user can open. **Write the Shot, and let the log carry only the judgement no field holds** — why this world, why this light, why this register.\n\n**Hard rule: never silently add weather, props, style, or camera movement.** Four of those are now FIELDS: put the value on the Shot (`prop`, `camera`, `shotSize`, `action`) so the user can read and change it, and put the *reason* in the log only when you invented it rather than being told it. The rule has not softened — it moved from narration into data, which is stronger, because a field can be corrected and a sentence in chat cannot.\n\n> ❌ **Do NOT turn this into a question gate.** Clarifying questions before optimizing directly fight the locked fast-path rule: *if intent is clear, generate immediately with sane defaults, don't ask questions; only ask for production intent, and batch every question into one message.* Log the decisions, then go. The log is an **output**, not an interrogation — surfaced alongside the plan, never as a separate ceremony, and never as a reason to wait.\n<!-- @end:decision-log -->\n\nA 4-10 shot script is where you invent the most on the user's behalf — time of day, wardrobe, weather, lens feel, camera moves the brief never mentioned. The log is what makes those visible while they are still free to change.\n\n### 2. Set up the project\n- `slates_create_project` named for the piece.\n- Recurring character? Build it properly — `slates_create_character` + the `slates-character-identity` recipe — so every frame references the same identity.\n- Recurring location? `slates_create_environment`.\n- One-off shots don't need character/environment records; skip the ceremony.\n\n### 3. Storyboard skeleton and the Shots (no generation yet)\n- `slates_create_storyboard`, `slates_add_scene` per script scene.\n- `slates_create_shot` per beat — the prompt, the model, the params and the references, with the roles they carry. **A Shot needs no image**, so the entire film exists as rows before anything is paid for.\n- `slates_get_shot` reads one back COMPOSED: the prompt the model will actually receive, its numbered references, and its exact quote. Audit your own work there — you cannot approve something the request will not contain.\n- Structure first, spend second — the user catches script problems on the free skeleton, not on burned credits.\n\n### 4. ONE aggregated cost approval — then hands-off\nThe Shots ARE the quote. `slates_generate_from_shots` without `confirm` returns one itemised total for the set plus the largest single item — no hand arithmetic, no `slates_estimate_generation_cost` per call:\n\n> Plan: 6 frames at 1k 16:9 + 5 × 8s Kling 3.0 std + 1 × 8s Seedance 2 hero shot ≈ N credits total, largest single N. Proceed with the batch?\n\nPer `slates-cost-discipline` 3b: that single OK authorizes `confirm=true` for **every enumerated call in the batch** — no per-call re-asking. Re-confirm only if a call's price overruns the plan >25% or new calls get added (extra retakes, new shots).\n\n### 5. Generate frame images\nFire the image Shots with `slates_generate_from_shots` (`confirm: true` — step 4 authorized it). Slates names each reference inline as \"image N\"; you never hand-write a role label or a number. Evaluate every result inline against the beat. Bind keepers via `slates_add_frame`, then `slates_update_shot` with `attachFrameId` so the recipe travels with the picture.\n\n**Multi-take where it matters:** for the hook shot and any shot the whole film hangs on, generate 2-4 variants (cheap model or 1k), pull them back with `slates_get_assets_batch`, pick the strongest on composition + identity, discard the rest. Don't multi-take filler shots.\n\n### 6. Generate video per Shot\nFork each bound frame's image Shot with `slates_duplicate_shot` (`model:` the video model — that is the A/B lever the op takes inline), then `slates_update_shot` the copy with `firstFrameAssetId` = the bound frame. Two calls, because `slates_duplicate_shot` forks the prompt, the model and the params; **attachments are changed with `slates_update_shot`.** Then fire the set with `slates_generate_from_shots`.\n\n⚠️ **It runs SEQUENTIALLY and blocks until the last clip lands** — a 6-shot film is one long wait, and it will usually outlast the HTTP timeout while the run keeps going. When that happens, poll `slates_get_shot` for each Shot's `generationIds` and then `slates_get_generation_status`; **never re-fire, that double-spends.** (Concurrent batch firing needs a real queue — concurrency limiting, per-item failure isolation, partial-billing semantics — and is deliberately not built yet.)\n\n**Model mixing — route per `slates-model-selection`** (details in the per-model guides):\n- **Kling V3** (`slates-prompting-kling-v3`): the DEFAULT for most shots — 16:9 / 9:16 / 1:1, 3-15s, strong start-frame adherence; std is the workhorse, Omni for multi-character dialogue.\n- **Seedance 2** (`slates-prompting-seedance`): the PREMIUM tier — any shot where physics/effects/scale remotely matter, plus the hero shot; audio included, first+last frame guidance, native 4K (4K video is Pro-only).\n- **MiniMax H3** (`slates-prompting-minimax-h3`): route here when a shot's SOUND is part of the writing — a line delivered a particular way, scene sound under it, score that must stay outside the characters' world. It authors all three in one pass, which **collapses a shot's audio pass into its video pass** and removes the separate `slates_generate_audio` step for that shot. 5-15s, 480p/768p/2K/4K. Its sibling `minimax-h3-max` is faster but capped at 768p, takes no references, and costs MORE at 768p — a deliberate speed pick, never a saving.\n- **Veo 3.1** (`slates-prompting-veo-3`): niche, never the default — only when native synced audio must generate WITH the video in one gen; 16:9 or 9:16, 4/6/8s (8s only at 1080p/4K or with reference images).\n\nFailed gen? The run continues past it and **nothing is retried automatically**. Read the per-Shot error in the result, fix that Shot with `slates_update_shot`, and re-fire only it (a retry beyond the plan = announce the delta cost).\n\n### 7. Assemble the timeline\n- `slates_get_timeline` once to get the lay of the land.\n- `slates_add_clip_to_timeline` for each completed video asset **in story order** — defaults append back-to-back on the first video track, which is exactly an assembly cut.\n- Order wrong? `slates_reorder_clips` with the full clip-id list. Dropped a shot? `slates_remove_clip`, then reorder to close the gap.\n\n### 8. Export + deliver\n- Output path: ask the user, or default to `<slates_get_project_directory>/exports/<name>.mp4`.\n- `slates_export_video` (absolute path, `.mp4`; blocks while ffmpeg renders — minutes for long timelines).\n- `slates_reveal_file` so the file is literally in front of them.\n- Offer the finishing path: `slates_export_timeline_xml` → DaVinci Resolve (File → Import → Timeline) for grading, sound, and titles.\n\n### 9. Report\nShots delivered, total spent vs. approved plan, the export path, and the single best next lever (\"re-take shot 3 with a tighter prompt\" / \"add a CTA end-card\").\n\n## Hard rules\n\n- **Two checkpoints only.** Creative plan (step 1) and total cost (step 4). Everything else runs without asking — that's the product promise.\n- **Skeleton before spend.** Project + storyboard structure are free; generation isn't.\n- **Look at everything.** Every image inline, every video via `slates_get_asset_video_frames` if a clip seems off. Never assemble a timeline from clips you haven't evaluated.\n- **3-strike rule per shot.** Three failed takes on one shot = stop, show the user what you tried, ask.\n- **Consistency comes from references, not luck.** Same identity asset on every character frame; same environment refs across a location's shots.\n- **Plan in Shots, not in chat.** Every decision that ends up in a sentence you have to remember is a decision the user cannot see, price, fork or re-fire. A Shot is a row: it survives the conversation, and the user can open it in the app and fix one reference without you.\n",
     "slates-previs-blocking": "---\nname: slates-previs-blocking\ndescription: Build a 3D blocking pass in Blender, render it grey-box, and use it as a reference video so the generated shot follows a camera path you designed instead of one the model invented. Use when the user wants precise camera control, a multi-cut sequence, a one-take move, spatial consistency across shots, or says the camera keeps drifting / they keep burning credits re-rolling.\n---\n\n# Previs blocking — design the shot, then generate it\n\nThe spine of the whole workflow. Read this first; the other four previs skills are branches off it.\n\n## The mechanism (why this works at all)\n\nA text prompt asks the model to *invent* camera motion, so it invents differently every roll. You cannot iterate on a variable you do not control, so you re-roll and pay again.\n\nA **reference video** removes the invention. You build the shot in Blender as untextured proxies — a neutral grey set with colour-coded figures, free, instant, deterministic — render the camera's path to mp4, and hand the model that clip alongside the prompt. **Blender locks the motion; the model builds the world.** Iteration moves to the free half, and the paid half usually lands first try.\n\nTwo halves, and keeping them separate is the whole discipline:\n\n| Half | Lives in | Changes when |\n|---|---|---|\n| **Structure** — cuts, camera, timing, who is where | the blocking clip | you re-block |\n| **Style** — what any of it looks like | references + prompt text | you restyle (see `slates-restyle-from-blocking`) |\n\n## Before you start\n\n1. `slates_blender_status` — confirms the bridge is up and returns fps, frame range, existing camera. If it reports `connected: false`, relay its hint and stop; nothing else here works.\n2. Settle **format first**, because the blocking render *is* the film's format: fps, aspect, duration. 24fps is the default and makes cut times land on clean frames. Duration ≤ 30s (seedance-2.5's reference-video ceiling; 15s on the others).\n3. Know the shot count. \"One take\" and \"19 cuts\" are different builds.\n\n## Build order\n\nDo these in order. Each stage is verifiable on its own, and a camera built before the geometry has nothing to frame.\n\n### 1. Set the format\n\n```python\nscene = bpy.context.scene\nscene.render.fps = 24\nscene.render.fps_base = 1.0\nscene.render.resolution_x, scene.render.resolution_y = 1920, 1080\nscene.frame_start, scene.frame_end = 1, 720   # 30s at 24fps\nresult = {\"seconds\": 720 / 24}\n```\n\nFrame maths, stated once so you never redo it in your head: **frame = seconds × fps + 1**. A cut at 7.79s is frame 188.\n\n### 2. Geometry and light — grey set, coded figures, named\n\nProxies only. A person is a box or a capsule with a sphere head. A car is a stretched cube. A can is a cylinder. **The SET is neutral grey — one light, a floor and enough wall that the space reads.** Colour is reserved for the figures, where it carries meaning (below); a grey set is what makes those few colours legible as notation rather than décor. Anything you spend on materials here you pay for twice, because the model repaints every surface anyway.\n\n**Name every object for what it *is* in the story**, not `Cube.003`. The name is how you refer to it later, and it is how you keep your own timeline honest.\n\nTwo conventions that cost nothing now and save a re-roll later:\n\n- **Colour is identity.** Give each character a distinct viewport colour and *write the mapping down* — `red = the boss, green = the kid, blue = the driver`. The generation prompt will restate that mapping so the model knows which grey body is which person across cuts. Without it, characters swap.\n- **Encode facing on featureless proxies.** A box has no front. Mark one face red, the back black, the sides green, and say so in the prompt: `RED face = the direction he faces`. Otherwise the model guesses which way people are looking.\n- **Checker a surface when SCALE or SPEED has to read.** Flat grey gives a model no parallax cue, so a fast move over a featureless floor reads as slow, and a big room reads as a small one. A black-and-white checker on the ground (or the wall a camera races past) gives it something to measure against. ⚠️ **Build it as GEOMETRY, never as a Checker Texture node.** The blocking render is Workbench, which draws one flat colour per material and never evaluates a shader node tree — a `TEX_CHECKER` comes out flat grey and you lose the cue without being told. Subdivide the plane and alternate `material_index` per face. Like every other colour here it is notation, so it goes in the translation list and gets dressed over.\n\n```python\n# Two materials, alternated per face. `TILE` is the square size in metres.\ndark = bpy.data.materials.new(\"Checker_Dark\")\ndark.diffuse_color = (0.05, 0.05, 0.05, 1.0)\nlight = bpy.data.materials.new(\"Checker_Light\")\nlight.diffuse_color = (0.80, 0.80, 0.80, 1.0)\nfloor.data.materials.append(dark)    # material_index 0\nfloor.data.materials.append(light)   # material_index 1\n# Subdivide first (edit mode or a Subdivide modifier applied) so there ARE\n# faces to alternate — a 2-triangle plane can only ever be one colour.\nfor face in floor.data.polygons:\n    cx, cy = face.center.x, face.center.y\n    face.material_index = (int(cx // TILE) + int(cy // TILE)) % 2\n```\n\nAnd the identity colour on each proxy:\n\n```python\nmat = bpy.data.materials.new(\"ID_Red\")\nmat.diffuse_color = (0.8, 0.1, 0.1, 1.0)   # what the blocking render draws\nobj.data.materials.append(mat)\nobj.color = (0.8, 0.1, 0.1, 1.0)           # same value, for viewport parity\n```\n\nThe blocking render pins Workbench to `MATERIAL` shading, so **`mat.diffuse_color` is the value that reaches the clip** — and an object with no material at all falls back to a neutral grey, which is why an unpainted set still reads correctly. Set `obj.color` to the same value anyway: it costs one line, it makes the user's viewport match what renders, and keeping the two equal means you never have to remember which one is authoritative.\n\n### 3. Camera\n\nThe whole of `slates-camera-language`. Build the rig, then keyframe it. Then **read back what you built** with `slates_blender_scene` — its `cutSeconds` is your cut list, and it is the number you will write timings against. That field is the authoritative one on EITHER rig — marker frames when cameras are bound to markers, the active camera's own keyframes when they are not. `camera.keyframeSeconds` is empty on a marker-bound edit, which is the rig `slates-camera-language` recommends for anything past a handful of cuts.\n\n### 4. Handheld, last\n\nAdd it after the moves are right, never before — noise on top of a wrong path just hides the wrong path.\n\n### 5. Verify the cuts\n\nThe one check that catches the most damage: on a multi-cut blocking, camera position, target and focal length must all change **exactly on the cut frame, with no transition frame between**. One interpolated frame reads as a whip-pan the model will faithfully reproduce.\n\n```python\n# Every camera f-curve keyframe on a cut frame must be CONSTANT out of the\n# previous key, or the cut smears.\nfor fc in cam.animation_data.action.fcurves:\n    for kp in fc.keyframe_points:\n        if int(kp.co[0]) in CUT_FRAMES:\n            kp.interpolation = 'CONSTANT'\n```\n\nAlso check nothing interpenetrates — proxies through floors, clones through the hero object, letters through each other. The model renders intersections as faithfully as it renders everything else.\n\n### 6. Save a backup after every stage\n\nCheap, and blocking is iterative by nature.\n\n```python\nbpy.ops.wm.save_as_mainfile(filepath=path, copy=True)\n```\n\n## Render and generate\n\n```\nslates_blender_render_blocking { projectId, fps: 24 }\n```\n\nRenders the **scene camera** through scene settings — never the user's viewport, so the result does not depend on where they left their mouse — imports the mp4 into the project, and returns `assetId` + `durationSeconds`.\n\nThen:\n\n```\nslates_generate_video {\n  model: \"seedance-2.5\",\n  videoReferenceAssetIds: [<the blocking asset>],\n  videoReferenceSecondsEach: [<durationSeconds>],\n  characterAssetIds: [...], environmentAssetIds: [...], styleAssetIds: [...],\n  prompt: <written per slates-blocking-to-prompt>\n}\n```\n\n**Four inputs, and that is the entire stack:** a character sheet each, one location/style reference, the blocking clip, and a prompt written against the blocking. Resist adding a fifth.\n\nModel note: seedance-2.5 is the seat for this — 10 reference videos at up to 30s each. seedance-2 and minimax-h3 take 3 at 15s. Route per `slates-model-selection`.\n\n## Leaving holes on purpose\n\nWhere the model outperforms any blockout you could build — liquid, smoke, fire, cloth — **block a black gap instead** and say so in the prompt: `CUT 7 (14.5-17.0, black gap in the reference)`. You are reserving a slot, not forgetting one.\n\n## What not to do\n\n- **Don't texture, light or material the blocking.** Grey is the specification. The reference supplies motion; the references supply look.\n- **Don't animate what you don't need.** Heads especially — a proxy head turning wrong is worse than one that never turns.\n- **Don't build the camera before the geometry.** It has nothing to aim at, and every value you set gets redone.\n- **Don't skip reading the scene back.** Write timings from `slates_blender_scene`'s `cutSeconds`, never from what you intended to build.\n- **Don't exceed the model's reference-video ceiling.** A 40s blocking against a 30s cap silently truncates.\n\n## Related\n\n`slates-camera-language` (rigs and moves) · `slates-blocking-to-prompt` (writing the prompt against the clip) · `slates-dialogue-blocking` (multi-character continuity) · `slates-restyle-from-blocking` (one blocking, many worlds) · `slates-model-selection` (routing)\n",
     "slates-project-organization": "---\nname: slates-project-organization\ndescription: Use when the user names an asset by code (\"use IMG-A36\"), asks what a code means, or is organizing or navigating a project. Covers the asset short-code system (IMG-A12 / VID-V3 / AUD-S1 badges on every gallery card), folders for film STRUCTURE, and the typed tabs for reusable references.\n---\n\n# Organizing a Slates project\n\nSlates already gives each REUSABLE reference type its own home — the **Characters**, **Environments**, and **Styles** tabs, each with its own generation + `@mention`/`#ref` behavior. Do NOT recreate those as folders. Folders are for **structure**, never type.\n\n**Folders = where an asset sits in the FILM**, and they mirror to real subfolders on disk (`projects/<id>/…`), so a human can open the project in Resolve/Finder and navigate it like an edit. Use them for work product, not references.\n\nCreate with `slates_create_folder`; file assets with `slates_move_assets_to_folder`. Generations land in the project's active folder, so set it before a batch.\n\nConventions by project type:\n- **Short film / narrative:** `Shots` (scene stills) · `Clips` (generated video) · `Final` (the export). Use one folder per scene (`Scene 1`, `Scene 2`, …) instead when the piece has distinct locations/beats.\n- **Ad / UGC:** `Hooks` · `B-roll` · `Talking-head` · `Final`.\n\nRules of thumb:\n- Reusable cast / sets / look → leave in the Characters/Environments/Styles tabs. Don't fold them.\n- Scene stills, clips, and the final cut → file into the structural folder they belong to, as you make them.\n- One folder per asset (folders are structure). Cross-cutting status (hero take, reject, variant) is a tag concern, not a folder.\n- Keep the gallery legible: work product lives in folders; the reference scaffolding (sheets, plates, style images) stays in its tabs.\n\n## Asset codes — the shared vocabulary (IMG-A12 / VID-V3 / AUD-S1)\n\nEvery asset gets a short, stable code the moment it lands in a project, and the user sees it as the badge in the **top-left corner of every image and video card** in the gallery. This is the shared vocabulary between you and the user — it exists so neither of you ever has to quote a UUID.\n\n**The scheme:**\n- `IMG-A{n}` = images · `VID-V{n}` = videos · `AUD-S{n}` = audio.\n- Numbering is **per project, per type**, counts up from 1, and **numbers are never reused** — deleting IMG-A12 doesn't renumber anything, so a code always means the same asset forever.\n- Each asset also carries a **label**: the first ~4 meaningful words of its prompt, title-cased. Chat format is code + label: `IMG-A12 — Beach Sunset`.\n\n**How to use it:**\n- **User names a code** (\"use IMG-A36 as the reference\", \"animate VID-V3's last frame\") → resolve it via `slates_list_assets` (match the `code` field) to get the assetId, confirm back in the same vocabulary: \"Got it — IMG-A36 — Marcus Rooftop Close-Up as the first frame.\"\n- **You name assets** → ALWAYS code + label, never UUID, never \"the beach one\" (which of three?). The user matches your words to the badge by eye.\n- **User seems confused** about what a code is or how to point you at an image → explain it in one line: \"Every image and video in your gallery has a code badge in its top-left corner — like IMG-A36. Just say that code and I'll know exactly which one you mean.\"\n- **Ambiguity** (\"the sunset image\" when several exist) → pull candidates with `slates_get_assets_batch` and offer the codes: \"I see IMG-A12, IMG-A19, and IMG-A24 with sunsets — which one?\"\n",
     "slates-prompting-elevenlabs": "---\nname: slates-prompting-elevenlabs\ndescription: How to prompt ElevenLabs Sound Effects v2 in Slates. Read before calling slates_generate_audio with model eleven-sfx — ONE short effect with an EXACT duration, or a seamless loop, billed per second. Covers describing an effect by its physical cause, the one-sound-per-generation rule, picking a duration, loops, prompt_influence, and when to use Seed Audio instead.\n---\n\n# ElevenLabs Sound Effects v2 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — ElevenLabs Sound Effects v2.** ONE short sound with an exact length, or a seamless loop. The only Slates audio surface with a real duration control and a real loop mode.\n\n**The five levers**\n1. **Describe the physical CAUSE, not the label** — `heavy oak door slams shut`, `boot scuffs on grit`, `a latch drops home`.\n2. **Name the material and the space.** The material decides the timbre and the space decides the tail: `on wet concrete`, `in a tiled stairwell`, `across an empty warehouse`.\n3. **One sound per generation.** A room with dialogue AND clatter AND ambience is one Seed Audio pass, not three effects.\n4. **Pick the duration from the cut**, not from a feeling: roughly 0.5-1s for an `impact`, 2-4s for a `whoosh`, 8-22s for a `loopable bed`.\n5. **Ask for a loop explicitly** — `seamless loop` — when the sound has to lie under a whole scene, and keep it featureless enough to survive the seam.\n\n**Examples**\n- `A heavy oak door slams shut in a stone hallway, brief reverberant tail.` (1.5s)\n- `Steady rain on a tin awning, no thunder, no wind gusts, seamless loop.` (18s)\n\n**Hard constraint:** it is billed per second and the duration is never left for the model to pick — that would make the charge non-deterministic. It is NOT a speech surface: a line in a specific voice is `inworld-tts-2`, and dialogue inside a scene is Seed Audio, which casts and performs the line in the room.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** — a label is not a sound; describe the physical cause:\n- `door sound`, `whoosh`, `footsteps`, `impact`, `ambience` standing alone\n<!-- @banned:end -->\n\nOne short sound with an exact length, carried on fal (`fal-ai/elevenlabs/sound-effects/v2`). This is the only Slates audio surface with a real duration control and a real loop mode.\n\n## Where it routes\n\n- **A single hit that has to land on a known frame** — door slam, whoosh, impact, UI blip, riser.\n- **A seamless loop** you can lay under a whole scene — rain, engine hum, crowd murmur, machine noise.\n- **NOT** layered scenes. A room with dialogue *and* clatter *and* ambience is one `seed-audio` pass, not three SFX generations.\n- **NOT** speech. Dialogue, narration and scratch VO are `seed-audio` — it casts and performs the line inside the scene.\n- **AUDIO-ONLY.** It cannot produce images or video.\n\n## THE RULES\n\n### 1. Describe the physical CAUSE, not the label\n\n```\n✗  door sound\n✓  heavy oak door slams shut in a stone hallway\n\n✗  whoosh\n✓  a thick rope swung fast past a microphone, low air displacement\n\n✗  footsteps\n✓  boots on wet gravel, slow, one person\n```\n\nMaterial + weight + surface + room. Naming all four is the difference between a usable effect and a stock-library shrug. Cap is 450 characters — you will not need them.\n\n### 2. One sound per generation\n\nThis surface makes a single event. A door, then footsteps, then a siren is three generations layered on the timeline — or one `seed-audio` scene, which is usually cheaper and always more coherent.\n\n### 3. Duration is always explicit, and it is the price\n\nSlates **always sends** `durationSeconds`. (Left null the model picks, which makes the charge non-deterministic — so it is never left null.) The window it must fall in:\n\n<!-- @inject:thresholds -->\n<!-- GENERATED from @slatesvideo/shared — do not edit between the markers.\n     Source: CONFIRM_CREDITS, DEVIATION_FACTOR and the audio bounds in\n     packages/shared/src/operations/index.ts. Every number here is REFUSED by an\n     op when a prompt gets it wrong, which is why none of them is typed by hand\n     any more: this block replaced four claims that contradicted the code. -->\n\n**The thresholds, from the code that enforces them:**\n\n- **Confirm gate:** above **17 credits** an op returns `requires_confirm` and will not\n  proceed until you re-call with `confirm: true`. Below it, announce the cost once and go.\n- **Deviation pause:** the desktop Studio Agent stops and re-asks when projected generation spend\n  exceeds the approved plan by more than **20%**. You do not trigger this; the app does.\n- **Seed Audio duration:** **3–120 seconds.** There is no duration\n  parameter on the model — the number you pass is written into the prompt AND is what the user is\n  billed. Outside that range the op refuses rather than clamping.\n- **Sound Effects duration:** **1–22 seconds**, billed per second, never left for the\n  model to pick.\n\nNever quote a credit figure from memory: `slates_estimate_generation_cost` returns the real one.\n<!-- @end:thresholds -->\n\n| Kind of sound | Ask for |\n|---|---|\n| impact, hit, click | 0.5–1s |\n| whoosh, riser, transition | 2–4s |\n| loopable bed | 8–22s + `loop: true` |\n\nOver-asking pads the tail with room tone you then trim. Under-asking clips the decay.\n\n### 4. Loops\n\n`loop: true` tiles without a seam — rain, engine hum, crowd murmur, machine noise. Combine with a longer duration so the loop point is not obvious.\n\nFor a bed longer than 22s, this is the wrong surface: `seed-audio` runs to 120s in one pass.\n\n### 5. Prompt influence\n\n`promptInfluence` 0–1, default 0.3. Higher hugs your wording with less variation between takes; lower explores. Raise it when a re-roll keeps wandering off the brief; lower it when every take sounds like the same take.\n\n## Iterating\n\n- Re-rolls that keep missing = the prompt named a **label** instead of a **cause**. Rewrite it as a physical event.\n- A hit that lands but sounds wrong in the scene is usually a *room* problem — name the space (\"in a stone hallway\", \"in a padded studio\", \"outdoors, no reflections\").\n- Three failed takes means the prompt is wrong, not the seed.\n\n## Content notes\n\nElevenLabs applies its own moderation. See slates-content-policy.\n",
     "slates-prompting-flux-2-max": "---\nname: slates-prompting-flux-2-max\ndescription: How to prompt FLUX.2 Max (Black Forest Labs image model). Read before calling slates_generate_image with model flux-2-max, or slates_edit_image with editModel flux-2-max. FLUX.2 wants front-loaded structure, real camera vocabulary, and positive-only phrasing — no negative prompts, no tag soup.\n---\n\n# FLUX.2 Max — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — FLUX.2 Max.** Word order is weight: it attends hardest to the start. Structure: `Subject + Action + Style + Context`, then secondary detail. Length 10-30 words for a concept test, 30-80 for most work, 80+ only for a genuinely complex scene.\n\n**The five levers**\n1. **Front-load the subject and the one action.** Anything after the first clause is a modifier, and it is read as one.\n2. **Name real gear** — `Shot on Hasselblad X2D, 80mm, f/2.8, natural light`, `Kodak Portra 400, natural grain`. This is the single biggest realism lever.\n3. **Era cues as a package** — `early digital camera, slight noise, flash photography, candid` reads 2000s; `film grain, warm cast, soft focus` reads 80s.\n4. **Bind every hex colour to an object.** `a #1B4D3E enamel mug` lands; an unbound colour does not.\n5. **For portraits add texture words** — `natural skin texture, realistic pores, subtle imperfections, soft diffused lighting`.\n\n**Examples**\n- `A chef plating in a steel kitchen pass. Shot on Hasselblad X2D, 80mm, f/2.8. Overhead fluorescents plus warm spill from the line. Natural skin texture, subtle imperfections. Muted steel and #7A3B2E copper.`\n- `An empty municipal pool at dusk, 35mm, deep focus, early digital camera with slight noise and flash falloff. Cracked #4A7C8C tiles. Candid, unstaged.`\n\n**Hard constraint:** no negative prompting. Every \"no X\" must be rewritten as the positive state — `no blur` becomes `sharp focus throughout`, `no people` becomes `empty scene`, `no harsh shadows` becomes `soft, diffused lighting`.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** — there is no negative prompting, so each of these has a positive form:\n- `no blur` (say `sharp focus throughout`), `no people` (say `empty scene`), `no harsh shadows` (say `soft, diffused lighting`)\n- an unbound hex colour — bind it to an object or it lands inconsistently\n- `masterpiece`, `best quality`, `trending on artstation`, `8k`\n<!-- @banned:end -->\n\nBlack Forest Labs' top image model, routed via fal.ai. In Slates: `slates_generate_image` with `model: flux-2-max` (REQUIRES projectId — no headless path), priced per resolution (1k/2k/4k — call `slates_estimate_generation_cost` for current numbers, never quote from memory). Strengths vs Nano Banana 2: photoreal texture, less censored, precise hex-color control, strong typography. Reference images route through FLUX's edit endpoint and carry a lower per-model cap than NB2's 14.\n\n## Core structure — front-load what matters\n\n```\nSubject + Action + Style + Context\n```\n\nWord order is weight. FLUX.2 attends hardest to the start of the prompt: main subject → key action → critical style → essential context → secondary details.\n\n**Length:** 10-30 words for concept tests, 30-80 words for most work, 80+ only for genuinely complex scenes.\n\n## Photorealism: name real gear, not \"professional photo\"\n\nThe single biggest realism lever is concrete camera vocabulary:\n\n```\nShot on Hasselblad X2D, 80mm lens, f/2.8, natural lighting\nShot on Sony A7IV, 35mm, golden hour, shallow depth of field\nKodak Portra 400, natural grain, organic colors\n```\n\nEra cues work the same way: \"early digital camera, slight noise, flash photography, candid\" reads 2000s digicam; \"film grain, warm color cast, soft focus\" reads 80s.\n\nFor portraits add: natural skin texture, realistic pores, subtle imperfections, soft diffused lighting.\n\n## No negative prompts — reframe positively\n\nFLUX.2 has no negative prompt support. Describe the presence you want, not the absence:\n\n- ❌ \"no blur\" → ✅ \"sharp focus throughout\"\n- ❌ \"no people\" → ✅ \"empty scene\"\n- ❌ \"no harsh shadows\" → ✅ \"soft, diffused lighting\"\n\n## Hex colors — bind them to objects\n\nFLUX.2 matches hex codes, but only when each code is attached to a specific object:\n\n```\nwalls in hex #C4725A, sofa in #1B6B6F, accent pillows #E8A847\ngradient starting with color #02eb3c and finishing with color #edfa3c\n```\n\n❌ \"use #FF0000 somewhere\" — unbound colors land inconsistently.\n\n## Text rendering\n\nQuote the exact text, then place and style it:\n\n```\nThe text 'OPEN' appears in red neon letters above the door\nLogo text 'ACME' in color #FF5733, ultra-bold decorative serif, centered\n```\n\nSpecify placement relative to other elements, font family feel (serif / sans / script), and relative size (\"large headline,\" \"small body copy\").\n\n## JSON prompting for production work\n\nFor multi-element scenes that must come out exactly right (product shots, infographics, brand work), FLUX.2 parses structured JSON prompts:\n\n```json\n{\n  \"scene\": \"Professional studio product photography on polished concrete\",\n  \"subjects\": [{ \"description\": \"matte black ceramic mug with steam\", \"position\": \"center foreground\" }],\n  \"style\": \"commercial product photography\",\n  \"color_palette\": [\"#1B1B1B\", \"#E8A847\"],\n  \"lighting\": \"three-point softbox, soft diffused highlights\",\n  \"camera\": { \"lens-mm\": 85, \"f-number\": \"f/5.6\" }\n}\n```\n\nUse natural language for exploration, JSON when the layout is locked and you're matching a spec.\n\n## Reference images (edit path)\n\nIn Slates, pass `referenceAssetIds` on `slates_generate_image` — FLUX routes them through its edit endpoint. Slates names each reference inline in the prompt (\"the subject (image 1), the style (image 2)\") in the order it sends them, so you don't hand-write role labels; the name carries the role and unnamed-by-position blending is avoided. For surgical changes to one existing image use `slates_edit_image` with `editModel: flux-2-max` (note: FLUX edits ignore extra referenceAssetIds — that's NB2-only).\n\n### Reference rules (the verified ones)\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### For FLUX.2 Max specifically\n\n- **FLUX caps references well below NB2's 14, so rule 1's \"2-4\" is a ceiling here, not a starting point.** Be deliberate about which roles earn a slot.\n- **Rule 9 has a hard edge on this model:** `slates_edit_image` with `editModel: flux-2-max` ignores extra `referenceAssetIds` — that is NB2-only. A FLUX edit sees the source image and the prompt, nothing else.\n- **FLUX has no memory between generations, so rule 7 is enforced by repetition.** Define the character exhaustively once and repeat those exact descriptors verbatim in every subsequent prompt — see Character consistency across a series below.\n\n## Character consistency across a series\n\nDefine the character exhaustively once, then repeat those exact descriptors verbatim in every subsequent prompt. FLUX has no memory between generations — the repeated description IS the consistency mechanism.\n\n## Common failure modes + fixes\n\n| Failure | Fix |\n|---|---|\n| Generic \"AI look\" on photoreal | Name a camera body + lens + f-stop instead of \"professional photo\" |\n| Colors drift from brand spec | Bind each hex code to a named object |\n| Text garbled | Quote the exact string, specify font feel + placement + size |\n| Multi-reference blend chaos | Name each reference inline (Slates does this from your @mentions/referenceAssetIds) — the same name for one entity, distinct names per role |\n| Wanted element missing | Move it earlier in the prompt — order is weight |\n\n## Pre-flight: references arrive inline, refer by code\n\nWhen you pass `referenceAssetIds`, the first call returns the references **inline as image content blocks** with a cost estimate and `requires_confirm: true`. Look at them — revise the prompt if they suggest a different composition or style — then re-call with `confirm=true`. Refer to each asset by its short code (`IMG-A12 — Beach Sunset`) when talking to the user; it matches the badge on their gallery thumbnail.\n\n## Sources\n\n- [Black Forest Labs — FLUX.2 Prompting Guide](https://docs.bfl.ml/guides/prompting_guide_flux2)\n- [fal.ai — FLUX.2 [max] Prompt Guide](https://fal.ai/learn/devs/flux-2-max-prompt-guide)\n",
-    "slates-prompting-gpt-image-2": "---\nname: slates-prompting-gpt-image-2\ndescription: Prompting GPT Image 2 — the readable-text / character-sheet / shot-grid engine AND the current photoreal front-runner. Read before calling slates_generate_image with model gpt-image-2. Covers the quality tiers (medium default, high for max text precision), resolution classes (1k/2k=1080p/3k=1440p/4k), text-accuracy prompting, panel/grid layout direction, and when to route to the Banana line instead.\n---\n\n# GPT Image 2 — sheets, grids, and text that actually reads\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — GPT Image 2.** The readable-text, ordered-panel and exact-placement engine, and (measured 2026-08-24 at `quality: high`) the photoreal front-runner for people. Structure: subject and action, then the exact copy in quotes, then layout, then light.\n\n**The five levers**\n1. **Quote every string that must render verbatim** — `the sign reads \"OPEN 24 HOURS\"`. Quoted strings render most reliably.\n2. **Font FEEL, never a font name** — `clean geometric sans, high contrast`, `hand-painted brush lettering`.\n3. **Order dense copy explicitly** — `Line 1: \"...\" Line 2: \"...\"`. It respects the ordering.\n4. **Name the layout as a grid** for sheets and panels — `a 3x2 grid of panels, reading left to right, equal gutters`.\n5. **Set `quality` deliberately.** `medium` is sharp text at a quarter of the price and is within a hair of `high` in blind tests; reach for `high` only when tiny type, dense diagrams or many labelled elements ARE the job.\n\n**Examples**\n- `A 2x3 character turnaround sheet on a neutral grey field, equal gutters, reading left to right: front, three-quarter, profile, back, three-quarter back, top. One woman, mid-30s, cropped dark hair, olive field jacket. Flat even studio light, no cast shadows. Small caption under each panel naming the angle.`\n- `Photoreal portrait, natural window light from camera-left, visible skin texture and pores, 85mm compression. A man in his 50s in a charcoal knit, half-smile, looking just past lens.`\n\n**Hard constraint:** keep total on-image text under about 30 words for perfect accuracy — beyond that it degrades, gracefully but really. It has its own content filter, distinct from Gemini's.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use:**\n- a font NAME — describe the feel (`clean geometric sans, high contrast`) instead\n- a reference role essay (`Reference image 1 is a photograph of a woman. Use that exact woman.`) — name the subject inline instead\n- `8k`, `masterpiece`, `best quality`, `highly detailed` — quality incantations do nothing here either\n<!-- @banned:end -->\n\nGPT Image 2's edge is **character-level text accuracy** (~99% on English), ordered panels, and exact element placement — the jobs where every other model garbles a word or shuffles a layout.\n\n🚨 **It is ALSO the photoreal front-runner, and this file said the opposite until 2026-08-24.** **Receipts:** Eric's direct call, plus a head-to-head on the Higgsfield rail where GPT Image 2 at `quality: high`, 2K beat both Nano Banana rails on skin realism for photoreal people — that result is why the whole AI-influencer ad lane generates its plates here. **Route photoreal to this model, not away from it.**\n\n**What the Banana line still owns:** edit-heavy work and the 14-reference ceiling.\n\n**What would kill this:** a head-to-head at the intended crop going the other way. Per `slates-model-selection` § The meta-rule, re-run the evidence test when the roster changes — never carry a ranking forward on reputation. That rule is exactly what this correction failed.\n\n## Quality tiers — always set explicitly\n\n- **medium** (default) — sharp text, fast, the value seat: half NB2's price at the 1080p class. Blind benchmarks put it within a hair of high at a quarter of the cost. Start here.\n- **high** — ~4× the price; max text precision + reasoning. A deliberate premium pick when tiny type, dense diagrams, or many labeled elements ARE the job.\n\nNever rely on the provider default (it's high — the priciest tier). The Slates ops send medium unless you say otherwise.\n\n## Resolution classes\n\n`1k` = 1024²-class · `2k` = 1920×1080-class · `3k` = 2560×1440-class · `4k` = 3840×2160-class. Pick 2k for most sheets/panels; 4k for print-density grids. 4K exists at BOTH tiers and is API-only — even paid ChatGPT can't render it.\n\n## Prompting for text accuracy\n\n- **Quote every string that must render verbatim**: `the sign reads \"OPEN 24 HOURS\"` — quoted strings render most reliably.\n- Specify font *feel*, not font names: \"clean geometric sans, high contrast\", \"hand-painted brush lettering\".\n- For dense text (posters, UI mocks), list the copy as ordered lines: `Line 1: \"...\" Line 2: \"...\"` — GPT Image 2 respects ordering.\n- Keep total on-image text under ~30 words for perfect accuracy; beyond that, accuracy degrades gracefully but degrades.\n\n## Panels, sheets, and grids\n\n- State the grid explicitly and number the cells: \"a 2×3 grid of panels, numbered 1–6, reading left-to-right, top-to-bottom\".\n- Give each cell ONE content clause: \"Panel 3: the character mid-jump, side view\".\n- Character identity sheets: GPT Image 2 holds both the structured panel layout AND photoreal skin, which is why the influencer-ad lane builds its sheets here at `quality: high`, 2K. Reach for NB2/NB Pro when the sheet needs many reference images folded in (14-ref ceiling) or when it is an edit of an existing sheet.\n\n## References & editing\n\nReference images route through the edit endpoint (up to ~10). The composed \"image N\" naming applies as everywhere else. Mask-based inpainting exists at the API level but isn't surfaced — describe the change instead.\n\n## 🚨 WHAT GETS YOU BLOCKED — read before writing a prompt with a person in it\n\n**Receipt: 24 consecutive attempts on one character, 2026-08-24, same project and same rail.** Eleven were refused with `content_policy_violation` on the fal edit endpoint. The refusals were never about the scene — one of the blocked prompts was a woman standing at a kitchen counter with her hand on it. **Two phrasings were hard blocks, 5 for 5 each, and neither ever passed:**\n\n**1. Never describe the reference as a photograph of a real person.**\n\n> ❌ `Reference image 1 is a photograph of a woman. Use that exact woman.`\n> ✅ `Reference image 1 is a character identity sheet showing one woman across several panels — the face in the large portrait panel is the authority for her identity. Use that exact woman.`\n\nThe first reads to the filter as *recreate this real person's likeness*, which is a hard refusal regardless of what the rest of the prompt says. The second signals a fictional character and passes. **This is a wording change only — the reference image can be the same file either way.** One plate flipped from refused to accepted on this single sentence with nothing else altered.\n\n**2. Never attach a reference sheet containing a headless body panel.** A sheet whose full-body panels are cropped above the neck is refused every time, even with the correct opener. Regenerate the sheet with the head visible in every panel. Related, and already in this file's sheet guidance: phrase a cropped panel as *framing* (`cropped at the collarbone`), never as *absence* (`the head not shown`).\n\n**On top of those, ordinary content triggers still apply** and they stack independently — a correct opener does not rescue them:\n\n| Refused | Why, and the fix |\n|---|---|\n| A woman sitting on a bed in a bedroom | Domestic + bed reads as intimate. Move her to a chair, a rug, another room. |\n| A knife, even lying flat on a chopping board next to a lemon | The object is the trigger, not the framing. Swap it — a cast-iron pan cleared instantly. |\n\n**🚨 Refusals are PROBABILISTIC. Retry once before rewriting a word.** In the same session an identical prompt, identical reference, identical params was refused and then accepted on a straight re-fire. A rejected job returns no file and costs nothing, so a retry is free and a rewrite is not — rewriting first is how you end up changing four variables and learning nothing. **Only redesign after two or three refusals.**\n\n**And change ONE thing at a time.** The eleven refusals above took far longer to diagnose than they should have because a reference swap and an opener rewrite shipped in the same call. Isolate on the prompt you actually want, so a pass leaves you with a usable asset instead of a data point.\n\n## Filter regime\n\nOpenAI moderate — a third regime distinct from Gemini (NB family) and ByteDance (Seedream). Real-face references pass more readily than Gemini; violence/brand rules are similar. `slates-content-policy` applies unchanged.\n",
+    "slates-prompting-gpt-image-2-5": "---\nname: slates-prompting-gpt-image-2-5\ndescription: Prompting GPT Image 2.5 (Flare and Sunburst) — the readable-text / character-sheet / shot-grid engine AND the photoreal front-runner. Read before calling slates_generate_image with model gpt-image-2-5-flare or gpt-image-2-5-sunburst. Covers picking the variant, the five quality tiers (high is the default and the everyday seat), resolution classes (1k/2k=1080p/3k=1440p/4k), reference-image roles, text-accuracy prompting, panel/grid layout direction, edit constraints, and when to route to the Banana line instead.\n---\n\n# GPT Image 2.5 — sheets, grids, and text that actually reads\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — GPT Image 2.5.** The readable-text, ordered-panel and exact-placement engine, and the photoreal front-runner for people. Structure: subject and action, then the exact copy in quotes, then layout, then light.\n\n**Pick the seat first.** `flare` = the small/FAST seat, quality *comparable to* GPT Image 2 — drafts, exploration, volume. `sunburst` = OpenAI's *most capable*, higher quality, slower, same price — finals, hero frames, photoreal, multi-reference edits. Explore on Flare, finish on Sunburst.\n\n**The six levers**\n1. **Quote every string that must render verbatim** — `the sign reads \"OPEN 24 HOURS\"`. Quoted strings render most reliably.\n2. **Font FEEL, never a font name** — `clean geometric sans, high contrast`, `hand-painted brush lettering`.\n3. **Order dense copy explicitly** — `Line 1: \"...\" Line 2: \"...\"`. It respects the ordering.\n4. **Name the layout as a grid** for sheets and panels — `a 3x2 grid of panels, reading left to right, equal gutters`.\n5. **Give every reference image a ROLE** — subject / style / clothing / background. New emphasis in 2.5 and the highest-leverage change for multi-reference work.\n6. **Set `quality` deliberately.** Five rungs — `low`, `medium`, `high` (default), `xhigh`, `max` — spanning ~36× end to end, in UNEVEN steps: `max` is 4× `high`, but `xhigh` only ~1.8× it. `medium` is the draft seat; `high` is the everyday tier; reach past it only when tiny type, dense diagrams or many labelled elements ARE the job. 🚨 **Coming from GPT Image 2, the names moved one rung:** its `medium` is this `high`, its `high` is this `max` — same money, renamed ladder. Carrying an old value over silently buys a cheaper picture.\n\n**Examples**\n- `A 2x3 character turnaround sheet on a neutral grey field, equal gutters, reading left to right: front, three-quarter, profile, back, three-quarter back, top. One woman, mid-30s, cropped dark hair, olive field jacket. Flat even studio light, no cast shadows. Small caption under each panel naming the angle.`\n- `Photoreal portrait, natural window light from camera-left, visible skin texture and pores, 85mm compression. A man in his 50s in a charcoal knit, half-smile, looking just past lens.`\n\n**Hard constraint:** keep total on-image text under about 30 words for perfect accuracy — beyond that it degrades, gracefully but really. It has its own content filter, distinct from Gemini's.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use:**\n- a font NAME — describe the feel (`clean geometric sans, high contrast`) instead\n- a reference role essay (`Reference image 1 is a photograph of a woman. Use that exact woman.`) — name the subject inline instead\n- `8k`, `masterpiece`, `best quality`, `highly detailed` — quality incantations do nothing here either\n<!-- @banned:end -->\n\nGPT Image's edge is **character-level text accuracy** (~99% on English), ordered panels, and exact element placement — the jobs where every other model garbles a word or shuffles a layout. 2.5 inherits all of it and is better at each.\n\n## Which variant\n\n**Speed → Flare. Quality → Sunburst.** That is OpenAI's own routing rule, quoted from its image-prompting guide: *\"start with GPT Image 2.5 Flare when speed is the priority, or GPT Image 2.5 Sunburst when demanding quality requirements are the priority.\"* Same price either way, so the trade is purely latency against quality.\n\n🚨 **FLARE IS NOT AN UPGRADE OVER GPT IMAGE 2 — IT IS THE FAST ONE.** OpenAI, verbatim: *\"GPT Image 2.5 Flare is the small model, optimized for speed, with image quality **comparable to** GPT Image 2. GPT Image 2.5 Sunburst is the base model, optimized for quality, with **higher image quality than** GPT Image 2.\"* Their model pages agree: Flare is *\"our fastest model for high-quality, everyday image generation\"*, Sunburst *\"our most capable model for image generation and editing.\"* **Sunburst is the seat that beats what we had; Flare is the one that holds it at half the latency.** An earlier revision of this file called Flare \"better than GPT Image 2\" and sent Sunburst only to multi-reference edits — both wrong, corrected 2026-09-09 against the vendor docs.\n\n**The production pattern: explore on Flare, finish on Sunburst.** Drafts, layout checks and volume go to Flare. Finals, hero frames, photoreal people and any edit that must preserve identity or geometry go to Sunburst.\n\n**Sunburst's widest lead is multi-reference editing** — several references all surviving into one frame, the character-consistency-across-shots problem. Reach for it there first, but that is not the only place it belongs.\n\n⚠️ **The LMArena receipt, scoped.** At launch Arena had Sunburst #1 and Flare #2 across text-to-image, single-image edit and multi-image edit, with margins over GPT Image 2 of **+81 / +47** on multi-image edit (Image Edit Arena: Sunburst 1520, Flare 1491, GPT Image 2 1461). Two caveats were missing and both matter: the baseline is **GPT Image 2 at `medium`, which is this model's `high`** — not its top tier — and the boards were **preliminary, a few thousand votes each**. Arena says Flare beats GPT Image 2; OpenAI says comparable. Route on OpenAI's wording and treat the board as a tiebreaker, not a spec.\n\n🚨 **The GPT Image line is ALSO the photoreal front-runner, and this file said the opposite until 2026-08-24.** **Receipts:** Eric's direct call, plus a head-to-head on the Higgsfield rail where GPT Image 2 at `quality: high`, 2K beat both Nano Banana rails on skin realism for photoreal people — that result is why the whole AI-influencer ad lane generates its plates here. **Route photoreal to this line, not away from it.**\n\n⚠️ **Which SEAT reproduces it follows from the two facts above, and it is not the obvious one.** The receipt was measured on GPT Image 2 at `high`, which is this model's **`max`** — the ladder was renamed, not repriced (see `slates-model-selection`). Flare is only *comparable* to GPT Image 2, so **Flare at `max` is the floor: it holds the measured result rather than beating it.** Sunburst is documented as higher quality than GPT Image 2, which makes **Sunburst at `max` the seat most likely to exceed it** — and a photoreal final is exactly the \"quality outranks speed\" case OpenAI routes to Sunburst. Nobody has re-run the head-to-head on either seat, so this is reasoning from the vendor's positioning, not a measurement. **Run Flare-max against Sunburst-max on one plate before committing the lane, and write the result here.**\n\n**What the Banana line still owns:** edit-heavy work, and holding many subjects coherently in one frame. **Not the reference ceiling any more** — that line was true until 2026-09-09, when GPT Image went to its documented 16 against Banana's 14. Route on which model keeps them all recognisable, not on the count.\n\n**What would kill this:** a head-to-head at the intended crop going the other way. Per `slates-model-selection` § The meta-rule, re-run the evidence test when the roster changes — never carry a ranking forward on reputation. That rule is exactly what the 2026-08-24 correction failed, and exactly what the two ⚠️ notes above are honouring.\n\n## Quality tiers — always set explicitly\n\nAll five rungs are exposed, and they span ~36× end to end (2k class: $0.0044 → $0.158), which makes this the single biggest cost lever on the model. **The steps are UNEVEN — do not reason about them as a constant multiplier:** ~2.3× `low`→`medium`, ~3.9× `medium`→`high`, ~1.8× `high`→`xhigh`, ~2.25× `xhigh`→`max`. The same ratios hold at every OFFERED resolution class (2k/3k/4k); unoffered 1k differs slightly.\n\n| Tier | Use it for |\n|---|---|\n| `low` | Roughest pass — layout and composition checks, throwaway comps. |\n| `medium` | The draft seat. Cheaper than NB2 Lite and available up to 4K, which is why the draft lane moved here. |\n| `high` | **Default.** The everyday tier. Blind benchmarks on GPT Image 2 put this rung — which it called `medium` — within a hair of `max` (which it called `high`) at a quarter of the cost. Inherited from the old ladder, never re-run on 2.5, and it says nothing about `xhigh`. |\n| `xhigh` | One rung short of the top at about half its price (2k: 4 cr against `max`'s 8). Worth trying before `max`. |\n| `max` | Top of the ladder. Tiny type, dense diagrams, many labelled elements. |\n\n⚠️ **A tier label means different things on different models.** OpenAI: *\"The same quality label does not imply the same image quality or response time across models.\"* Flare at `max` and Sunburst at `max` are not the same picture, and neither matches Nano Banana's idea of \"high\".\n\n🚨 **The tier NAMES moved between versions and the strings did not.** GPT Image 2's `medium` is this model's `high`; its `high` is this model's `max` — same money, one rung of renaming. So a recipe, a doc or a memory that says \"GPT Image at medium\" means **`high` here**. Getting this backwards costs picture quality silently: nothing errors, the bill is correct for what was asked, and the image is just worse.\n\nNever rely on the provider default. fal's default is `high`, which is correct today — but it is the third rung of five rather than the top of two, so leaning on it means a fal-side change silently reprices you. The Slates ops send `high` unless you say otherwise.\n\n**Find the tier from the top down, then walk back.** OpenAI's own procedure: *\"If the output falls short, test a higher quality setting. Once it meets your requirements, test lower settings to see whether they preserve acceptable quality while reducing latency. Use `xhigh` or `max` only when they improve an unmet quality requirement within your latency budget.\"* A higher rung does **not** guarantee a better result on a given prompt. Compare `medium` against `high` when the job is small or dense text; that is where the rungs separate most visibly.\n\n## Resolution classes\n\n`1k` = 1024²-class · `2k` = 1920×1080-class · `3k` = 2560×1440-class · `4k` = 3840×2160-class. Pick 2k for most sheets/panels; 4k for print-density grids. 4K exists at every tier and is API-only — even paid ChatGPT can't render it.\n\n`1k` is not offered, and the reason is not its price: it is strictly dominated. At 1k you pay more for fewer pixels than at 2k, at **all five tiers**. Don't ask for it.\n\n⚠️ **Above 2560×1440 you are on a path OpenAI marks EXPERIMENTAL.** Verbatim: *\"Outputs with more than 3,686,400 total pixels ('2560x1440') are experimental.\"* That is the whole **4k** class (≈8.0 MP) plus 3k at 4:3/3:4 (≈3.70 MP). It bills normally and it works — but prove the shot at 2k or 3k 16:9 first, and do not be surprised by an odd frame at 4k.\n\n**Hard size bounds**, from fal's schema verbatim: each edge ≤ 3840 px, both edges multiples of 16, longer:shorter ratio ≤ 3:1, total pixels between 655,360 and 8,294,400. **The pixel ceiling is the one that actually bites** — the multiple-of-16 rule is documented but NOT enforced, and we have the receipt: 1920×1080 fails it (1080 = 67.5 × 16), is one of fal's own six priced canonical sizes, and metered clean. Slates picks sizes that respect the ceiling; these matter only if you hand-build a request.\n\n🚨 **THE ASPECT RATIO CHANGES THE PRICE ON THIS MODEL, and on no other image model.** OpenAI bills image OUTPUT TOKENS and the count tracks the frame's SHAPE, so at the same resolution class **`1:1` costs about 1.8× and `4:3`/`3:4` about 1.37× what `16:9` costs**; `9:16` costs the same as `16:9`. Metered 2026-09-09 and priced into the cost key, so the quote you get before generating is the real number — but if you are choosing between shapes and the budget is tight, **16:9 or 9:16 is the cheap one.** Every other image model charges the same whatever the shape.\n\n## Reference images — give every one a role\n\n**Assign a role to every reference image: subject, style, clothing, or background.** This is new emphasis in 2.5 and the highest-leverage change for the 16-reference character lane. An unroled pile of references makes the model guess what each one is for, and it guesses differently every run — which is the drift people mistake for a consistency failure.\n\nReference images route through the edit endpoint, **up to 16** — fal's documented `maxItems`, and the highest reference ceiling of any image seat in Slates (the Banana line takes 14). It was capped at 10 until 2026-09-09, which was never anybody's limit, just a number nobody had checked. The composed \"image N\" naming applies as everywhere else. Mask-based inpainting exists at the API level but is not surfaced: a mask is something the user has to paint, and there is no painting surface — describe the change instead.\n\n## Editing — separate the change from the constraints\n\n**State the change, then list what must survive.** \"Change only X,\" then name the invariants explicitly: identity, geometry, lighting, labels. For precise local edits also pin saturation, contrast, camera angle and surrounding objects — anything you do not pin is fair game for the model to move.\n\n**One change per iteration, and restate the constraints every turn.** Cross-turn drift is the named failure mode in OpenAI's own guidance: constraints do not persist across turns by themselves, so a multi-turn refinement that stops restating them will slowly rewrite the frame. This applies directly to multi-turn shot refinement.\n\n## Prompting for text accuracy\n\n- **Quote every string that must render verbatim**: `the sign reads \"OPEN 24 HOURS\"` — quoted strings render most reliably.\n- Say the text appears **once**, and give its position and typography.\n- Spell unusual words letter-by-letter.\n- Add `no extra text, no watermarks`.\n- Specify font *feel*, not font names: \"clean geometric sans, high contrast\", \"hand-painted brush lettering\".\n- For dense text (posters, UI mocks), list the copy as ordered lines: `Line 1: \"...\" Line 2: \"...\"` — it respects ordering.\n- **Don't bundle unrelated instructions into a text-rendering request.** A prompt that also redesigns the scene competes with the text for attention.\n- Keep total on-image text under ~30 words for perfect accuracy; beyond that, accuracy degrades gracefully but degrades.\n\n## Transparent backgrounds\n\nIf you need a cut-out rather than a scene, **ask for it explicitly and check the alpha**. OpenAI: request `background=transparent` and use PNG or WebP, then *\"check the decoded image's alpha channel, including hair, glass, shadows, and object edges\"* — a painted-white backdrop is the common failure and it is not transparency. Say what must NOT appear: *\"no solid backdrop, no checkerboard, no scenery, no watermark\"*, and do not let the product get restyled while the background is removed. **On every follow-up edit, repeat the transparency requirement** or it gets dropped. (Slates always requests PNG, so the format half is handled for you. **`background` IS surfaced now** — the Background control on the prompt bar, and `backgroundMode` on `slates_generate_image` / `slates_edit_image`. It is free: fal prices this family on size × quality alone.)\n\n## When an edit must not touch a region at all\n\nPrompting alone cannot guarantee pixel-identical pixels. OpenAI's own instruction: if a region must stay exactly as it was, **composite the approved edit back into the original image** rather than asking the model to preserve it. Treat \"preserve\" language as a strong bias, never a lock.\n\n## Structure a complex prompt in labeled sections\n\nFor anything with several requirements, OpenAI recommends organising the prompt as **scene, subject, details, constraints** with labeled sections. Same content, easier to read and to change one part without disturbing the rest — which is what makes the one-change-per-iteration rule practical.\n\n**Say \"photorealistic\" or \"real photograph\" when that is the goal.** It is not inferred from a detailed description; ask for it directly, then describe framing and texture.\n\n## Concrete visuals beat mood words\n\nName materials, lighting, colour and medium. Mood words are cues only — \"cinematic\", \"moody\", \"epic\" tell the model almost nothing on their own. Give scale, atmosphere and colour instead. Camera specs (`85mm`, `f/1.4`) are appearance hints, not a physical simulation; they bias the look, they do not compute optics.\n\n**For people, state body framing and scale**: \"full body visible, feet included\", \"hands naturally gripping the handlebars\". This is also the safest way to phrase a crop — see the blocked-phrasings section below.\n\n**No special syntax is required.** Prose, JSON and tagged blocks all work equally well, so pick whatever stays maintainable in the caller.\n\n## Panels, sheets, and grids\n\n- State the grid explicitly and number the cells: \"a 2×3 grid of panels, numbered 1–6, reading left-to-right, top-to-bottom\".\n- Give each cell ONE content clause: \"Panel 3: the character mid-jump, side view\".\n- Character identity sheets: GPT Image holds both the structured panel layout AND photoreal skin, which is why the influencer-ad lane builds its sheets here. Reach for NB2/NB Pro when it is an edit of an existing sheet, or when many subjects have to stay recognisable at once — not for the reference count, which GPT Image now leads at 16.\n\n## 🚨 WHAT GETS YOU BLOCKED — read before writing a prompt with a person in it\n\n**Receipt: 24 consecutive attempts on one character, 2026-08-24, same project and same rail.** Eleven were refused with `content_policy_violation` on the fal edit endpoint. The refusals were never about the scene — one of the blocked prompts was a woman standing at a kitchen counter with her hand on it. **Two phrasings were hard blocks, 5 for 5 each, and neither ever passed:**\n\n**1. Never describe the reference as a photograph of a real person.**\n\n> ❌ `Reference image 1 is a photograph of a woman. Use that exact woman.`\n> ✅ `Reference image 1 is a character identity sheet showing one woman across several panels — the face in the large portrait panel is the authority for her identity. Use that exact woman.`\n\nThe first reads to the filter as *recreate this real person's likeness*, which is a hard refusal regardless of what the rest of the prompt says. The second signals a fictional character and passes. **This is a wording change only — the reference image can be the same file either way.** One plate flipped from refused to accepted on this single sentence with nothing else altered.\n\n**2. Never attach a reference sheet containing a headless body panel.** A sheet whose full-body panels are cropped above the neck is refused every time, even with the correct opener. Regenerate the sheet with the head visible in every panel. Related, and already in this file's sheet guidance: phrase a cropped panel as *framing* (`cropped at the collarbone`), never as *absence* (`the head not shown`).\n\n⚠️ **These refusals were measured on GPT Image 2, not on 2.5.** The classifier belongs to OpenAI rather than to a model version, so the phrasing rules carry — but they are inherited, not re-measured. If Flare or Sunburst accepts one of the blocked phrasings, that is a new receipt to write down here, not a reason to delete this one.\n\n**On top of those, ordinary content triggers still apply** and they stack independently — a correct opener does not rescue them:\n\n| Refused | Why, and the fix |\n|---|---|\n| A woman sitting on a bed in a bedroom | Domestic + bed reads as intimate. Move her to a chair, a rug, another room. |\n| A knife, even lying flat on a chopping board next to a lemon | The object is the trigger, not the framing. Swap it — a cast-iron pan cleared instantly. |\n\n**🚨 Refusals are PROBABILISTIC. Retry once before rewriting a word.** In the same session an identical prompt, identical reference, identical params was refused and then accepted on a straight re-fire. A rejected job returns no file and costs nothing, so a retry is free and a rewrite is not — rewriting first is how you end up changing four variables and learning nothing. **Only redesign after two or three refusals.**\n\n**And change ONE thing at a time.** The eleven refusals above took far longer to diagnose than they should have because a reference swap and an opener rewrite shipped in the same call. Isolate on the prompt you actually want, so a pass leaves you with a usable asset instead of a data point.\n\n## Filter regime\n\nOpenAI moderate — a third regime distinct from Gemini (NB family) and ByteDance (Seedream). Real-face references pass more readily than Gemini; violence/brand rules are similar. `slates-content-policy` applies unchanged.\n",
     "slates-prompting-inworld-tts": "---\r\nname: slates-prompting-inworld-tts\r\ndescription: How to use Inworld Realtime TTS-2, the VOICE seat. Read before calling slates_generate_audio with model inworld-tts-2. Speech in a SPECIFIC voice, billed per character - the prompt is the words spoken, verbatim. Covers the identity-versus-acoustics rule (what a reference clip does and does not carry), how to write a line so it is performed rather than read, when to reach for seed-audio instead, and the voice-consent rule.\r\n---\r\n\r\n# Inworld Realtime TTS-2 — the voice seat\r\n\r\n<!-- @card:start -->\r\n<!-- slates-only -->\r\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\r\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\r\n     model, so it is the ONE piece of positive craft guidance the agent cannot\r\n     skip. Keep it under 2,400 characters (the build fails above that) and keep\r\n     the rationale and the worked examples in the body below. -->\r\n<!-- /slates-only -->\r\n**Card — Inworld TTS-2.** Speech in a SPECIFIC voice. The prompt is the words spoken, verbatim — not a description of them. Text length determines the bill.\r\n\r\n**IDENTITY, NOT ACOUSTICS — the rule that decides whether cloning works**\r\nA reference carries WHO is speaking: timbre, pitch, accent, age, vowel shape. It does NOT carry WHERE they are — room tone, distance, phone EQ, reverb and mic character are *acoustics*, and this model reproduces the identity while discarding the room. So:\r\n1. **A noisy reference does not give a noisy read — it gives a WORSE identity.** Music, a second speaker or heavy reverb corrupt what is being extracted. Use a clean single-speaker recording.\r\n2. **You cannot get \"on a payphone\" by cloning a payphone recording.** Acoustics come from the MIX, or from `seed-audio` which renders a room.\r\n\r\n**DIRECTION GOES IN SQUARE BRACKETS. PARENTHESES ARE SPOKEN ALOUD.** `[whispering] I hope nobody notices` is whispered; `(quietly) I hope nobody notices` says the word \"quietly\" out loud. Verified by ear — the easiest way to ruin a take.\r\n\r\n- **Plain English works inside them** — it is natural-language steering, not a fixed vocabulary: `[very quiet]`, `[whisper in a hushed style]`, `[very slow]`, `[say excitedly]`. Non-verbals are their own tags: `[laugh]`, `[sigh]`, `[breathe]`, `[clear throat]`.\r\n- **A tag it does not recognise is still consumed, and still changes the read.** Never spoken, never an error — so a mistyped tag fails SILENTLY and only listening catches it.\r\n- **Tags persist across sentences** until changed; `[reset]` returns to normal.\r\n- **Punctuation is the timing.** `Wait. Stop.` differs from `Wait, stop.`\r\n- **One line, one take.** Split a paragraph so a bad clause costs one re-roll.\r\n- **Spell numbers and titles aloud:** `twenty twenty-six`, `Doctor Reyes`.\r\n\r\n**Route elsewhere when:** the scene needs dialogue mixed with effects and room tone in one pass (`seed-audio`), or it is a single non-speech sound (`eleven-sfx`). This surface makes ONE voice saying ONE thing, cleanly.\r\n\r\n**Hard constraints:** no duration parameter — length falls out of the text. Exactly one voice source: a preset `voiceId` from `slates_list_voices`, a clip as `voiceReferenceAssetId` (a character's voice clip to speak AS the character, or any clean clip of one speaker), or `voiceDescription`.\r\n<!-- @card:end -->\r\n\r\n<!-- @banned:start -->\r\n<!-- slates-only -->\r\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\r\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\r\n     estimate, and every submitted prompt is matched against it. Keep entries\r\n     backticked and prose outside the backticks. -->\r\n<!-- /slates-only -->\r\n**Never use** — the prompt on this surface is SPOKEN ALOUD, so anything that describes the audio instead of being the audio gets read out as words:\r\n\r\n- `SFX`, `Ambient noise`, `Background music` as labels — this model speaks; it does not render a scene. Use `seed-audio` for those.\r\n- shot language: `wide shot`, `slow push in`, `warm tungsten` — video-prompt words, and here they would literally be said aloud\r\n- `voiceover`, `narrator says`, `he says` as stage directions wrapping the line — write only the words that should come out of the speaker\r\n<!-- @banned:end -->\r\n\r\n## Why the prompt is not a prompt\r\n\r\nOn every other surface in Slates the prompt DESCRIBES what you want and the model interprets it. Here the prompt IS the deliverable: each character is spoken aloud and each character is billed. `a gravelly man says he is tired` produces a voice saying the words \"a gravelly man says he is tired\".\r\n\r\nThat also means the two numbers a user cares about are the same number. The text length sets the price (in 250-character buckets) and sets the length of the audio. There is nothing to choose and nothing to reconcile.\r\n\r\n## Steering the delivery\r\n\r\n`VERIFIED BY EAR, 2026-09-05.` Every claim in this section was listened to, not\r\ninferred — an earlier draft of this skill documented tag forms that had only been\r\nprobed for an HTTP 200, which proves the request was accepted and nothing about\r\nwhether it was obeyed.\r\n\r\n**Square brackets are consumed. Parentheses are read aloud.** That is the whole\r\nrule, and getting it wrong is not a subtle degradation — the audience hears a\r\nnarrator say the word \"quietly\" in the middle of your line.\r\n\r\n| Written | What comes out |\r\n|---|---|\r\n| `[whispering] I really hope nobody notices that.` | whispered, tag not spoken ✅ |\r\n| `[very quiet] I really hope nobody notices that.` | very quiet, tag not spoken ✅ |\r\n| `[whisper in a hushed style] …` | hushed, tag not spoken ✅ |\r\n| `[very slow] …` | slowed right down, tag not spoken ✅ |\r\n| `[laugh] …` | an actual laugh, then the line ✅ |\r\n| `(quietly, under his breath) …` | 🚨 **the words \"quietly, under his breath\" are SPOKEN** |\r\n\r\n**Plain English works — it is natural-language steering, not a fixed vocabulary.**\r\nBoth the documented phrasings (`[whisper in a hushed style]`) and ordinary adverbs\r\n(`[whispering]`) were obeyed. Write the direction the way you would say it to an\r\nactor.\r\n\r\nThe eight dimensions the model steers on, with a working example of each:\r\n\r\n| Dimension | Example |\r\n|---|---|\r\n| Emotion | `[say excitedly]`, `[sound sad]`, `[sound terrified]` |\r\n| Articulation | `[say with force]`, `[articulate clearly]` |\r\n| Intonation | `[say with a rising pitch]` |\r\n| Volume | `[very quiet]`, `[very loud]` |\r\n| Pitch | `[say in a low tone]` |\r\n| Range | `[say playfully]`, `[say with no pitch variation]` |\r\n| Speed | `[very fast]`, `[very slow]` |\r\n| Vocal style | `[whisper in a hushed style]`, `[give a nasal quality]` |\r\n\r\nNon-verbals sit inline where they happen: `[laugh]`, `[sigh]`, `[cough]`,\r\n`[breathe]`, `[yawn]`, `[clear throat]`.\r\n\r\n### Four rules that are not obvious\r\n\r\n1. 🚨 **A tag it does not recognise is still consumed, and still changes the read.**\r\n   `[zzzqqq]` is not spoken and does not error — it produces a different, arbitrary\r\n   delivery. So a typo in a tag is SILENT: there is no rejection, no warning, and no\r\n   way to catch it except listening to the take. Treat an unexpected performance as\r\n   a possible misspelled tag before you blame the voice.\r\n2. **Tags persist across sentences.** A `[very slow]` at the top governs everything\r\n   after it until something changes it. Use `[reset]` to go back to normal rather\r\n   than assuming the next sentence starts clean.\r\n3. **Do not stack opposing directions.** `[whisper in a hushed style]` together with\r\n   `[very loud]` produces unpredictable results — the model is resolving a\r\n   contradiction, and which side wins is not something you can rely on.\r\n4. **Tags COUNT toward the billed characters**, even though they are never spoken.\r\n   They are part of the text sent to the vendor, so the vendor charges for them and\r\n   so do we — billing what was actually sent is the only honest basis. It rarely\r\n   matters (a 13-character tag inside a 250-character bucket), but a line sitting\r\n   just under a bucket boundary can be pushed into the next one by a long\r\n   direction. Prefer `[very slow]` over `[say this one very slowly please]`.\r\n\r\n## Identity versus acoustics, at length\r\n\r\nThis is the distinction that decides whether the feature feels good, and it is worth being precise about because the failure is quiet — you get a usable clip that is subtly not the person.\r\n\r\n**What a reference clip transfers:** vocal timbre, pitch range, accent and regional vowels, apparent age, speech rate tendencies, and the particular rasp or breathiness of the source speaker.\r\n\r\n**What it does not transfer:** the room, the microphone, the codec, the distance from the mic, any processing on the source, and any other sound present in it.\r\n\r\nSo the ideal reference is boring: one person, close to a microphone, no music, no second speaker, no heavy reverb, five to fifteen seconds, speaking normally rather than performing. A phone voice memo in a quiet room beats a beautifully produced clip with a music bed underneath it.\r\n\r\n**Two failure modes, both common:**\r\n\r\n- *\"I cloned my podcast intro and it doesn't sound like me.\"* The intro had music under it. The model averaged the music into the identity. Re-clone from a clean stretch.\r\n- *\"I want the line to sound like it's coming through a car radio.\"* Clone the clean voice, then EQ and process the returned clip on the timeline. A radio-sounding reference makes a worse voice, not a radio effect.\r\n\r\n## Getting the voice onto the call\r\n\r\nExactly one source per call, and none of them requires a character to exist first:\r\n\r\n- **A preset:** `slates_list_voices` lists stock voices with gender, age, accent and tags — filter by any of them, or search the descriptions (\"gravelly\", \"narration\"). Pass the chosen `voiceId`. Presets clone nothing, so they are the fastest path and avoid the clone-creation rate ceiling.\r\n- **Speak AS a character:** `voiceReferenceAssetId: <its voiceAssetId>` (the clip on the row `slates_list_characters` returns). The seat clones the clip for that take and discards the vendor voice afterwards, so there is nothing to reconcile — but cloning shares a ceiling of two new voices a minute across every Slates user, so a run of lines in one cloned voice pauses between takes rather than failing. Send each line once; do not re-send one that already came back. Any other clean clip of one speaker works the same way.\r\n- **A voice with no recording:** `voiceDescription` (7–1000 characters of words). If it will be used again, keep the returned clip on a character with `slates_update_character` (`voiceAssetId`) so later lines clone the same clip instead of designing a new voice each time — a convenience, never a requirement.\r\n\r\n## Consent\r\n\r\nCloning a real person's voice needs that person's explicit, documented permission, scoped to what you are making. Clone from original human recordings only — never from another model's output. This is the same gate the real-face route applies to likeness, and it applies here for the same reason.\r\n\r\n## Worked examples\r\n\r\n**A line with a direction**\r\n\r\n```\r\n[very quiet] I heard what you said in there. I'm not going to pretend I didn't.\r\n```\r\n\r\n**A line that needs its numbers spoken**\r\n\r\n```\r\nThe vote was three hundred and twelve to eighty-nine. It carried at four minutes past midnight.\r\n```\r\n\r\n**A paragraph, split into three takes** — so one bad clause costs one re-roll:\r\n\r\n```\r\n1. You keep asking me why I stayed.\r\n2. It wasn't loyalty. It wasn't even fear, not by the end.\r\n3. [very slow] It was that I couldn't picture the version of me that left.\r\n```\r\n\r\n**What NOT to send**\r\n\r\n```\r\n(gravelly, tired) a tired old man narrates the opening of the film, wide shot, warm tungsten\r\n```\r\n\r\nEvery word of that is spoken aloud — **including the parenthetical**, which is the\r\ntrap: it looks like a stage direction and is treated as dialogue. Describe the voice when you are CHOOSING one (`voiceDescription`, or the desktop's voice picker); the prompt is only ever the words.\r\n",
     "slates-prompting-kling-v3": "---\nname: slates-prompting-kling-v3\ndescription: How to prompt Kling V3.0 (Kuaishou). Read before calling slates_generate_video with kling-v3.0-std, kling-v3.0-pro, or kling-v3.0-omni. Kling has dialogue + SFX + ambient native syntax (Omni adds multi-character dialogue and language codes). Multi-shot rules differ from Seedance/Veo — don't cross syntaxes.\n---\n\n# Kling V3.0 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Kling V3.0.** The general default. Define the core subjects clearly at the START and keep those descriptions identical across shots. Up to 15s, up to 6 cuts, and the strongest image-to-video identity hold in the catalogue.\n\n**The five levers**\n1. **Dialogue in quotes** — `Character says, \"exact words here\"`. On Omni, direct the voice with `Gender + Age + Voice quality + Speech rate + Emotional tone + Language`: `[Character A: Detective, mid-40s, raspy, slow cadence, weary]: \"I've seen this before.\"`\n2. **Unique speaker labels, no pronouns after the introduction.** `he`, `the agent`, any synonym causes voice drift.\n3. **Sound has real syntax** — `SFX: heavy boots on wet pavement, distant siren wailing`, `Ambient noise: city traffic`, `Background music: low cello`. Always physical-cause specific; `SFX: footsteps` is not enough.\n4. **Motion adverbs modulate energy directly** — `slowly`, `rapidly`, `gently`, `explosively`. One primary camera move per shot, never stacked.\n5. **On image-to-video, do NOT re-describe the image.** It is an anchor; prompt how the scene EVOLVES from it — movement, camera, environmental change.\n\n**Examples**\n- `A detective in a wet grey overcoat stands under a stairwell light. He steps forward slowly as the light flickers. [Character A: Detective, mid-40s, raspy voice, slow cadence, weary]: \"I've seen this before.\" SFX: heavy boots on wet concrete, distant siren wailing. Ambient noise: rain on metal.`\n- `Camera tracks right alongside a cyclist crossing a bridge at dusk. She rises out of the saddle rapidly as the grade steepens. Ambient noise: wind, tyres on wet asphalt, distant traffic.`\n\n**Hard constraint:** `Immediately` (Omni only) removes the natural conversational beat between speakers — use it when timing matters and leave it out when it does not. Kling has a real `negativePrompt` field, unlike Seedance; start from the standard block and layer scene-specific suppressions.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use:**\n- `SFX: footsteps` and any label-only effect — physical-cause specificity or nothing\n- a pronoun or synonym for a speaker after the first introduction (`he`, `the agent`) — it causes voice drift; repeat the full label\n- `single continuous take` — Seedance's phrase, and it fights Kling's multi-shot\n<!-- @banned:end -->\n\nKuaishou's video model. Three tiers: `kling-v3.0-std` (general use, no audio), `kling-v3.0-pro` (higher visual quality, no audio), `kling-v3.0-omni` (multi-character dialogue + audio-visual co-generation).\n\nUp to 15s. Multi-shot supported (up to 6 cuts in 15s total). Strong on image-to-video — preserves identity, layout, and text from the input image well.\n\n## Subject definition rule (verbatim, fal blog)\n\n> \"Define your core subjects clearly at the beginning of the prompt and keep descriptions consistent across shots.\"\n\n## Dialogue syntax\n\n```\nCharacter says, \"exact words here\"\n```\n\nUse quotation marks for precise speech. Languages (Omni only): EN, ZH, JA, KO, ES.\n\n## Voice direction formula (Omni)\n\n```\nGender + Age Range + Voice Quality + Speech Rate + Emotional Tone + Language\n```\n\nExample:\n```\n[Character A: Detective, mid-40s, raspy voice, slow cadence, weary]: \"I've seen this before.\"\n```\n\nTone phrases that fire:\n- `speaking in a hushed, trembling whisper`\n- `shouting with commanding authority`\n- `clear, fearful voice`\n- `with a trembling voice, \"I'm scared\"`\n\n## The `Immediately` keyword (Omni only)\n\nWithout `Immediately`, Kling adds a natural conversational beat between speakers. With it, dialogue is back-to-back. Use when timing matters.\n\n```\n[Alice]: \"Get down!\" Immediately, [Bob]: \"Where?\"\n```\n\n## Speaker label discipline\n\nUnique labels per character. **No pronouns or synonyms after first introduction** — they cause voice drift.\n\n✅ `[Character A: Black-suited Agent]` ... `[Character A: Black-suited Agent]: \"Stop.\"`\n❌ `[Agent]... then he says...`\n\n## Multi-character dialogue (Omni)\n\n```\nAlice says in English, \"Hello!\" Then Bob replies in Spanish, \"¡Hola!\"\n```\n\n## Sound effects, ambient noise, music\n\n```\nSFX: thunder cracks, footsteps approaching\nAmbient noise: city traffic, birds chirping, ocean waves\nBackground music: tense orchestral strings, low cello\n```\n\nSFX accepts physical-cause specificity:\n- ✅ `SFX: heavy boots on wet pavement, distant siren wailing`\n- ❌ `SFX: footsteps`\n\n## Image-to-video guidance\n\n**Verbatim (fal blog):**\n> \"Treat the input image as an anchor. Kling 3.0 excels at preserving the identity, layout, and text details. Focus prompts on how the scene evolves *from* the image: subtle movements, camera motion, or environmental changes.\"\n\n**Don't re-describe what's already in the image.** Focus on motion, changes, evolution.\n\n## Multi-shot — what makes them hit\n\n**Hard cap: total duration ≤ 15s across all shots. Max 6 cuts.**\n\nHit conditions:\n- Shot labels are explicit: `Shot 1:`, `Shot 2:`\n- One primary action per shot\n- Subject described identically in each shot block\n- Camera move per shot is **one verb**, not a chain\n- Per-shot blocks: 30-60 words\n\nMiss conditions:\n- Compressing narrative into one paragraph\n- Pronoun-only references after the first shot\n- Mixing camera moves within a shot (\"pan then orbit then push in\")\n- Extreme wide → extreme close in adjacent shots without reference images\n\n## Element references (Omni)\n\nUpload 2-4 multi-angle reference photos per character/object. Tag inline:\n\n```\n@element1 is the protagonist (refs: front, side, back angles).\n@element2 is the antagonist.\n```\n\n## Reference discipline (character / environment refs)\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### For Kling specifically\n\n- **Kling's consistency lever is \"lock the subject with a fixed label reused verbatim.\"** That is Kling's phrasing for rules 2 and 3, and it is stricter than the others: **pronoun and synonym drift breaks it**, so the exact same label must appear on every single mention — not \"he\", not \"the detective\" after you named him. Reusing the label verbatim is the whole game. Slates composes this for you from `@mentions`.\n- **Element references are the transport for rule 1** — 2-4 multi-angle photos per character/object, tagged `@element1` / `@element2` (see Element references above). The cap is 4 combined refs on the edit path.\n\n## Negative prompting — has a real field\n\nKling exposes `negative_prompt` on the fal endpoint (different from Seedance which has none). Default block to start from:\n\n```\nblurry, low quality, watermark, text overlay, distorted hands, extra fingers,\nduplicate limbs, unnatural skin texture, overly saturated colors, lens flare,\nfloating objects, inconsistent shadows, jittery, flickering, morphing face\n```\n\nLayer scene-specific suppressions on top.\n\n## Cinematic tactics\n\n- **Motion adverb precision** modulates motion energy directly: `slowly`, `rapidly`, `gently`, `explosively`\n- **Camera vocabulary that registers as instructions:** profile shot, tracking, following, freezing, panning, \"moving in sync with the subject\"\n- **One primary camera move per shot** — never stack\n\n## Tier choice\n\n- **Standard**: general use, no audio\n- **Pro**: higher visual quality, no audio\n- **Omni**: multi-character dialogue, audio-visual co-gen, language codes, `@elementN` references\n\nPick by capability: need dialogue/audio → Omni; need maximum visual quality silent → Pro; everything else → Standard. Prices change — check current numbers before choosing a tier<!-- slates-only -->; call `slates_estimate_generation_cost` or `slates_list_available_models`<!-- /slates-only -->.\n\n## Benchmark prompt structure\n\n```\n[Character A: <role>, <voice quality>]: \"<line>.\" Immediately, [Character B: <role>, <voice quality>]: \"<reply>.\"\nAmbient noise: <soundscape>.\nCamera <single move>.\n```\n\nCinematic example (paraphrasing fal blog patterns):\n> \"Shot 1: Wide establishing shot of a neon-lit alleyway in heavy rain, steam rising from grates. Camera slowly tracks forward.\n> Shot 2: Medium shot of a detective in a trench coat ducking under an awning, water dripping from his hat brim. [Detective: weary, raspy]: 'I knew she'd come back.' Ambient noise: distant traffic, rain on metal.\n> Shot 3: Close-up on his eyes, narrowing as headlights flash across his face.\"\n\n<!-- slates-only -->\n## Pre-flight: references arrive inline, refer by code\n\nWhen you call `slates_generate_video` with `firstFrameAssetId` or `ingredientAssetIds`, the first call returns those references **inline as image content blocks** alongside cost + `requires_confirm: true`. Look at them, revise prompt if needed, then re-call with `confirm=true`. Kling Omni multi-character with several ingredient images especially benefits — confirm each character image lands cleanly before spending.\n\nWhen talking to the user about the gen, refer to each reference by its short code: `IMG-A12 — Detective Closeup`. The user sees that code as a gallery badge.\n\n- ✅ \"I'm anchoring on **IMG-A12** as the detective and **IMG-A18** as the alleyway environment — Omni will handle the line delivery in EN.\"\n- ❌ \"I'm using the detective image and the alley one...\" (which alley? Three exist.)\n<!-- /slates-only -->\n\n## Video-to-video EDIT<!-- slates-only --> (`slates_edit_video`)<!-- /slates-only --> — @Video1 / @ElementN / @ImageN\n\nKling O3 edit takes an EXISTING 3-15s clip and changes only what the prompt names — character swap, environment change, style transfer — in one pass, no masking. Original motion, camera, and audio are preserved by default. Its notation is Kling's own, different from the \"image N\" naming used everywhere else:\n\n- **`@Video1`** — the source clip (always; the transport anchors the instruction to it).\n- **`@Element1..`** — subjects to swap IN. Each element = one frontal image + up to 3 angle images<!-- slates-only --> (pass as `characterAssetIds`; @mention names in the prompt compile to @ElementN automatically)<!-- /slates-only -->.\n- **`@Image1..`** — style/appearance references<!-- slates-only --> (pass as `styleAssetIds`)<!-- /slates-only -->.\n- Max **4 combined** element + image refs per edit.\n\n**Prompt shape — the change, not the whole scene:**\n\n```\nReplace the man in @Video1 with @Element1, keeping his walk cycle, the camera move, and the rain unchanged.\n```\n\n```\nEdit @Video1: turn the daytime street into a neon-lit Tokyo alley at night, wet asphalt reflections. Apply the visual style of @Image1. Keep the subject and camera motion exactly as they are.\n```\n\nRules:\n- Name what CHANGES; explicitly state what stays (\"keep the motion / camera / everything else unchanged\") — the model preserves better when told to.\n- One edit intent per pass. Chain passes for compound changes (each output is itself an editable clip, linked to its parent).\n- Billing is per second of OUTPUT ≈ the clip length, rounded UP to the next second. A 7.3s clip bills as 8s.\n- Clip constraints: 3-15s, 720-3840px, MP4/MOV. Agents can pre-trim on the timeline when a clip runs long.\n- Routing: Kling edit is the default edit tool (element lock + audio intact); Seedance edit/relocate wins style-transfer-heavy re-imaginings<!-- slates-only --> — see `slates-model-selection`<!-- /slates-only -->.\n\n## Sources\n\n- [fal.ai — Kling 3.0 Prompting Guide](https://blog.fal.ai/kling-3-0-prompting-guide/)\n- [Vidguru — Kling 3.0 Omni Guide](https://www.vidguru.ai/blog/kling-3.0-omni-guide.html)\n- [AcceptPrompt — Kling 3 Prompt Guide](https://www.acceptprompt.com/blog/kling-3-prompt-guide)\n- [DataCamp — Kling 3.0 Tutorial](https://www.datacamp.com/tutorial/kling-3-0)\n",
     "slates-prompting-lip-sync": "---\nname: slates-prompting-lip-sync\ndescription: How to set up lip-sync — Kling-only (dedicated lip-sync and avatar endpoints, 5-second outputs). Read before calling slates_generate_lip_sync. Two flows — video→video re-dub and image→video avatar — with different inputs, pricing, and gotchas. Voice catalog, framing rules, audio file constraints, and which tier to pick. Also covers the Seedance alternative, which is a normal video generation rather than a mode of this tool.\n---\n\n# Lip-sync — setup guide\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Lip-sync (Kling only).** Two different flows with different inputs and different prices; every output is 5 seconds.\n\n**The five levers**\n1. **Pick `sourceType` deliberately** — `video` re-dubs an existing talking head (cheapest); `image` animates a still portrait (avatar-standard, then avatar-pro only on the final selected take).\n2. **The `prompt` on the avatar flows is SCENE CONTEXT, not motion direction.** Ambience, lighting, micro-expression: `Soft rim light`, `warm office`, `cool blue evening light through a window`, `gentle confident smile between sentences`, `focused intent expression`.\n3. **Clean the audio before uploading** — `noise-reduced`, `levelled`. Lip detection is sensitive, and a raw recording is the most common cause of a bad take.\n4. **Iterate on the SOURCE or the AUDIO, never on a refinement prompt** — there is not one. If the output is wrong, change the input.\n5. **Use avatar-standard for first-pass dialogue takes**, and switch to pro only once the line is locked. Facial fidelity is not visible until then.\n\n**Examples**\n- `Soft rim light, warm office, gentle confident smile between sentences.`\n- `Cool blue evening light through a window, focused intent expression.` (Or `.` — an empty prompt is fine when you have nothing to add.)\n\n**Hard constraint:** it is Kling-only and always 5 seconds. For a generated PERFORMANCE instead — head movement, gesture, delivery energy, with the dialogue as a native conditioning signal — that is a normal Seedance video generation with the clip attached as a video reference, not a mode of this tool. A real recording, or a cloned/cast voice rendered on `inworld-tts-2`, for production; this tool's built-in TTS is for scratch.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** — the avatar prompt is scene context and motion verbs are ignored:\n- `turns her head`, `raises an eyebrow`, `hand gestures`, `nods`, `walks`\n- `reader_en_m-v1` — listed in fal's docs, returns \"Voice id not found\" in production\n<!-- @banned:end -->\n\n**This tool is Kling-only.** It wraps Kling's dedicated lip-sync and avatar endpoints; every entry is a real endpoint and every output is 5 seconds.\n\n| Flow | Source | Model | Cost | Use case |\n|------|--------|-------|-----------|----------|\n| Re-dub | video clip | kling-lip-sync-video | ~4 credits / 5s | Replace dialogue on an existing talking head |\n| Avatar standard | still image | ai-avatar/v2/standard | ~14 credits / 5s | Animate a portrait into a talking avatar |\n| Avatar pro | still image | ai-avatar/v2/pro | ~29 credits / 5s | Higher facial fidelity for hero shots |\n\nPick `sourceType` deliberately — it decides the pricing tier and the underlying endpoint.\n\n## Want Seedance instead? That is a video generation, not a mode here\n\nSeedance can generate the performance rather than bolting a mouth onto finished pixels — head movement, gesture, delivery energy, with the dialogue as a native conditioning signal, and a video source keeps its own voice. **It is not an engine switch on this tool.** Run a normal `slates_generate_video` on `seedance-2` with the clip (or portrait) attached as a video/ingredient reference and the dialogue written into the prompt yourself.\n\nThat is the same endpoint the old `engine=seedance-2` branch called — it just built the sentence for you, invisibly, and it presupposed a \"video 1\" that might not exist. Writing the prompt is the whole difference, and it is the part you want control of.\n\n- Driving clips must be 2–15s; output duration is whatever you set (4–15s).\n- Video references bill COMBINED input+output seconds (`seedance-2*-vref-*` keys) — pass the clip duration and quote before confirming.\n- Faces go through the normal cascade: `seedanceFace` for a character, `[REAL_FACE_DETECTED]` → `seedanceRealFace` + `realFaceConsent` for a real person.\n\nEverything below is about the Kling tool.\n\n## Choosing video vs avatar\n\nUse **video** (re-dub) when:\n- A talking-head clip already exists (Slates-generated, recorded, or imported)\n- The mouth/face is already moving and only the audio needs to change\n- ~4 credits is hard to beat for short dialogue replacement\n\nUse **avatar** when:\n- Only a still portrait exists\n- The character needs to come alive from a single image\n- Identity + face fidelity matter (avatar-pro for hero shots, standard for everything else)\n\n## Source asset constraints\n\n### Video flow (`sourceType: 'video'`)\n- Format: mp4 or mov\n- Duration: 2–10s (lip-sync output is always 5s — long videos get trimmed)\n- Resolution: 720p or 1080p (480p will be rejected)\n- Max file size: 100MB\n- Face must be visible and roughly facing camera. Profile shots fail.\n- Existing audio is replaced.\n\n### Avatar flow (`sourceType: 'image'`)\n- Min 512×512, PNG/JPG/WebP\n- **Face occupies 60–70% of frame.** This is the single biggest avatar quality lever.\n- Eyes open, mouth neutral, looking near-camera. Side profile = bad output.\n- Single subject, clean background. Group photos confuse the face anchor.\n\n## Audio source\n\nTwo ways to drive the lips:\n\n### TTS (`audioMethod: 'tts'`)\n- Pass `ttsText` (the words spoken)\n- Optional: `ttsVoice` (default `oversea_male1`), `ttsLanguage` (default EN), `ttsSpeed` (default 1.0)\n- **Hard cap: 120 characters of text.** Longer = silently truncated.\n- Languages: EN, ZH, JA, KO, ES\n\n### Upload (`audioMethod: 'upload'`)\n- Pass `audioFilePath` — absolute path to an audio file on the user's machine\n- Format: mp3, wav, m4a, ogg, aac\n- Max 5MB\n- Duration: 2–60s (output is 5s — longer audio gets trimmed)\n- Single clean voice. Music underneath, multiple speakers, or noisy mics produce garbage lips.\n\nPrefer upload for production-quality voice. TTS for fast iteration / placeholder dialogue.\n\n## Voice catalog (TTS)\n\nReliable English voices (verified working on the fal endpoint as of 2026):\n\n| Voice ID | Description |\n|----------|-------------|\n| `oversea_male1` | Male, English — default, stable |\n| `commercial_lady_en_f-v1` | Female commercial English |\n| `uk_boy1` | Young man, UK accent |\n| `uk_man2` | Man, UK accent |\n| `uk_oldman3` | Older man, UK accent |\n| `calm_story1` | Storyteller / narrator |\n\nAvoid `reader_en_m-v1` — listed in fal.ai docs but returns \"Voice id not found\" in production.\n\nFull 48-voice list (ZH, JA, KO included): https://fal.ai/models/fal-ai/kling-video/lipsync/text-to-video/api\n\n## Speech-rate notes\n\n`ttsSpeed` range 0.5–2.0:\n- 0.8–1.0: natural conversational\n- 1.1–1.3: punchy ad delivery\n- 1.4+: rushed, clips consonants\n- 0.6–0.7: slow, weighty (good for dramatic lines)\n\nDefault 1.0 unless the line specifically calls for slower or faster cadence.\n\n## Avatar prompt usage\n\nThe `prompt` parameter on avatar-v2 (standard + pro) is **scene context**, not motion direction. The mouth animation comes from the audio — the prompt sets ambiance, lighting, micro-expression.\n\nGood:\n- `Soft rim light, warm office, gentle confident smile between sentences.`\n- `Cool blue evening light through a window, focused intent expression.`\n\nBad (the model ignores motion verbs):\n- ❌ `She turns her head, raises an eyebrow, then speaks.`\n- ❌ `Hand gestures while talking.`\n\nDefault `\".\"` is fine if you have nothing useful to add.\n\n## Tier selection — avatar standard vs pro\n\n**Use standard** when:\n- Drafts, A/B testing voices, internal review reels\n- Wide / medium shots where face isn't the focal point\n- Cost matters more than micro-expression fidelity\n\n**Use pro** when:\n- Final ads where the avatar's face fills the screen\n- The character is named / branded — identity drift kills the take\n- You're already paying tens of credits for the surrounding video pipeline\n\nDon't default to pro. The ~15-credit delta per take adds up across iteration.\n\n## Common failure modes\n\n| Symptom | Likely cause | Fix |\n|---------|--------------|-----|\n| Lip movement looks \"rubber\" / disconnected | Source face <60% of frame | Re-crop the still tighter |\n| Voice doesn't match character age/gender | Default voice id used | Pick from voice catalog |\n| Output truncated mid-word | TTS text >120 chars | Shorten or chain two takes |\n| Garbled mouth on uploaded audio | Background music / multi-voice | Use clean dialogue-only audio |\n| \"Voice id not found\" 422 | Hit `reader_en_m-v1` | Switch to `oversea_male1` |\n| Avatar eyes drift / cross | Source had closed/angled eyes | Pick a frame with neutral open eyes |\n| Generation completes but lips don't move | Profile shot / face >70° off-axis | Use a near-frontal portrait |\n\n## Cost discipline\n\n- Video re-dub at ~4 credits is the cheapest dialogue iteration in the entire Slates stack — use it for voice A/B testing\n- Avatar standard at ~14 credits is fine for medium use\n- Avatar pro at ~29 credits trips the confirm gate — explicit user OK required every time\n- All 5s. There is no shorter option.\n\n## Workflow patterns\n\n**Voice A/B test (cheap):**\n1. Generate one base talking-head video clip with Veo or Seedance (~40 credits)\n2. Run `slates_generate_lip_sync` with `sourceType: 'video'` against 3–5 different `ttsVoice` values\n3. Total cost: ~40 + (5 × ~4) ≈ 60 credits to compare voices\n\n**Brand avatar from a single portrait:**\n1. Generate or upload the hero portrait (face fills frame, eyes open, neutral mouth)\n2. Avatar standard for first-pass dialogue takes\n3. Avatar pro only on the final selected take\n\n**Avoid:**\n- Avatar pro on first iteration (waste — facial fidelity isn't visible until you've locked the line)\n- TTS for final ads (production should use real voice or cloned voice — the upload flow)\n- Uploading raw recordings — clean noise + level the file first, lip detection is sensitive\n\n## Confirm gate: cost + codes, no inline preview\n\nLip-sync is mechanical — the model re-syncs the chosen source to the chosen audio. The confirm response carries the source asset's code so you can announce it in chat.\n\n- ✅ \"Lip-syncing **IMG-A12 — Founder Portrait** to the new line. ~29 credits on avatar-pro. Confirm?\"\n- ❌ \"Using the founder image...\" (which? Three exist.)\n\nDon't second-guess the source. If the output is wrong, iterate on source choice or audio, not on a refinement prompt (there isn't one).\n\n## Sources\n\n- [fal.ai — Kling LipSync API](https://fal.ai/models/fal-ai/kling-video/lipsync/text-to-video/api)\n- [fal.ai — AI Avatar v2 Standard](https://fal.ai/models/fal-ai/kling-video/ai-avatar/v2/standard/api)\n- [fal.ai — AI Avatar v2 Pro](https://fal.ai/models/fal-ai/kling-video/ai-avatar/v2/pro/api)\n",
     "slates-prompting-ltx-2-5": "---\nname: slates-prompting-ltx-2-5\ndescription: How to prompt LTX-2.5 and LTX-2.5 Pro. Read before calling slates_generate_video with model ltx-2-5 or ltx-2-5-pro. LTX scores the picture on the same pass that draws it, so SOUND IS THE FIRST THING YOU WRITE — Lightricks ranks the prompt sound, camera, character detail, shot type and scene, then scene dressing, all in one flowing paragraph. It is also the catalogue's native MULTISHOT seat: one generation carries two to four connected shots holding character, light and voice across the cuts. Base ltx-2-5 is the distilled build — 720p/1080p/1440p/4K, clips of 6 to 20 seconds in EVEN steps, and the cheapest native 1080p second in Slates; ltx-2-5-pro is the full diffusion build and is NOT a superset, reaching only 1080p and 10 seconds for about a third more money. Three hazards live here: durations are even numbers only from six (there is no 5s or 7s clip), the model has NO reference endpoint at all so identity references are unavailable, and any sound not anchored to something in frame gets invented for you.\n---\n\n# LTX-2.5 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — LTX-2.5.** It scores the picture on the same pass that draws it, so SOUND IS THE FIRST THING YOU WRITE. Lightricks' own priority order: sound, camera, character detail, shot type and scene, then scene dressing. One flowing paragraph, not labelled sections. When a prompt sprawls, cut from the bottom.\n\n**The five levers**\n1. **Lead with sound, and anchor every sound to something in frame.** The test is \"visible, or at least locatable\" — `the rope creaks against the cleat`, `the hull knocking hollow against the fenders`, `rain on the awning`. A distant whistle is fine IF you have named the post it comes from.\n2. **Camera second**, because framing decides the visual weight of the shot — `low camera at the gunwale`, `slow drift right`, `static medium behind the counter`.\n3. **Character detail as physical ACTION**, not as adjectives about a person — `she braces a boot on the rail and hauls`, `his hands counting notes`.\n4. **It is the native MULTISHOT seat** — one generation carries two to four connected shots holding character, light and voice across the cuts. Write the cuts.\n5. **Quote dialogue and name the language and accent** — `in English with a slight German accent` — `\"We should not have come back,\" in English with a slight German accent.`\n\n**Examples**\n- `The rope creaks against the cleat as she leans back, gulls calling somewhere off the port bow, the hull knocking hollow against the fenders. Low camera at the gunwale, slow drift right. She braces a boot on the rail and hauls, twice, then stops.`\n- `A till drawer bangs shut, a fan ticks against its cage, rain on the awning outside. Static medium behind the counter, then cut to a close-up of his hands counting notes, then cut wide as he looks up at the door.`\n\n**Hard constraint:** durations are EVEN numbers from six — there is no 5s or 7s clip. There is NO reference endpoint at all, so identity references are unavailable; use MiniMax H3 or Kling when a character must hold across shots. Any sound not anchored to something in frame gets invented for you. And never write mood adjectives as sound: \"tense atmosphere\", \"a sense of dread\" and \"ominous ambience\" produce nothing usable — the fix is one more moving object in frame with a sound attached to it.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** — mood adjectives standing in for sound produce nothing usable:\n- `tense atmosphere`, `a sense of dread`, `ominous ambience`, `eerie silence`\n- an unanchored sound: name the thing in frame it comes from, or cut it\n<!-- @banned:end -->\n\nLTX-2.5 generates picture and sound **in a single pass**, with a Gemma-4 12B text encoder reading\none flowing paragraph. That single fact drives everything below: the prompt is not a shot\ndescription with audio bolted on, it is **a scene where the sound is load-bearing** — and\nLightricks' own priority order puts sound first, ahead of the camera.\n\nTwo seats, and the naming is a trap:\n\n| | `ltx-2-5` (base) | `ltx-2-5-pro` |\n|---|---|---|\n| Build | Distilled, 8-step | Full diffusion (\"Diffusion Fidelity Rendering\") |\n| Resolutions | 720p / 1080p / **1440p** / 4K | 720p / 1080p |\n| Durations | 6–20s, even steps | 6 / 8 / 10s |\n| Price | $0.09–$0.30 per second | $0.12–$0.17 per second |\n| Reach for it when | iterating, long takes, 4K delivery, batch volume | one dense final render inside 1080p and 10s |\n\n**Pro is not \"base plus more.\"** It buys picture quality on a *narrower* envelope — it cannot make\na 1440p frame and it cannot make a 12-second clip. Reaching for it out of habit costs a third more\n*and* takes away the reach.\n\n---\n\n## 1. The six parts, in priority order, in one paragraph\n\nLightricks ranks the elements of an LTX prompt like this. When a prompt sprawls, **cut from the\nbottom.**\n\n1. **Sound** — highest priority; the model scores the picture as it draws it.\n2. **Camera** — framing decides visual weight and the feel of the shot.\n3. **Character detail** — expressed as physical action.\n4. **Shot type and scene** — the action itself.\n5. **Scene dressing** — the first thing to trim.\n\nWrite it as **one flowing paragraph**, not a list of labelled sections. LTX is not Seedance (eight\nengineering slots) and not H3 (three separate audio layers) — it wants continuous prose.\n\n---\n\n## 2. Sound: anchor it or it gets invented\n\n**Write the audio line last, then go back and check every cue has a source you could point at.**\nAnything unanchored, the model invents for you.\n\nThe test is **\"visible, or at least locatable.\"** A distant whistle is fine *if* you have named the\nmarshal's post it comes from. A \"distant whistle\" with nothing to attach to is a coin flip.\n\n> the rope creaks against the cleat as she leans back, gulls calling somewhere off the port bow,\n> the hull knocking hollow against the fenders\n\n**Never write mood adjectives as sound.** \"Tense atmosphere\", \"a sense of dread\" and \"ominous\nambience\" produce nothing usable. If a scene feels thin, the fix is **one more moving object in\nframe with a sound attached to it** — never another adjective.\n\n### Dialogue\n\nQuote it, and name the language and accent:\n\n> \"We should not have come back,\" in English with a slight German accent.\n\nTwo rules that decide whether the lip sync lands:\n\n- **Give the character a beat of stillness before they speak.** The sync needs something to lock\n  against; a character already mid-motion when the line starts drifts.\n- **Describe the beat structure** — when they look, how long they wait, when they speak, where they\n  look afterwards.\n\nSlates pins the frame rate at 25fps, which is also what Lightricks recommends for dialogue: at 50fps\nthe performance \"pulls toward a video look.\"\n\n---\n\n## 3. Character emotion is physical\n\nThe model renders actions. It does not render adjectives.\n\n| Instead of | Write |\n|---|---|\n| she looks anxious | her jaw sets, she turns the ring on her finger twice |\n| he seems exhausted | he blinks slowly and lets his shoulder take the doorframe |\n| a tense standoff | neither moves; his thumb finds the strap and stays there |\n\n---\n\n## 4. Multishot — the thing this model is uniquely for\n\n**One LTX generation can carry several connected shots**, holding character, environment, lighting,\nvoice and style across every cut. Nothing else in the catalogue does this natively; everywhere else\nyou generate separate clips and stitch them, and identity drifts between them.\n\n**Working range is two to four shots.** Three is the comfortable stopping point.\n\nAt **every** transition you must supply four things:\n\n1. **Name the edit in the prose** — \"hard cut\", \"dissolve\", \"match cut\".\n2. **Re-establish the shot completely** — scale, angle, lens and light all reset at a cut. A cut is\n   not a continuation.\n3. **Re-identify recurring characters by their original descriptor.** \"The woman in the bronze\n   gown\", never \"she\". Pronouns lose the character across a cut — this is the single most common\n   multishot failure.\n4. **State what the sound does at the cut.** Silence is not assumed; if the room tone should drop\n   out, say so.\n\nA shape that works:\n\n> Wide establishing shot of the workshop, dust in the window light, a lathe turning somewhere off\n> frame — hard cut — macro close-up of the brass fitting as it seats, the turning noise gone,\n> replaced by a single dry click — match cut — medium shot of the woman in the bronze gown stepping\n> back, the room tone returning underneath her.\n\n---\n\n## 5. Camera: write it, don't enumerate it\n\nfal exposes a `camera_motion` enum (dolly in/out/left/right, jib up/down, static, focus shift).\n**Slates does not surface it, deliberately** — and prose is the better instrument anyway:\n\n- **A written move can be tied to a specific moment.** \"A slow push-in that settles as she reaches\n  the door, then holds\" is not expressible as an enum value.\n- **For multishot it would be actively wrong** — one enum value would impose a single camera\n  behaviour on three shots that each want their own.\n\nSo name the lens, the framing, the move, and **the moment the move resolves**.\n\n---\n\n## 6. The hard constraints\n\n### Durations are even numbers only, starting at six\n\n**6, 8, 10, 12, 14, 16, 18, 20.** There is no 5-second LTX clip and no odd duration of any length.\nAsking for 7s is not a rounding matter — that generation does not exist.\n\n**And the long end is 1080p-and-below only.** At 1440p and 4K the ceiling drops to **6, 8 or 10**.\n\nfal's own default is `auto`, which lets the model pick the length from the described action.\n**Slates always sends an explicit length instead**, so what you choose is what you are billed for.\nChoose the length the beat needs.\n\n### Aspect ratios: 16:9 and 9:16, and nothing else\n\nThe narrowest set in the catalogue alongside Veo. Square, 4:5 and 21:9 are not available on this\nmodel at any resolution.\n\n### Frames, not references\n\nLTX takes a **start frame** and an **optional end frame** (which generates a transition between the\ntwo). It has **no reference-to-video endpoint at all** — no identity references, no style\nreferences, no environment references, no reference video, no reference audio.\n\n**For character consistency across separate shots, use MiniMax H3 or Kling.** Within a single LTX\ngeneration, use multishot instead — that is precisely the gap it fills.\n\nIn image-to-video, **do not cut away from the opening frame too early.** You have paid for that\nframe; let it play before the first move.\n\n### Do not ask for text on screen\n\nNeither the spelling nor its stability from frame to frame can be relied on. Signage, labels,\ncaptions and lower-thirds belong in post.\n\n---\n\n## 7. Audio is free here, and that changes the routing\n\nNative synchronised audio is **included at every resolution on both seats**, with no surcharge and\nno toggle that costs money — unlike Kling, where sound is a paid dimension. A 6-second 1080p LTX\nclip **with sound** is 39 credits.\n\nCombined with 1080p at $0.13/s — the cheapest native 1080p second in Slates — this makes LTX **the\ncoverage seat**: the one to reach for when the job is many takes rather than one hero shot, when a\nsequence needs its own sound, or when the credit budget is the binding constraint.\n\nRoute away from it when you need identity references (H3, Kling), a ratio other than 16:9 or 9:16\n(Seedance, Kling), or authored multi-layer audio direction (H3).\n",
     "slates-prompting-minimax-h3": "---\nname: slates-prompting-minimax-h3\ndescription: How to prompt MiniMax H3 and MiniMax H3 Max. Read before calling slates_generate_video with model minimax-h3 or minimax-h3-max. H3 is the only Slates video seat where AUDIO IS AUTHORED rather than toggled — synchronised dialogue, scene sound and an audience-only score are three separate sections of the prompt, generated in one pass — and the only one where a reference carries a DECLARED RELATIONSHIP (kept whole, partly kept, transferred onto a different subject, or a loose echo). Base minimax-h3 runs 480p/768p/2K/4K and reads 9 images + 3 video + 3 audio references; minimax-h3-max is fal's faster post-train, capped at 768p, and costs MORE than base H3 at 768p — a deliberate speed pick, never the default and never the cheap one; it still animates start and end frames, but it has no reference-to-video endpoint, so the omni-reference set is base-H3 only. Two hazards live here: reference images past the fifth cost 4 credits each on the base row, and audio written into the wrong section is dropped or duplicated.\n---\n\n# MiniMax H3 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — MiniMax H3.** The only seat where audio is AUTHORED rather than toggled: dialogue, scene sound and score are three separate sections of the prompt, generated in one pass, and putting a sound in the wrong section drops or doubles it.\n\n**The five levers**\n1. **Write the three audio layers separately** — `Scene sound:` for what is in the room, `Score:` for what only the audience hears, and the dialogue quoted inline. Section decides attribution.\n2. **Quote dialogue and name the language** — `says in English`, `speaks in Spanish`. Eleven languages are stably supported; the language is part of the instruction, not an afterthought.\n3. **Declare the reference RELATIONSHIP**, which no other seat has: `kept whole`, `partly kept`, `transferred`, or `a loose echo`. An undeclared reference is a guess.\n4. **Give a beat of stillness before a line** — `sits still for a beat, then looks up`. The sync needs something to lock against; a character already mid-motion when the line starts drifts.\n5. **Describe the beat structure** — `waits`, `then speaks`, `under the last three seconds`. H3 is a timeline, so write one.\n\n**Examples**\n- `A woman sits still at a kitchen table for a beat, then looks up. She says in English, \"You said Tuesday.\" Scene sound: a fridge hum, a spoon set down on formica. Score: none.`\n- `Two mechanics either side of an open bonnet. The younger one wipes his hands, waits, then speaks in Spanish, \"No es el alternador.\" Scene sound: a socket wrench, a radio two bays over. Score: a low sustained cello under the last three seconds, audience only.`\n\n**Hard constraint:** the two seats differ in what the ENDPOINT accepts, not in grammar. Base H3 reaches 2K/4K and takes references; `minimax-h3-max` is capped at 768p, has NO reference endpoint at all (start and end frames still work), and costs MORE at the tier they share — it is a speed pick, never the cheap one. H3's top two resolution tiers are UPSCALES of the native render: judge at native. Reference images past the fifth are a paid dimension of the cost key — declare the count when quoting.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use:**\n- a sound written into the wrong audio section — it is dropped, doubled, or attributed to the wrong layer\n- `background music` as a bare instruction: the score is its own authored layer, audience-only, and it is named as such\n- an undeclared reference relationship — say kept whole, partly kept, transferred, or a loose echo\n<!-- @banned:end -->\n\nH3 is an **omni transformer**: it generates picture and sound in the same pass, at 24fps with\n32kHz stereo, 5–15 seconds, in 11 stably-supported languages (Arabic, Chinese, English, French,\nGerman, Italian, Japanese, Korean, Portuguese, Russian, Spanish). That single fact drives\neverything below — the prompt is not a shot description with sound bolted on, it is a **timeline\nwith three audio layers you author separately**.\n\n**Two seats, one grammar.** Everything in this file applies to both. They differ only in what the\nendpoint accepts:\n\n| | `minimax-h3` | `minimax-h3-max` |\n|---|---|---|\n| Resolution | 480p / 768p / **2K / 4K** | 480p / 768p |\n| References | 9 images + 3 video + 3 audio (12 files) | **none — no reference endpoint exists.** Frames still work; see the row below |\n| Frames | start and/or end | start and/or end |\n| Price at 768p | **$0.060/s** | $0.080/s |\n| Why pick it | resolution, references, and the cheaper second | **speed** — a 5s 768p clip in **4.8s** vs **57s** (measured) |\n\n**Max is the premium seat, not the budget one.** It is 33% dearer at the one tier they share and it\ntops out lower. Route there when a fast turnaround on a text-to-video or start-frame shot is worth\npaying for; route to base H3 for anything needing resolution, references, or the same tier cheaper.\n\n**The speed is measured, not claimed** (2026-08-27, same prompt and params on both rows): a 5-second\n768p text-to-video finished in **4.8 seconds** on Max against **57 seconds** on base H3 — roughly\n**12x**, queue to finished file. fal advertises \"under 3 seconds\"; the literal claim did not hold at\n4.8s wall-clock, but the order of magnitude did. For iteration loops and client-present work that gap\nis the entire reason the seat exists.\n\n---\n\n## The one thing that makes H3 different: audio is a THREE-LAYER instruction\n\nEvery other video seat treats sound as on or off. H3 splits it, and the split is enforced by where\nyou write each thing. Get the section wrong and the sound is dropped, doubled, or attributed to the\nwrong source.\n\n| Layer | What belongs in it | Where it goes |\n|---|---|---|\n| **Synchronised events** | dialogue, singing, and any sound tied to a specific shot or action | the **body** of the prompt, on the beat it lands |\n| **Scene sound** | ambience and physical sounds that run across the whole clip — room tone, rain, traffic, a ventilation hum | the **soundscape** section |\n| **Score** | music the characters cannot hear; audience-only | the **music** section |\n\n**Three rules, all from MiniMax's own guide:**\n\n1. **Dialogue and singing NEVER go in the soundscape section.** They are synchronised events; they\n   belong in the body, at the moment they happen.\n2. **Diegetic music — music the characters can hear** (a radio in the scene, a busker) — also\n   belongs in the **body**, not in the score section. The score section is audience-only.\n3. **Write the score in instrumental terms, not mood words.** Name the instruments, the tempo, and\n   how it develops. *\"A restrained solo-piano score at a slow tempo, sustained low cello underneath,\n   no swell\"* — not *\"emotional music\"*.\n\nUse **N/A** for a section only when silence or absence is genuinely what the shot wants. An empty\nscore section is a real choice; a vague one is a wasted layer.\n\n### The shape, in the one prompt field\n\nSlates sends one prompt string, so write the three layers as labelled paragraphs in this order:\n\n```\n[Shot 1] Live-action, cinematic. A medium-wide shot frames a baker opening the shutters of a\nsmall street bakery before sunrise. The camera pushes in with small amplitude at slow speed as\nthe middle-aged baker with a calm, slightly raspy voice places a fresh loaf on the counter and\nsays: \"First batch of the morning.\" [Shot 2] At 00:05.000, the camera cuts to a close-up of\nsteam rising from the sliced bread while his final words carry over from the previous shot.\n\nSoundscape: wooden shutters scrape open over a quiet street, trays clink softly inside, a\ndoorbell rings once, then light footsteps and the crisp sound of bread being sliced.\n\nScore: a soft acoustic-guitar pattern at a moderate tempo, joined by sparse upright-bass notes,\ngentle fade at the end.\n```\n\n**Body target: 350–500 words** for a reference-carrying shot. Dialogue-heavy content prioritises\nfitting the complete spoken timeline over hitting a word count.\n\n🚨 **Slates disables the provider's prompt expander.** H3's API can rewrite your prompt before\ngeneration; Slates turns that off, because a model rewriting the user's words invisibly is banned\noutright (prompt transparency: what the composer shows is what the model gets). The practical\nconsequence is on you: **nothing will pad a thin prompt.** Write the whole body.\n\n---\n\n## Shots and timing\n\nThe first shot carries **no timestamp**. Every later shot opens with the bracket and a cut time\nthat increases and stays inside the clip length:\n\n```\n[Shot 2] At 00:03.500, the camera cuts to ...\n```\n\nTransition verbs the model knows: **cuts to · transitions to · changes to · switches to**.\n\n**Dialogue that continues across a cut** needs the continuity said out loud — *\"his final words\ncarry over from the previous shot\"* — or the line restarts. **Speech that ends abruptly** should be\ndescribed as cut off rather than trailed off.\n\n---\n\n## Camera — write the move into the sentence\n\nThe model has a named motion vocabulary:\n\n> Zoom In / Zoom Out · Push In / Pull Out · Pan Left / Pan Right · Truck Left / Truck Right ·\n> Tilt Up / Tilt Down · Pedestal Up / Pedestal Down · Arc Shot · Tracking Shot · Static Shot ·\n> Shake Slightly / Shake Strongly · POV · Roll Clockwise / Roll Counterclockwise\n\nModify with **amplitude** (`with small amplitude` / `with large amplitude`) and **speed**\n(`at slow speed` / `at fast speed`).\n\n🚨 **Integrate the motion into the sentence — never stack labels.** MiniMax's own example:\n*\"The camera pushes in with small amplitude at slow speed toward the folded letter in her hands.\"*\nNot *\"Push In. Small amplitude. Slow.\"*\n\n---\n\n## Speakers and dialogue\n\nGive each speaking character a stable identity in the prose and keep it: describe the voice once\n(*\"a young woman with a quiet, breathy voice\"*), then refer back to the same description at every\nline. Identification, delivery and action sit **outside** the quoted line; the line itself is only\nthe words.\n\n```\nThe young woman with a quiet, breathy voice says: \"I get off at the next station.\"\n```\n\n**Voiceover** needs two things — the phrase *\"says in an off-screen voiceover\"* **and** an explicit\nstatement that the lips stay closed. Without the second half the model animates a mouth.\n\n```\nThe man says in an off-screen voiceover: \"I still remember that road.\" — his lips remain\ncompletely closed.\n```\n\n**On-screen text** — signs, banners, labels, subtitles, neon — goes in double quotes with the\noriginal wording preserved exactly: *A red neon sign reading \"Open Late\" glows above the doorway.*\n\n---\n\n## References — H3's real differentiator is the declared RELATIONSHIP\n\n*(Base `minimax-h3` only. `minimax-h3-max` has no reference endpoint — Slates refuses references\non that row rather than dropping them silently.)*\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### Cite references by number — Slates already does it for you\n\nH3 on fal takes references as **typed slots** and expects the prompt to name them by modality and\norder: **`image 1`, `image 2`, `video 1`, `audio 1`**. That is exactly what the Slates composer\nemits from your `@mentions` and `#tags` (`Marcus (image 1) in the workshop (image 2)`), in the\nexact order it sends them.\n\n🚨 **Do NOT hand-write angle-bracket reference tags.** MiniMax's own model-card grammar uses\n`<Subject N>` / `<Picture N>` / `<Video N>` / `<Audio N>` labels; the fal endpoints Slates calls do\nnot — they build the binding from the typed slots and ask for plain numbered prose. Typing the tags\nyourself puts literal angle brackets in the prompt the model reads.\n\n### State how much of each reference survives\n\nThis is the lever no other model in the catalogue gives you. Say, in plain words, what each\nreference is FOR and how much of it should carry through:\n\n| Intent | Say something like |\n|---|---|\n| Keep it whole | *\"Keep the woman in image 1 exactly as she appears — hair, cardigan, necklace.\"* |\n| Keep part of it | *\"Use the café in image 2 for the brick wall and the sofa; the lighting is late evening, not daylight.\"* |\n| **Move a trait onto someone else** | *\"Give the man in image 3 the weathered leather texture of the jacket in image 4.\"* |\n| Loose echo | *\"Match the general palette and grain of image 5; nothing else from it.\"* |\n\nThe third row is the one with no equivalent anywhere else in Slates: **transferring a characteristic\nonto a different subject** is a first-class thing H3 understands. Reach for H3 when that is the job.\n\n**Audio references** bind a voice or a texture without copying the words. Say which speaker an\naudio reference is for (*\"the woman in image 1 speaks in the voice timbre of audio 1\"*), and when\nyou are referencing only the timbre, **do not carry the reference clip's original dialogue into your\nprompt** — write the new line. When you genuinely want the same words re-performed, quote them\nexactly and say so.\n\n**An audio reference cannot travel alone** — H3 refuses a reference set that is audio only. Pair it\nwith at least one image or video reference.\n\n### 💸 Reference images past the fifth cost 4 credits each\n\nThe first **5** reference images are free. Each additional image — the model takes **9** — adds\n**4 credits** to the generation, at every resolution and every length. Four extra images on a 10s\n768p clip add 16 credits to a 30-credit generation: **more than half again**, for references that\noften make the output worse rather than better (see the 2–4 rule above).\n\nAttach the references the shot needs, not the ceiling. Call\n`slates_estimate_generation_cost` with `referenceImages` set to the real count before a\nreference-heavy job — a quote that omits it under-reports the bill.\n\n---\n\n## Frames\n\n`minimax-h3` and `minimax-h3-max` both take a **start frame**, an **end frame**, or both. With an\nend frame, land it explicitly: describe the final pose, spacing and composition as the thing the\nshot **settles into** at the end, rather than hoping the model finds it.\n\n> *\"…she rotates the handle into the final angle and settles into the pose, spacing and composition\n> of image 2 at the end of the shot.\"*\n\n**Frames and references are mutually exclusive** on both rows — they are different endpoints, and\nthe reference endpoint has no frame slots at all. Slates refuses the combination rather than\ndropping one side.\n\n---\n\n## Cost discipline\n\n| Combination | Credits |\n|---|---:|\n| `minimax-h3` · 768p · 5s | 15 |\n| `minimax-h3` · 768p · 10s | 30 |\n| `minimax-h3` · 2K · 10s | 65 |\n| `minimax-h3` · 4K · 10s | 80 |\n| `minimax-h3-max` · 768p · 10s | 40 |\n| every reference image past the fifth | **+4** |\n\n**768p is the default for a reason.** It is the tier the model natively generates.\n\n🚨 **2K and 4K are UPSCALES of a 768p render, not larger generations.** fal's own schema says so:\n*\"480P and 768P are native generation modes; 2K and 4K upscale a 768P base result.\"* The upscaler\n(H3-Regenerate-2K) is a separate stage bolted onto a finished take — it can enlarge detail but it\ncannot add information.\n\n**In our own test (2026-08-27, same prompt, same seed) the 2K pass came back with MORE artifacting\nthan the 768p original it was built from**, while costing 33 credits for a 5-second take against 15,\nand taking nearly twice as long to return. One shot, so treat it as a warning rather than a law —\nbut the mechanism explains it, and the burden of proof is on 2K.\n\n**So: generate at 768p and judge it at 768p.** Reach for 2K or 4K only when a delivery spec demands\nthe pixels, and expect to be paying for size rather than quality — a post-production upscale from a\nclean 768p master is very often the better result. **4K video is Pro-only** (the server returns\n`PRO_REQUIRED` for a base account); 2K is open to every tier.\n\n---\n\n## Quick checklist\n\n- Body written as a timeline, first shot untimestamped, later shots on `[Shot N] At MM:SS.mmm`.\n- Camera motion written **into** a sentence with amplitude and speed.\n- Dialogue and diegetic music in the body; ambience in the soundscape section; audience-only score\n  in the score section, described by instrument and tempo.\n- Voiceover carries both the off-screen phrase and the closed-lips statement.\n- References cited as `image 1` / `video 1` / `audio 1`, each with a stated job and a stated degree\n  of retention. No angle-bracket tags.\n- Reference count is deliberate — you are paying 4 credits for each one past the fifth.\n- Frames **or** references, never both.\n- The prompt is the prompt: no expander will fill it out for you.\n",
     "slates-prompting-motion-transfer": "---\nname: slates-prompting-motion-transfer\ndescription: How to set up motion transfer — Kling Motion Control only (std and pro tiers, 5-second outputs). Read before calling slates_generate_motion_transfer. Reference image (character) + driving video (motion source) → new video of the character performing the motion. Asset selection rules, character_orientation, tiers, and prompt usage. Also covers the Seedance alternative, which is a normal video generation rather than a mode of this tool.\n---\n\n# Motion transfer — setup guide\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Motion transfer (Kling Motion Control only).** A target IMAGE (your character) plus a source VIDEO (the motion) produces your character performing that motion. Always 5 seconds.\n\n**The five levers**\n1. **The target image must show body proportions clearly** and the character must occupy more than about 5% of the frame. A tiny figure in a wide shot has nothing to drive.\n2. **Single character in the target.** A group image breaks the identity anchor.\n3. **Choose `characterOrientation` on purpose** — `video` takes the source clip's framing, `image` preserves the portrait's. It is the most-missed choice here.\n\n4. **The prompt is atmosphere only** — `Soft afternoon sunlight, dust motes in the air, vintage warm color grade.` Motion verbs are ignored; the motion is already in the driving video.\n5. **Pick the best 5 seconds of the source up front**, and write only atmosphere: `soft afternoon sunlight`, `vintage warm color grade`, `clean studio backdrop`. The output is 5s regardless, so a long driving clip just wastes the choice.\n\n**Examples**\n- `Soft afternoon sunlight, dust motes in the air, vintage warm color grade.`\n- `Clean studio backdrop, sharp focus on the character.` (Or leave it empty.)\n\n**Hard constraint:** cartoon driving videos fail, and a cropped or partial target character drifts. std is fine while the motion-and-framing combination is still moving; switch to pro once it is locked. For a REGENERATED shot instead — physical contact, cloth and hair, camera motion, native audio — that is a normal Seedance video generation with the driving clip as a video reference, not a mode of this tool.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** — the motion is already in the driving video, so motion verbs are ignored:\n- `spins faster`, `jumps higher`, `add more energy`, `moves quicker`\n<!-- @banned:end -->\n\nTake a still **target image** (your character) and a **source video** (the motion you want), produce a new video of your character performing the source video's motion. **This tool is Kling-only** — it wraps Kling Motion Control and nothing else.\n\n| Tier | Cost | Use case |\n|------|-----------|----------|\n| Kling std (`kling-mc-std-5s`) | ~32 credits / 5s | General motion transfer, budget lane |\n| Kling pro (`kling-mc-pro-5s`) | ~42 credits / 5s | Cleaner anatomy, better identity preservation |\n\nBoth tiers trip the confirm gate. User OK required every time. (Prices are approximate — `slates_estimate_generation_cost` returns the exact credit total.)\n\n## Want Seedance instead? That is a video generation, not a mode here\n\nKling MC retargets a skeleton onto a finished image; Seedance *generates* the shot with the motion as a conditioning input — the difference shows on fast choreography, physical contact, cloth/hair, and camera motion, and the output carries native audio. **It is not an engine switch on this tool.** Run a normal `slates_generate_video` on `seedance-2` with the driving clip attached as a video reference and the character image as an ingredient, then write the prompt yourself:\n\n```\nThe character from image 1 performs the exact motion, choreography, and camera\nmovement from video 1. Preserve the character's identity, appearance, and outfit.\n```\n\nThat is the same endpoint the old `motionModel=seedance-2` branch called — it just wrote that sentence for you, invisibly. Add style/setting/camera direction freely; Seedance re-generates the whole shot.\n\n- **Driving clip must be 2–15s** (all providers cap reference video at 15s). Longer clips: trim first, or use Kling MC (`characterOrientation: 'video'` takes up to 30s).\n- **Billing = combined input+output seconds** (the vref keys). The server probes the clip and corrects an understated key — quote via the confirm gate before spending.\n- **Faces route through the face cascade**: `seedanceFace` for a character, `[REAL_FACE_DETECTED]` → confirm consent → `seedanceRealFace=true, realFaceConsent=true` (premium realface vref pricing).\n- `characterOrientation` has no Seedance equivalent; framing follows the prompt + `aspectRatio`.\n\nEverything below is about the Kling tool.\n\n## Inputs\n\n- `sourceVideoAssetId` — driving video. **Must be a realistic human** with clear proportions. Anime/cartoon/CG driving videos fail.\n- `targetImageAssetId` — character to be animated. Can be any style (cartoon, anime, realistic, painted).\n- Both must already exist as assets in the project. Use `slates_list_assets` to find them or upload first.\n\n## Source video constraints\n\n- Realistic human (not animated, not CG)\n- Entire body OR upper body visible — head must not be obstructed\n- Subject occupies a clear share of the frame\n- Single primary subject. Multi-person driving videos confuse the motion anchor.\n- Clean motion — choppy / cut-edited driving videos produce jittery output\n\nGood driving video sources:\n- Reference dance footage with one subject\n- Walking / gesture / posing clips\n- Talking-head footage when paired with character_orientation: 'video'\n\nBad driving video sources:\n- Music videos with multi-shot edits\n- Anime / animation clips\n- Heavily stylized footage with smoke / particles obscuring the body\n- Footage where the subject's head leaves frame mid-clip\n\n## Target image constraints\n\n- Character body proportions clearly visible\n- Character occupies >5% of image area (not a tiny figure in a wide shot)\n- Single character. Group images break the identity anchor.\n- Any artistic style works — cartoon, anime, painted, realistic, 3D render\n\nAvoid:\n- Extreme close-up of just the face (no body to drive)\n- Character partially cropped at the waist when the driving video is full-body\n- Multiple characters\n\n## character_orientation — the most-missed choice\n\nThis single parameter changes the output dramatically. Pick deliberately.\n\n| Value | Output framing | Max source duration | Best for |\n|-------|----------------|---------------------|----------|\n| `video` | Matches driving video framing | Up to 30s source | Complex full-body motion (dance, action, athletics) |\n| `image` | Matches target image framing | Up to 10s source | Camera moves, simpler motion, preserving original composition |\n\n**Default `video`** when the driving video has the look you want (most cases).\n\nSwitch to `image` when the target image's composition is the brand asset and the motion is secondary (e.g., a hero shot of a character that needs subtle gesture, not a full performance).\n\n## Tier choice — std vs pro\n\n**std (~32 credits)** for:\n- Drafts, motion exploration, blocking\n- Group scenes where the character isn't a hero shot\n- When the budget is tight and the motion is the focus\n\n**pro (~42 credits)** for:\n- Final hero takes\n- Branded characters where identity drift = unacceptable\n- Anatomically complex motion (limbs crossing, fast direction changes)\n- Anime / cartoon target images — pro handles non-realistic styles better\n\nDon't default to pro. The ~10-credit delta compounds fast across iteration.\n\n## Prompt usage (optional)\n\nThe `prompt` field is **scene/style refinement**, not motion direction. The motion comes from the driving video — the prompt sets ambiance, lighting, additional detail.\n\nGood:\n- `Soft afternoon sunlight, dust motes in the air, vintage warm color grade.`\n- `Clean studio backdrop, sharp focus on the character.`\n\nBad (model ignores motion verbs — they're already in the driving video):\n- ❌ `She spins faster and jumps higher.`\n- ❌ `Add more energy to the dance.`\n\nLeave it empty if you don't have a specific atmospheric note.\n\n## Common failure modes\n\n| Symptom | Likely cause | Fix |\n|---------|--------------|-----|\n| Limbs distort / extra fingers | std tier, complex motion | Switch to pro |\n| Character identity drifts | Target image cropped too tight | Use a fuller-body target |\n| Output looks \"stuck\" / minimal motion | Driving video subject too small in frame | Pick a driving video where the subject fills more of the frame |\n| Cartoon target turns realistic | std tier on stylized art | Switch to pro — handles non-realistic styles better |\n| Garbled output entirely | Anime / CG driving video | Use realistic human driving footage |\n| Wrong framing on output | character_orientation set wrong | Try the other value |\n| Background bleeds through character | Target image had complex background | Use a target with cleaner background separation |\n\n## Workflow patterns\n\n**Reference dance to brand character:**\n1. Generate or upload the brand character as a still image (clean background, full body, single subject)\n2. Find driving footage — a clean reference video of the dance you want\n3. Upload both as project assets\n4. Run motion transfer with `motionModel: 'kling-mc-pro'`, `characterOrientation: 'video'`\n5. Total cost: ~42 credits per 5s take\n\n**Subtle motion on a hero portrait:**\n1. Use the locked hero portrait as the target image\n2. Pick a driving video with subtle gesture (head turn, slight posture shift)\n3. `characterOrientation: 'image'` to preserve the portrait's framing\n4. std tier is fine for this case — motion isn't dramatic\n\n**Avoid:**\n- Pro tier on first iteration — waste, switch to it once the motion + framing combo is locked\n- Cartoon driving videos — guaranteed failure\n- Cropped or partial target characters — identity will drift\n- Long driving videos when output is 5s — pick the best 5s of the source upfront\n\n## Cost discipline\n\n- 5 seconds, no shorter option\n- Both tiers trip the confirm gate — every call needs explicit user OK\n- Iteration is expensive: 4 takes at pro ≈ 168 credits. Lock framing + driving video before tier-up to pro.\n- Always run a single std take first to validate the motion + framing combo before committing to pro\n\n## Confirm gate: cost + codes, no inline preview\n\nMotion transfer is mechanical — the model deterministically applies source motion to target image. Both tiers trip the confirm gate; the response includes the asset codes for source and target so you can announce them in chat.\n\n- ✅ \"Transferring motion from **VID-V3** onto **IMG-A12 — Detective Closeup**. ~42 credits, confirm?\"\n- ❌ \"Using the walk video and the detective image...\" (multiple of each in the project.)\n\nDon't second-guess the assets the user picked — the model executes the transfer. If the output is wrong, iterate on motion source or target choice, not on a refinement prompt.\n\n## Sources\n\n- [fal.ai — Kling Motion Control V3 Standard](https://fal.ai/models/fal-ai/kling-video/v3/standard/motion-control)\n- [fal.ai — Kling Motion Control V3 Pro](https://fal.ai/models/fal-ai/kling-video/v3/pro/motion-control)\n",
-    "slates-prompting-nano-banana-2": "---\nname: slates-prompting-nano-banana-2\ndescription: How to write prompts that produce cinematic, photorealistic results from Nano Banana 2 (Google Gemini 3.1 Flash Image, accessed via fal-ai/nano-banana-2). Read this before calling slates_generate_image when the user wants film-quality, real-world, or cinematic output. Skip for stylized / illustrated / cartoon work — the rules differ.\n---\n\n# Nano Banana 2 — cinematic & photorealistic prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Nano Banana 2 (Gemini 3.1 Flash Image).** Brief it like a creative director, not a tag list. Structure: `Film still from [director] [genre]. Shot on [camera] with [lens]. [Subject and action]. [3-5 specific visual details]. [Lighting — direction + quality]. [Color palette]. [Film stock]. [1-2 word tone].`\n\n**The five levers**\n1. **Named lens + aperture** beats \"shallow depth of field\" — `85mm f/1.4`, `135mm f/2.8` (the cheat code for skin), `Panavision anamorphic`, `400mm telephoto`.\n2. **Light by direction and quality**, never \"good lighting\" — `hard sidelight from a single window, deep falloff`, `overcast north light`, `practical tungsten spill`.\n3. **A named film stock or sensor** carries a whole palette — `Kodak Portra 400`, `Cinestill 800T`, `ARRI Alexa 65`.\n4. **Composition as a shot** — `low angle`, `aerial view`, `rule of thirds with the subject camera-left`, `foreground occlusion`.\n5. **Positive framing only.** Describe what is there. \"Empty street\", never \"no cars\"; \"unstaged documentary photography\", never \"not anime\".\n\n**Examples**\n- `Film still from a Denis Villeneuve thriller. Shot on ARRI Alexa 65, 85mm f/1.4. A woman in a charcoal wool coat stands at a rain-slick bus stop, breath visible. Hard sodium light from a single overhead lamp, deep falloff into blue night. Kodak Vision3 500T. Isolated.`\n- `Editorial still life on seamless bone paper. 100mm macro, f/8. A cracked ceramic bowl holding three figs. Soft north light from camera-left, one gentle shadow. Muted earth palette. Portra 400 grain. Quiet.`\n\n**Hard constraint:** there is no `negativePrompt` field. Suppress by reframing positively, or inline `without` / `free of`. Knowledge cutoff January 2025 — anything later needs reference images.\n<!-- @card:end -->\n\nNano Banana 2 is **Gemini 3.1 Flash Image**.<!-- slates-only --> It is the default model behind `slates_generate_image` — the op also exposes `flux-2-max` and `seedream-5-lite`, each with its own prompting skill.<!-- /slates-only --> It is **not** Gemini 3 Pro Image; that is Nano Banana **Pro** (`nano-banana-pro`), a separate model with its own seat.<!-- slates-only --> Verified against the runtime slug map in `slate/src/main/api/google.ts`.<!-- /slates-only --> NB2 is a language model that outputs pixels — brief it like a creative director, not like a Stable-Diffusion tag-soup tool. The single biggest lever for realism: **specificity that mimics how real photographers and cinematographers describe their work**.\n\nKnowledge cutoff: January 2025. Anything after needs explicit reference images.\n\n## Google's 4 official rules (verbatim)\n\n1. **Be specific.** Provide concrete details on subject, lighting, and composition.\n2. **Use positive framing.** Describe what you want, not what you don't want.\n3. **Control the camera.** Use photographic and cinematic terms like \"low angle\" and \"aerial view.\"\n4. **Iterate.** Refine images with follow-up prompts in a conversational manner.\n\n## Official prompt formula\n\n```\n[Subject] + [Action] + [Location/context] + [Composition] + [Style]\n```\n\nFor the cinematic / photoreal use case, expand to:\n\n```\nFilm still from [DIRECTOR] [GENRE]. Shot on [CAMERA] with [LENS]. [SUBJECT and action]. [3-5 specific visual details]. [LIGHTING — direction + quality]. [COLOR PALETTE]. [FILM STOCK or sensor language]. [1-2 word emotional tone].\n```\n\n## Photorealism positives — what consistently works\n\n> ⚠️ **This vocabulary is an IMAGE-model lever and a video-model anti-pattern — do not carry it across.**\n> Named lenses, apertures, film stocks and camera bodies (`85mm f/1.4`, `Kodak Portra 400`, `ARRI Alexa 65`) are correct and encouraged **here**. They are a **Seedance anti-pattern**: ByteDance's own guide uses shot sizes, camera moves, pacing words and its image-quality vocabulary throughout, and never once mentions fps, shutter angle, f-stop, or lens millimetres.\n> The leak happens in one specific way — you write an NB2 start frame, then write the video prompt to animate it and carry the look description straight across. **Translate instead of copying:** `85mm f/1.4, Portra 400` → `close-up, shallow depth of field, warm natural colors, cinematic texture, film-grain texture`. Full rule and the receipts: `slates-prompting-seedance` (Part 3, \"Don't cross-pollinate image-model syntax\").\n\n**Named lenses + apertures** beat generic \"shallow depth of field\":\n- `85mm f/1.4`, `135mm f/2.8` (the cheat code for skin texture), `50mm f/1.2`, `35mm f/2`\n- `Panavision anamorphic` for horizontal flares + cinematic width\n- `400mm telephoto` for compression + isolation\n- `24mm` for environmental interiors\n\n**Named cameras / sensors:**\n- `ARRI Alexa 65`, `Hasselblad X2D`, `Canon EOS R5`, `Sony A7III`, `Fujifilm X-T5`\n- \"Specific gear\" beats \"DSLR\"\n\n**Named film stocks** (one per prompt — never mix):\n- `Kodak Portra 400` — natural skin, warm\n- `Fuji Velvia 50` — saturated, landscape\n- `Ilford HP5 Plus` — black and white, gritty grain\n- `CineStill 800T` — tungsten night, halation\n\n**Physics-based lighting** (direction + quality):\n- `Single key light at 45 degrees from upper left`\n- `Late afternoon sun at 15 degrees above horizon`\n- `Color temperature 4500K` beats `slightly warm`\n- `Practicals only — no fill` for Deakins-style realism\n\n**Imperfection vocabulary** (forces away from AI-clean):\n- `visible pores`, `natural skin grain`, `peach fuzz`, `slight hyperpigmentation`\n- `unretouched raw photography`, `ISO noise`, `sweat beading`\n- `crisp catchlights in the eyes`, `skin micro-detail`\n\n**Director references** (use when locking style):\n| Director | Tone | Visual signature |\n|---|---|---|\n| Denis Villeneuve | Cold, vast, existential | Desaturated, overwhelming scale |\n| Roger Deakins | Precise motivated light | Single source, deep shadows, practicals |\n| Emmanuel Lubezki | Natural, spiritual | Available light, golden hour |\n| Bradford Young | Warm darkness | Underexposed, rich shadows, skin tones |\n\n**Genre cues that move the model:**\n- `unstaged documentary photography style`\n- `fashion magazine editorial, shot on medium-format analog film, pronounced grain`\n- `Film still from [Director] [genre]`\n\n## The anti-list — phrases that DEGRADE realism\n\nThese are Stable-Diffusion-era tag soup. The model treats them as low-signal noise. Measured success rate: ~60-70% with these vs ~95%+ with positive description.\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is extracted\n     by src/prompts/banned-tokens.ts, inlined verbatim into the slates_generate_image\n     op description (always in context on both surfaces), and matched against every\n     submitted prompt. Editing this list changes what the agent is told AND what it\n     is warned about — keep every entry backticked, and keep prose outside the\n     backticks. -->\n<!-- /slates-only -->\n**Never use:**\n- `8k`, `4k` (as a quality token)\n- `hyperrealistic`, `ultra-realistic`, `photorealistic` standing alone\n- `masterpiece`, `best quality`, `highly detailed`, `ultra-detailed`\n- `trending on ArtStation`, `award-winning`\n- `perfect skin`, `flawless`, `airbrushed`, `smooth skin`\n- `cinematic` standing alone — always specify *which cinema* (director, lens, era, stock)\n- `not anime, not cartoon, not 3D` — negation tag soup, replace with a positive style cue\n<!-- @banned:end -->\n\n## Negative prompting — there is no field\n\nNano Banana 2 has **no `negativePrompt` parameter**. Three patterns to suppress unwanted content:\n\n1. **Positive reframing (preferred):** \"empty street\" not \"no cars\". \"Unstaged documentary photography\" not \"not anime.\"\n2. **Inline `without` / `free of`:** \"without any people, vehicles, or man-made structures\", \"free of text overlays, logos, or watermarks.\"\n3. **Constraint clauses for anatomy/quality:** \"accurate anatomy with five fingers per hand, symmetrical features, natural proportions\"; \"sharp, well-exposed, free of blur or JPEG artifacts.\"\n\nDefault to #1. Reach for #2 only when positive framing can't suppress the unwanted element.\n\n## Reference images\n\n- **Hard limit: 14 images** (10 object-fidelity + 4 character-consistency). Categories don't trade — you can't use 14 object slots even if no characters are referenced.\n- **Name each reference inline — Slates does this for you.** When you `@mention` a subject/environment or `#mention` a style<!-- slates-only --> (or pass `referenceAssetIds`)<!-- /slates-only -->, Slates composes the prompt so each reference is named inline as \"image N\" — e.g. `Marcus (image 1) sits across from the woman (image 2) in the cafe (image 3)`, with a trailing `Render in the visual style of image 4.` The model does NOT infer a reference's role from its position; the NAME carries it. NB2's own consistency lever is literally **\"assign a distinct name to each character/object\"**. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render the scene's expression\") — that drags the sheet's wardrobe + studio lighting into the scene. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n\n### Reference rules (the verified ones)\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### For Nano Banana 2 specifically\n\n- **NB2's own consistency lever is \"assign a distinct name to each character/object.\"** That is Google's phrasing for rule 3 — cite each canonical identity inline by name.\n- **Rule 8 is a job you do, not one you delegate.** NB2 *is* the start-frame model — when a downstream video shot needs legible text, render it here and animate from this frame.\n- **Character consistency is officially \"not 100% perfect\"** per Google. Test before bulk generations. High-resolution, front-facing reference images help most.\n- **Injection is stochastic — budget 3-5 re-rolls per shot; re-roll, don't re-engineer.** First rolls miss faces/hands; the same prompt lands a clean one within a few tries.\n\n## Common failure modes + fixes\n\n**Hands:** Append `accurate anatomy with five fingers per hand, symmetrical features, natural proportions, relaxed open palm`. Avoid heavy jewelry, props intersecting fingers, motion blur in references.\n\n**Text in images:** Quote-wrap target text. Specify font (`Century Gothic, 12pt`). Long phrases work; small text degrades. Two-step works best — generate text concepts conversationally first, then ask for the image.\n\n**Left/right confusion:** Default is **viewer's perspective**, not subject's. Append `left and right are from the character's perspective, NOT the camera's` when scene-blocking matters.\n\n**Surreal / absurd prompts trip uncanny valley:** The model drags toward realism. If you want surrealism, lean hard into stylization keywords (`painted`, `illustrated`, `stop-motion`).\n\n**Soft faces / dead eyes:** Add `crisp catchlights in the eyes`, `skin micro-detail`, `peach fuzz visible`. Don't stack quality enhancers — single clean prompt beats multiple re-interpretations.\n\n**Post-cutoff content (anything after Jan 2025):** Use reference images. The model has no knowledge of recent franchises, products, events.\n\n## Resolution tactics\n\n- Resolution is priced: NB2 4k costs roughly 2x 1k. Prices change — check current numbers<!-- slates-only --> by calling `slates_estimate_generation_cost`<!-- /slates-only -->. Pick the cheapest resolution that serves the use case.\n- **At 2K and above, the model allocates more tokens to surface detail** — explicit texture vocabulary (pores, fabric weave, grain) compounds at higher resolution.\n- 1k for fast iteration / drafts; 2k for hero shots; 4k only when you need print-grade detail.\n- 2K generations vary 20-60s+. Don't time-budget tightly.\n\n## Boring vs cinema — examples\n\n❌ **Boring:** \"Wide shot of a man on a dock looking at the forest.\"\n\n✅ **Cinema:** \"Direct overhead drone shot on weathered dock surface. Single figure standing center frame, climbing up from frame bottom. Boot prints leading away from him toward shore. Pale winter light. Anamorphic lens flare from low sun. Desaturated blue and slate grey palette. Kodak Portra 400 grain. The path already walked by someone else. Map of threat.\"\n\n❌ **Boring:** \"Close up of a woman looking scared.\"\n\n✅ **Cinema:** \"Extreme close on subject's mouth and nose, 135mm f/2.8, shallow depth of field. Breath pluming out, catching cold light from upper-left key. Lips slightly parted, peach fuzz visible. The breath holds. CineStill 800T halation around catchlights. Waiting.\"\n\n## The 3-strike rule\n\nIf three iterations on the same prompt haven't produced what the user wants, stop. Hand back to the user with what you tried and what isn't working. The slot machine doesn't converge — the prompt structure is wrong, not the seed.\n\n## Family variants — Lite and Pro\n\nEverything in this skill applies to the whole Nano Banana family; two variants trade speed/ceiling around NB2 full:\n\n- **nano-banana-2-lite** — ~half the price, ~2.7× faster, **1K output only**, max 4 refs. The draft/iteration seat: explore compositions here, then re-run the winner on NB2 full at 2K/4K. Same Gemini filter.\n- **nano-banana-pro** — the hero-frame/typography ceiling (~2× NB2, 4K native). NB2 ≈ 95% of Pro; escalate only when spatial composition, cinematic lighting/skin, fine typography-in-scene, or deep multi-element frames must be perfect. Up to 14 refs — it takes a full subject library in one call.\n\n<!-- slates-only -->\nRouting between them (and vs GPT Image 2 / FLUX / Seedream): `slates-model-selection`.\n<!-- /slates-only -->\n",
+    "slates-prompting-nano-banana-2": "---\nname: slates-prompting-nano-banana-2\ndescription: How to write prompts that produce cinematic, photorealistic results from Nano Banana 2 (Google Gemini 3.1 Flash Image, accessed via fal-ai/nano-banana-2). Read this before calling slates_generate_image when the user wants film-quality, real-world, or cinematic output. Skip for stylized / illustrated / cartoon work — the rules differ.\n---\n\n# Nano Banana 2 — cinematic & photorealistic prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Nano Banana 2 (Gemini 3.1 Flash Image).** Brief it like a creative director, not a tag list. Structure: `Film still from [director] [genre]. Shot on [camera] with [lens]. [Subject and action]. [3-5 specific visual details]. [Lighting — direction + quality]. [Color palette]. [Film stock]. [1-2 word tone].`\n\n**The five levers**\n1. **Named lens + aperture** beats \"shallow depth of field\" — `85mm f/1.4`, `135mm f/2.8` (the cheat code for skin), `Panavision anamorphic`, `400mm telephoto`.\n2. **Light by direction and quality**, never \"good lighting\" — `hard sidelight from a single window, deep falloff`, `overcast north light`, `practical tungsten spill`.\n3. **A named film stock or sensor** carries a whole palette — `Kodak Portra 400`, `Cinestill 800T`, `ARRI Alexa 65`.\n4. **Composition as a shot** — `low angle`, `aerial view`, `rule of thirds with the subject camera-left`, `foreground occlusion`.\n5. **Positive framing only.** Describe what is there. \"Empty street\", never \"no cars\"; \"unstaged documentary photography\", never \"not anime\".\n\n**Examples**\n- `Film still from a Denis Villeneuve thriller. Shot on ARRI Alexa 65, 85mm f/1.4. A woman in a charcoal wool coat stands at a rain-slick bus stop, breath visible. Hard sodium light from a single overhead lamp, deep falloff into blue night. Kodak Vision3 500T. Isolated.`\n- `Editorial still life on seamless bone paper. 100mm macro, f/8. A cracked ceramic bowl holding three figs. Soft north light from camera-left, one gentle shadow. Muted earth palette. Portra 400 grain. Quiet.`\n\n**Hard constraint:** there is no `negativePrompt` field. Suppress by reframing positively, or inline `without` / `free of`. Knowledge cutoff January 2025 — anything later needs reference images.\n<!-- @card:end -->\n\nNano Banana 2 is **Gemini 3.1 Flash Image**.<!-- slates-only --> It is the default model behind `slates_generate_image` — the op also exposes `flux-2-max` and `seedream-5-lite`, each with its own prompting skill.<!-- /slates-only --> It is **not** Gemini 3 Pro Image; that is Nano Banana **Pro** (`nano-banana-pro`), a separate model with its own seat.<!-- slates-only --> Verified against the runtime slug map in `slate/src/main/api/google.ts`.<!-- /slates-only --> NB2 is a language model that outputs pixels — brief it like a creative director, not like a Stable-Diffusion tag-soup tool. The single biggest lever for realism: **specificity that mimics how real photographers and cinematographers describe their work**.\n\nKnowledge cutoff: January 2025. Anything after needs explicit reference images.\n\n## Google's 4 official rules (verbatim)\n\n1. **Be specific.** Provide concrete details on subject, lighting, and composition.\n2. **Use positive framing.** Describe what you want, not what you don't want.\n3. **Control the camera.** Use photographic and cinematic terms like \"low angle\" and \"aerial view.\"\n4. **Iterate.** Refine images with follow-up prompts in a conversational manner.\n\n## Official prompt formula\n\n```\n[Subject] + [Action] + [Location/context] + [Composition] + [Style]\n```\n\nFor the cinematic / photoreal use case, expand to:\n\n```\nFilm still from [DIRECTOR] [GENRE]. Shot on [CAMERA] with [LENS]. [SUBJECT and action]. [3-5 specific visual details]. [LIGHTING — direction + quality]. [COLOR PALETTE]. [FILM STOCK or sensor language]. [1-2 word emotional tone].\n```\n\n## Photorealism positives — what consistently works\n\n> ⚠️ **This vocabulary is an IMAGE-model lever and a video-model anti-pattern — do not carry it across.**\n> Named lenses, apertures, film stocks and camera bodies (`85mm f/1.4`, `Kodak Portra 400`, `ARRI Alexa 65`) are correct and encouraged **here**. They are a **Seedance anti-pattern**: ByteDance's own guide uses shot sizes, camera moves, pacing words and its image-quality vocabulary throughout, and never once mentions fps, shutter angle, f-stop, or lens millimetres.\n> The leak happens in one specific way — you write an NB2 start frame, then write the video prompt to animate it and carry the look description straight across. **Translate instead of copying:** `85mm f/1.4, Portra 400` → `close-up, shallow depth of field, warm natural colors, cinematic texture, film-grain texture`. Full rule and the receipts: `slates-prompting-seedance` (Part 3, \"Don't cross-pollinate image-model syntax\").\n\n**Named lenses + apertures** beat generic \"shallow depth of field\":\n- `85mm f/1.4`, `135mm f/2.8` (the cheat code for skin texture), `50mm f/1.2`, `35mm f/2`\n- `Panavision anamorphic` for horizontal flares + cinematic width\n- `400mm telephoto` for compression + isolation\n- `24mm` for environmental interiors\n\n**Named cameras / sensors:**\n- `ARRI Alexa 65`, `Hasselblad X2D`, `Canon EOS R5`, `Sony A7III`, `Fujifilm X-T5`\n- \"Specific gear\" beats \"DSLR\"\n\n**Named film stocks** (one per prompt — never mix):\n- `Kodak Portra 400` — natural skin, warm\n- `Fuji Velvia 50` — saturated, landscape\n- `Ilford HP5 Plus` — black and white, gritty grain\n- `CineStill 800T` — tungsten night, halation\n\n**Physics-based lighting** (direction + quality):\n- `Single key light at 45 degrees from upper left`\n- `Late afternoon sun at 15 degrees above horizon`\n- `Color temperature 4500K` beats `slightly warm`\n- `Practicals only — no fill` for Deakins-style realism\n\n**Imperfection vocabulary** (forces away from AI-clean):\n- `visible pores`, `natural skin grain`, `peach fuzz`, `slight hyperpigmentation`\n- `unretouched raw photography`, `ISO noise`, `sweat beading`\n- `crisp catchlights in the eyes`, `skin micro-detail`\n\n**Director references** (use when locking style):\n| Director | Tone | Visual signature |\n|---|---|---|\n| Denis Villeneuve | Cold, vast, existential | Desaturated, overwhelming scale |\n| Roger Deakins | Precise motivated light | Single source, deep shadows, practicals |\n| Emmanuel Lubezki | Natural, spiritual | Available light, golden hour |\n| Bradford Young | Warm darkness | Underexposed, rich shadows, skin tones |\n\n**Genre cues that move the model:**\n- `unstaged documentary photography style`\n- `fashion magazine editorial, shot on medium-format analog film, pronounced grain`\n- `Film still from [Director] [genre]`\n\n## The anti-list — phrases that DEGRADE realism\n\nThese are Stable-Diffusion-era tag soup. The model treats them as low-signal noise. Measured success rate: ~60-70% with these vs ~95%+ with positive description.\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is extracted\n     by src/prompts/banned-tokens.ts, inlined verbatim into the slates_generate_image\n     op description (always in context on both surfaces), and matched against every\n     submitted prompt. Editing this list changes what the agent is told AND what it\n     is warned about — keep every entry backticked, and keep prose outside the\n     backticks. -->\n<!-- /slates-only -->\n**Never use:**\n- `8k`, `4k` (as a quality token)\n- `hyperrealistic`, `ultra-realistic`, `photorealistic` standing alone\n- `masterpiece`, `best quality`, `highly detailed`, `ultra-detailed`\n- `trending on ArtStation`, `award-winning`\n- `perfect skin`, `flawless`, `airbrushed`, `smooth skin`\n- `cinematic` standing alone — always specify *which cinema* (director, lens, era, stock)\n- `not anime, not cartoon, not 3D` — negation tag soup, replace with a positive style cue\n<!-- @banned:end -->\n\n## Negative prompting — there is no field\n\nNano Banana 2 has **no `negativePrompt` parameter**. Three patterns to suppress unwanted content:\n\n1. **Positive reframing (preferred):** \"empty street\" not \"no cars\". \"Unstaged documentary photography\" not \"not anime.\"\n2. **Inline `without` / `free of`:** \"without any people, vehicles, or man-made structures\", \"free of text overlays, logos, or watermarks.\"\n3. **Constraint clauses for anatomy/quality:** \"accurate anatomy with five fingers per hand, symmetrical features, natural proportions\"; \"sharp, well-exposed, free of blur or JPEG artifacts.\"\n\nDefault to #1. Reach for #2 only when positive framing can't suppress the unwanted element.\n\n## Reference images\n\n- **Hard limit: 14 images** (10 object-fidelity + 4 character-consistency). Categories don't trade — you can't use 14 object slots even if no characters are referenced.\n- **Name each reference inline — Slates does this for you.** When you `@mention` a subject/environment or `#mention` a style<!-- slates-only --> (or pass `referenceAssetIds`)<!-- /slates-only -->, Slates composes the prompt so each reference is named inline as \"image N\" — e.g. `Marcus (image 1) sits across from the woman (image 2) in the cafe (image 3)`, with a trailing `Render in the visual style of image 4.` The model does NOT infer a reference's role from its position; the NAME carries it. NB2's own consistency lever is literally **\"assign a distinct name to each character/object\"**. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render the scene's expression\") — that drags the sheet's wardrobe + studio lighting into the scene. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n\n### Reference rules (the verified ones)\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### For Nano Banana 2 specifically\n\n- **NB2's own consistency lever is \"assign a distinct name to each character/object.\"** That is Google's phrasing for rule 3 — cite each canonical identity inline by name.\n- **Rule 8 is a job you do, not one you delegate.** NB2 *is* the start-frame model — when a downstream video shot needs legible text, render it here and animate from this frame.\n- **Character consistency is officially \"not 100% perfect\"** per Google. Test before bulk generations. High-resolution, front-facing reference images help most.\n- **Injection is stochastic — budget 3-5 re-rolls per shot; re-roll, don't re-engineer.** First rolls miss faces/hands; the same prompt lands a clean one within a few tries.\n\n## Common failure modes + fixes\n\n**Hands:** Append `accurate anatomy with five fingers per hand, symmetrical features, natural proportions, relaxed open palm`. Avoid heavy jewelry, props intersecting fingers, motion blur in references.\n\n**Text in images:** Quote-wrap target text. Specify font (`Century Gothic, 12pt`). Long phrases work; small text degrades. Two-step works best — generate text concepts conversationally first, then ask for the image.\n\n**Left/right confusion:** Default is **viewer's perspective**, not subject's. Append `left and right are from the character's perspective, NOT the camera's` when scene-blocking matters.\n\n**Surreal / absurd prompts trip uncanny valley:** The model drags toward realism. If you want surrealism, lean hard into stylization keywords (`painted`, `illustrated`, `stop-motion`).\n\n**Soft faces / dead eyes:** Add `crisp catchlights in the eyes`, `skin micro-detail`, `peach fuzz visible`. Don't stack quality enhancers — single clean prompt beats multiple re-interpretations.\n\n**Post-cutoff content (anything after Jan 2025):** Use reference images. The model has no knowledge of recent franchises, products, events.\n\n## Resolution tactics\n\n- Resolution is priced: NB2 4k costs roughly 2x 1k. Prices change — check current numbers<!-- slates-only --> by calling `slates_estimate_generation_cost`<!-- /slates-only -->. Pick the cheapest resolution that serves the use case.\n- **At 2K and above, the model allocates more tokens to surface detail** — explicit texture vocabulary (pores, fabric weave, grain) compounds at higher resolution.\n- 1k for fast iteration / drafts; 2k for hero shots; 4k only when you need print-grade detail.\n- 2K generations vary 20-60s+. Don't time-budget tightly.\n\n## Boring vs cinema — examples\n\n❌ **Boring:** \"Wide shot of a man on a dock looking at the forest.\"\n\n✅ **Cinema:** \"Direct overhead drone shot on weathered dock surface. Single figure standing center frame, climbing up from frame bottom. Boot prints leading away from him toward shore. Pale winter light. Anamorphic lens flare from low sun. Desaturated blue and slate grey palette. Kodak Portra 400 grain. The path already walked by someone else. Map of threat.\"\n\n❌ **Boring:** \"Close up of a woman looking scared.\"\n\n✅ **Cinema:** \"Extreme close on subject's mouth and nose, 135mm f/2.8, shallow depth of field. Breath pluming out, catching cold light from upper-left key. Lips slightly parted, peach fuzz visible. The breath holds. CineStill 800T halation around catchlights. Waiting.\"\n\n## The 3-strike rule\n\nIf three iterations on the same prompt haven't produced what the user wants, stop. Hand back to the user with what you tried and what isn't working. The slot machine doesn't converge — the prompt structure is wrong, not the seed.\n\n## Family variants — Lite and Pro\n\nEverything in this skill applies to the whole Nano Banana family; two variants trade speed/ceiling around NB2 full:\n\n- **nano-banana-2-lite** — ~half the price, ~2.7× faster, **1K output only**, max 4 refs. The draft/iteration seat: explore compositions here, then re-run the winner on NB2 full at 2K/4K. Same Gemini filter.\n- **nano-banana-pro** — the hero-frame/typography ceiling (~2× NB2, 4K native). NB2 ≈ 95% of Pro; escalate only when spatial composition, cinematic lighting/skin, fine typography-in-scene, or deep multi-element frames must be perfect. Up to 14 refs — it takes a full subject library in one call.\n\n<!-- slates-only -->\nRouting between them (and vs GPT Image 2.5 / FLUX / Seedream): `slates-model-selection`.\n<!-- /slates-only -->\n",
     "slates-prompting-omni-flash": "---\nname: slates-prompting-omni-flash\ndescription: How to prompt Gemini Omni Flash (Google, via fal). Read before calling slates_generate_video with omni-flash or slates_edit_video with omni-flash-edit. Cheap 720p tier with native synced audio included — 3-10s, 16:9/9:16 only; t2v, single-start-frame i2v, or reference-to-video with up to 7 reference images. The edit variant is the EDIT-FIDELITY WINNER for footage-synced VFX (receipt 2026-07-09) — but ONLY with short prompts: one change + \"Keep everything else the same.\" Long descriptive prompts destroy fidelity.\n---\n\n# Gemini Omni Flash — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Gemini Omni Flash.** Two different jobs with OPPOSITE prompt rules, and getting them the wrong way round is the whole failure mode.\n\n**The five levers**\n1. **Editing: short prompt, ONE change, nothing else.** Google's own doc says so and a 2026-07-09 receipt confirms it — a long \"keep every frame identical\" preamble produced WORSE drift than two sentences.\n2. **Editing: always end with `Keep everything else the same.`** — the one documented preservation lever.\n\n3. **Editing: describe the EFFECT, never a real object as a metaphor.** \"Candle-like flame\" rendered a literal candle in the subject's hand.\n4. **Editing: no conditional timing cues.** \"…when he calls it, as he walks…\" hard-fails with `invalid_request`. Collapse to one continuous action; the model syncs to the footage's own motion.\n5. **Generation: the opposite — describe fully.** Subject, action, setting, `camera tracking alongside`, `overcast flat light`, tone. Audio is prompt-driven with no parameters: dialogue in quotes, sound in plain language — `rain patters on the tin roof`, `spray from tyres`, `a horn somewhere behind`.\n\n**Examples**\n- Edit: `Small magical flames appear on his fingertips when he snaps his fingers, and vanish when he blows on them. Keep everything else the same.`\n- Generate: `A courier in a yellow shell jacket weaves between stalled cars on a wet arterial road, camera tracking alongside at shoulder height. Overcast flat light, spray from tyres. Rain patters on car roofs, a horn somewhere behind.`\n\n**Hard constraint:** it is a CHEAP DRAFT seat for generation and the EDIT-fidelity winner for footage-synced VFX — never a hero generation shot. Expect a possible jitter or doubled speech beat in the last half second of an edit: trim the tail rather than burning a re-roll.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use in an EDIT prompt** (each one has a receipt above):\n- a long preservation preamble — it produces WORSE drift than `Keep everything else the same.`\n- a real object as a metaphor: `candle-like`, `flame-like`, `laser-like`\n- a conditional timing cue: `when he`, `as she`, `once they` — these hard-fail, they do not merely drift\n- harm-to-person framing: `ignite`, `catch fire`, `on fire` applied to a person trips the safety filter\n<!-- @banned:end -->\n\nGoogle's fast video generation + editing model (\"Nano Banana Pro for video\" in creator slang — a nickname; it is NOT the NB Pro image model). Carried on fal (`google/gemini-omni-flash*`). 720p only, 24fps, 3–10 second clips, 16:9 or 9:16. **Audio is native and included** — dialogue, SFX, and ambient generate WITH the video at no extra cost.\n\n## Where it routes\n\n- **Video editing (`omni-flash-edit`) — its headline strength and the edit-lane default** for footage-synced VFX: verified 2026-07-09 head-to-head vs Kling O3 Edit on real phone footage (fire-on-fingertips on a talking take) — Omni Flash held lip movement perfectly, audio near-identical, and executed both action beats; Kling kept audio verbatim but drifted lips and missed the second beat. Full routing: slates-model-selection.\n- **Cheap drafts and iteration volume** — lowest-cost audio-native video seat (~6.4 cr/s at 720p).\n- **NOT hero GENERATION shots** — Kling 3.0 stays the general gen default, Seedance 2.0 the premium tier; Omni Flash's *generation* quality seat is still unproven.\n\n## Editing (`slates_edit_video`, model `omni-flash-edit`) — THE RULES (receipts, not theory)\n\n1. **SHORT PROMPT. One change. Nothing else.** Google's own doc: *\"Simple prompts work best for video editing. Overly descriptive prompts can lead to unintended changes.\"* Live receipt 2026-07-09: a long \"keep every frame/word/movement identical…\" preamble produced WORSE drift (re-synthesized performance, wrong timing); the winning prompt was two sentences: *\"Small magical flames appear on his fingertips when he snaps his fingers, and vanish when he blows on them. Keep everything else the same.\"*\n2. **Always end with \"Keep everything else the same.\"** — the one documented preservation lever.\n3. **Never name a real-world object as a metaphor.** \"Candle-like flame\" rendered a literal candle in his hand. Describe the effect itself (\"small magical flames on his fingertips\").\n3b. **No conditional timing cues — they HARD-FAIL, not drift.** Receipt 2026-07-09: \"a dragon appears behind him, flies onto his shoulder WHEN HE CALLS IT, and perches AS HE WALKS…\" → deterministic `invalid_request` (2×, \"could not generate with the given inputs\"); collapsing to one continuous action — \"A small photorealistic dragon flies in and perches on his shoulder, puffing a small breath of flame and smoke.\" — succeeded first try. The model syncs the change to the footage's own motion; it cannot take beat-by-beat stage directions cued to moments in the video.\n4. **Safety filter (Google's, strict about harm-to-person):** \"fingertips ignite / catch fire\" → `content_policy_violation`. Frame effects as magical/harmless VFX: \"small magical flames appear on his fingertips\" passed. See slates-content-policy §Gemini for the substitution patterns.\n5. **Expect a possible tail artifact** — jitter or a doubled final speech beat in the last ~0.5s. Plan to trim the tail on the timeline; don't burn a re-roll on it.\n6. **Prompt + source clip ONLY.** No element/style reference images — identity swaps that need refs go to `kling-v3.0-omni-edit`.\n7. Source clip 3–10s (trim longer clips first). Output length follows the source; billing per output second, rounded up. Voice editing unsupported — never ask it to change dialogue.\n8. **Ship via segment-splice** (the workflow, not the model): edit only the seconds where the change happens, splice back over the original on the timeline with the original audio underneath. Most of the deliverable stays untouched original footage — this is how the pro demos are actually assembled (gesture-only edited beats + voiceover in post).\n9. Chain edits one change at a time — each edit saves as a new asset linked to its parent.\n\n## Generation (`slates_generate_video`, model `omni-flash`)\n\n- **Inputs:** prompt only (t2v), prompt + ONE start frame (`firstFrameAssetId`, i2v), or prompt + up to **7 reference images** (ingredient/character/environment/style asset params — they merge into one reference list). No last frame, no video/audio references — the op rejects them.\n- Descriptive prompts are fine for GENERATION (the short-prompt law above is edit-specific). Structure like a shot brief: subject + action + setting + camera + lighting + tone.\n- **Name references inline** the standard Slates way (\"Marcus (image 1) walks…\"). The endpoint also accepts explicit `<IMAGE_REF_0>`-style binding tags (zero-indexed) — useful when a specific image must bind to a specific role.\n- **Audio is prompt-driven** — no audio parameters. Dialogue in quotes; direct sound in plain language (\"rain patters on the tin roof\"). Negative direction as plain instructions (\"Do not show text\").\n- Duration is an explicit 3–10s integer param; cost scales linearly per second.\n\n## Input conditioning (Slates handles this — know it exists)\n\nPhone footage stores rotation as a metadata flag; models ignore it and edit the raw sideways pixels. Clips must be rotation-normalized (and oversized sources downscaled) before upload — receipt 2026-07-09: a portrait Pixel clip came back sideways until conditioned. If an edit output comes back rotated, the source wasn't normalized.\n\n## Content notes\n\n- Google applies its own safety filters to input images/clips and output. Uploads containing recognizable real people are restricted by Google's policy — though own-footage editing of the uploader passed on our route 2026-07-09. See slates-content-policy.\n- Output carries an invisible SynthID watermark (Google-side, programmatic detection only).\n",
     "slates-prompting-seed-audio": "---\nname: slates-prompting-seed-audio\ndescription: How to prompt Seed Audio 1.0 (ByteDance, via fal). Read before calling slates_generate_audio with model seed-audio. The one-pass audio SCENE model — dialogue, SFX and ambience together from ONE plain sentence. CRITICAL - it has NO duration parameter, so length must be named IN THE PROMPT TEXT and Slates bills the duration you request. Covers the one-sentence doctrine, the crowd-size rule, why Kling \"SFX:\" syntax hurts here, and the audio-refs-XOR-image input rule.\n---\n\n# Seed Audio 1.0 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Seed Audio 1.0.** One plain sentence describing a SCENE, and it returns dialogue, effects and ambience together in one pass. For ONE named voice saying ONE line, `inworld-tts-2` is the seat instead; this one renders the whole room.\n\n**The five levers**\n1. **Write one sentence in plain language.** Describe the room and what is happening in it; the model casts and performs it.\n2. **Name the crowd size, the room size and the distance.** This is the highest-leverage single edit on any bed — unqualified nouns default BIG. \"Tiny applause of two or three people at an open mic\", not \"applause\".\n3. **Put the length in the prompt** and make it deliberate.\n4. **Dialogue is performed inside the scene** — write the line as spoken in the room, then re-roll until a take is right and lip-sync against it.\n5. **Describe sounds directly** — \"one coffee machine hissing, cutlery somewhere behind the counter\" — with the object, the action and where it is.\n\n**Examples**\n- `a diner at 2am, one coffee machine hissing, cutlery somewhere behind the counter, one man says quietly \"you're late again\". 15 seconds`\n- `nature soundscape, a wide open field, cicadas near, birds mid-distance, one loon far off across water. 20 seconds`\n\n**Hard constraint:** it has NO duration parameter — the length you request is written into the prompt AND is what you are billed, whatever comes back, so choose it deliberately. Kling's `SFX:` / `Ambient noise:` / `Background music:` labels have no parser here and measurably worsen the output. Shot language, camera moves and lighting are video-prompt words the model must ignore. It is not a music model and it cannot produce pixels.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** — Kling video syntax has no parser here and measurably worsens the output:\n- `SFX`, `Ambient noise`, `Background music` as labels — describe the sounds directly\n- shot language: `wide shot`, `slow push in`, `warm tungsten` — camera and lighting words are video-prompt words the model has to ignore\n<!-- @banned:end -->\n\nByteDance's one-pass audio scene model, carried on fal (`bytedance/seed-audio-1.0`). It generates dialogue, sound effects and ambience **together**, from a single plain sentence. 1–120 seconds. It is the default audio model in Slates and the workhorse for continuity beds.\n\n## Where it routes\n\n- **Scene audio, room tone, ambience beds, crowd/nature soundscapes** — anything where several sounds share a space. One generation, not three layered ones.\n- **Dialogue and scratch VO inside a scene.** The line is performed in the room, by a voice the scene casts. When WHO is speaking matters — a character's own voice, a clean narrator — that is `inworld-tts-2` instead. Lock the read by re-rolling until a take is right, then lip-sync against it with `slates_generate_lip_sync`.\n- **NOT** a single effect that must land on a known frame — that is `eleven-sfx`, which takes an exact duration.\n- **NOT** music. Slates has no music model; import a track and drop it on an audio track.\n- **AUDIO-ONLY.** It cannot produce images or video.\n\n## THE RULES\n\n### 1. 🚨 There is no duration parameter — the words set the length\n\nThis is the single most important fact about this model. Output length is driven by the prompt text (\"… 15 seconds\"), capped at 120s.\n\n**Slates handles this for you:** the `durationSeconds` param appends the duration to the prompt and **bills that number of seconds**. So:\n\n- Set `durationSeconds` to what you actually want.\n- **Do not also write a different length into your sentence.** Two numbers fight, and you pay for the one you selected, not the one you got.\n- If the returned clip is shorter than requested you still paid for the request — that is the deal that keeps the displayed price equal to the charge. Ask for what you need.\n\n<!-- slates-only -->\nThe server re-derives the billed key from `durationSeconds` (a client cannot under-bill), probes the returned `audio.duration` after completion, and logs `SEED AUDIO BILLING DRIFT` if the model overshot. No auto-charge, no refund — the request is the contract.\n<!-- /slates-only -->\n\n### 2. One plain sentence. No production jargon.\n\nField-proven (Higgsfield sprint, 2026-07-27/28). Working prompts look like this:\n\n```\ntiny applause of 2 or 3 people at an open mic. 15 seconds\nnature soundscape, wide open field cicadas and birds and a loon.\na diner at 2am, one coffee machine hissing, cutlery somewhere behind the counter\n```\n\nNot this:\n\n```\n✗ AMBIENCE: interior diner, night. SFX: espresso machine (hiss, 2s), cutlery.\n✗ Wide shot of a diner. Slow push in. Warm tungsten. Ambient noise: ...\n```\n\nShot language, camera moves and lighting belong to video prompts. Here they are just words the model has to ignore.\n\n### 3. Never bring Kling's audio syntax to this model\n\n`SFX:` and `Ambient noise:` prefixes and `Background music:` labels are **Kling 3.0 video** syntax. Seed Audio has no parser for them — it reads them as text in the scene and the output gets measurably worse. Describe the sounds directly instead.\n\n### 4. Name the crowd size, the room size, the distance\n\nThe highest-leverage single edit on any bed. Unqualified nouns default big:\n\n| Vague | What it returns | Fixed |\n|---|---|---|\n| `applause` | a full auditorium | `tiny applause of 2 or 3 people` |\n| `traffic` | a highway | `one car passing on a wet residential street` |\n| `crowd` | a stadium | `four people talking at the next table` |\n\nDistance words (`far off`, `muffled through a wall`, `right next to the mic`) work the same way and are how you build depth in one sentence.\n\n### 5. Beds must outlast the cut\n\nAsk for a few seconds more than the clip needs so the edit has handles to fade through. A bed that ends exactly on the cut always sounds clipped. This is a product requirement, not a preference — it is why the duration control exists at all.\n\n### 6. Dialogue goes in quotes, inside the same sentence as the room\n\n```\na tired bartender says, \"we closed twenty minutes ago\", glasses clinking behind him\n```\n\nPick a preset voice when a specific speaker matters. Leave `voice` unset and the scene casts itself — which is usually right for crowd and background dialogue.\n\nPreset voices (20): `vivi_mixed_en_zh_ja_es_id`, `mindy_en_es_id_pt_zh`, `kian_en_zh`, `cedric_en_zh`, `sophie_en_zh`, `jean_en_zh`, `magnus_en_zh`, `mabel_en_zh`, `nadia_en_zh`, `opal_en_zh`, `pearl_en_zh`, `quentin_en_zh`, `corinne_mixed_en_zh`, `esther_mixed_en_zh`, `lyla_mixed_en_zh`, `tracy_es_zh`, `sandy_es_mixed_en_zh`, `felix_zh`, `celeste_zh`, `monkey_king_zh`.\n\nSet `multilingual: true` for non-English or mixed-language lines.\n\n### 7. Inputs: up to 3 audio clips **XOR** one image. Never both.\n\n- **Audio references** — up to 3 clips, each ≤30s and ≤10MB (wav/mp3/pcm/ogg_opus). Refer to them in the prompt as `@Audio1`, `@Audio2`, `@Audio3`: *\"match the room tone of @Audio1\"*.\n- **Image reference** — one image (jpeg/png/webp ≤10MB). The model scores what it sees.\n- Sending both is rejected by the API. Pick the one that carries the intent.\n\n### 8. The knobs, and when to touch them\n\n| Param | Range | Reach for it when |\n|---|---|---|\n| `speed` | 0.5–2.0 | Dialogue is racing or dragging against picture. |\n| `volume` | 0.5–2.0 | Rarely — normalize on the timeline instead. |\n| `pitch` | −12…+12 semitones | Ageing or shifting a voice. Small moves only; ±3 is already a lot. |\n| `multilingual` | bool | Non-English or code-switched lines. |\n| `sampleRate` | 8k–48k | Leave at 24000 unless you are matching an existing stem. |\n| `outputFormat` | mp3 / wav / pcm / ogg_opus | wav when this is going into a mix; mp3 otherwise. |\n\n## Iterating\n\n- A bed that came back wrong is almost always a **scale** problem (crowd/room too big) or a **jargon** problem (the sentence reads like a spec). Fix those two before touching `speed`/`pitch`.\n- Three failed takes on the same sentence means the sentence is wrong, not the seed. Rewrite it the way you would say it out loud.\n- Generations are cheap enough at short durations that auditioning two phrasings beats agonizing over one.\n\n## Content notes\n\nProvider-side moderation applies to voices and to recognizable real people. See slates-content-policy.\n",
-    "slates-prompting-seedance-2-5": "---\nname: slates-prompting-seedance-2-5\ndescription: How to prompt Seedance 2.5 and Seedance 2.5 Edit. Read before calling slates_generate_video with model seedance-2.5, or slates_edit_video with model seedance-2.5-edit. 2.5 is a SECOND SEAT next to 2.0, not an upgrade — it buys 30-second takes, 30 image references, audio-only references and INTEGER-SECOND TIMESTAMPS, and it gives up native 4K and costs more than 2.0 at every resolution they share. Timestamps are the one grammar difference that matters: 2.0 ignores them and answers only to shot numbers, 2.5 acts on them. Otherwise it shares 2.0's grammar (read slates-prompting-seedance for subject binding, camera and constraint vocabulary); this file covers what is different, plus the two hazards unique to 2.5 — the prompt-intent task classifier and the cost trap that comes with 30-second takes.\n---\n\n# Seedance 2.5 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Seedance 2.5.** Shares 2.0's grammar exactly (subject binding, camera vocabulary, externalised emotion, inline constraints — read `slates-prompting-seedance` for those). Two things are different, and both matter.\n\n**The five levers**\n1. **Timestamps work here** — integer seconds, and the model acts on them: `[0-4] she reads the letter. [4-9] she folds it and looks up.` 2.0 ignores exactly this syntax.\n2. **Length is the reason to be here** — takes up to 30 seconds, where 2.0 stops at 15. Write the beats as `[0-6]`, `[6-12]`, `[12-18]`; do not hope for them.\n3. **Up to 30 image references**, and a multi-view image can serve as ONE subject reference (up to 5 subjects). 2.0 cannot do either.\n4. **Audio-only references are accepted** without an image or video alongside — the only Seedance seat that takes one.\n5. **Keep the 2.0 discipline**: one camera move per beat (`slow track right`, `handheld follow`), physical action instead of stated emotion, and quality asked for in the image-quality slot vocabulary — `rich details`, `natural colors`, `cinematic texture`, `soft lighting`.\n\n**Examples**\n- `[0-6] Wide shot, <Subject_1>@<Image_1> crosses an empty car park toward a idling van, slow track right. [6-12] Medium, she stops as the driver's window comes down. [12-18] Close-up, she looks off past the lens and does not answer. Rich details, natural colors. Keep it subtitle-free.`\n- `[0-10] A single continuous handheld follow behind a courier climbing a fire escape, rain. [10-20] She reaches the landing, turns, and the city opens behind her. Cinematic texture, soft lighting.`\n\n**Hard constraint:** it is the EXPENSIVE seat and it has NO 4K — 480p/720p/1080p only, and dearer than 2.0 at every resolution they share. It is a second seat, never an upgrade. Long takes multiply cost linearly: quote a 30-second take before you fire it.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** (2.5 reclassifies the task and fails a fresh generation on these):\n- `edit`, `extend`, `continue the video`, `same video but` — they make the provider read a fresh generation as an edit\n- `85mm`, `f/1.4`, `Portra 400` and any other lens, aperture, film-stock or camera-body token — image-model vocabulary, a Seedance anti-pattern on both seats\n<!-- @banned:end -->\n\n**Read `slates-prompting-seedance` first.** The prompt GRAMMAR is the same model family: the\n8-slot advanced formula, subject binding by `<Subject_N>@<Image_N>`, camera vocabulary,\nexternalised emotion, inline constraint words, the anti-twin fix. None of it is restated here.\nThis file is only what 2.5 changes — and the biggest change is that **2.5 acts on timestamps\nwhere 2.0 ignores them.**\n\n---\n\n## The one fact that decides whether you use it at all\n\n**Seedance 2.5 is the EXPENSIVE seat, not the cheap one — and it has no 4K.**\n\nIt runs at 480p, 720p or 1080p (1080p landed on all three routes on 2026-08-24), and at every\nresolution the two seats share it costs MORE than 2.0 — 720p $0.231/s against $0.15/s, **54% more**.\nSo 2.5 does not replace 2.0; it sits beside it, and you pay for what it buys:\n\n| | Seedance 2.0 | Seedance 2.5 |\n|---|---|---|\n| Resolution | 480p / 720p / 1080p / **native 4K** | 480p / 720p / 1080p — **no 4K** |\n| Price at 720p (faceless) | **$0.15/s** | $0.231/s |\n| Length | 4–15s | **4–30s in one take** |\n| Reference budget | 15 (9 image + 3 video + 3 audio) | **50 (30 image + 10 video + 10 audio)** |\n| Combined reference video/audio | ≤15s | **≤30s** |\n| Audio-only reference | ✗ (needs an image or video alongside) | **✓** |\n| **Timestamps in the prompt** | **✗ — ignored; shot numbers only** | **✓ — integer seconds, acted on** |\n| Multi-view image as ONE subject reference | ✗ (not recommended) | **✓ (up to 5 subjects)** |\n| Video edit as its own task type | ✗ | **✓ (`seedance-2.5-edit`)** |\n| Default video model | **yes** | no |\n\n**Route to 2.5 when the shot needs LENGTH, MANY REFERENCES, or an audio-only reference.\nRoute to 2.0 when resolution matters at all** — which, for anything a client will see full-screen,\nis most of the time.\n\n---\n\n## 🚨 Hazard 1 — the prompt-intent task classifier\n\nThis is the one that costs money and time, and it has no equivalent on 2.0.\n\n**Seedance 2.5 sorts every request into one of five task types** — text-to-video,\nreference-to-video, first/last-frame, **video edit**, **video extend** — from the reference roles\nattached **plus the intent of your sentence**. Each type then has its own parameter constraints,\nand a violation comes back **asynchronously**: the task queues, credits are reserved, and only then\ndoes it fail.\n\nThe trigger words are ordinary English:\n\n| Reclassified as | Words that do it (ByteDance's own list) |\n|---|---|\n| **video edit** | `edit video` · `add` · `insert` · `remove` · `delete` · `modify` · `replace` · `change to` |\n| **video extend** | `extend forward` · `extend backward` · `continue` · `continue from` · `extend the story` |\n\nSo a perfectly legitimate reference-to-video prompt — *\"a wide shot of the workshop, **remove** the\ntripod from frame\"* — gets classified as an edit and fails on constraints it never set.\n\n**What to do:**\n\n1. **If you mean to edit an existing clip, say so with the MODEL, not the sentence.** Call\n   `slates_edit_video` with `model: 'seedance-2.5-edit'`. That routes to a dedicated\n   task-typed endpoint and the classifier never has to guess.\n2. **If you mean a fresh shot, describe the finished frame rather than an instruction to change\n   one.** Not *\"remove the tripod\"* → *\"the workshop bench, clear and uncluttered\"*. Not\n   *\"add rain\"* → *\"heavy rain falling through the streetlight\"*. This is better prompting anyway:\n   the model renders what you describe, it does not take edits to an imagined draft.\n3. The trigger only fires when **references are attached**. A plain text-to-video prompt is safe\n   however it is worded.\n\n**Slates will warn you, and it will never rewrite your prompt.** When a 2.5 reference generation's\nprompt contains one of these words, the composer shows a warning that NAMES the words and the agent\nroute returns the same string. Silently editing the user's sentence to dodge a provider classifier\nis forbidden — the words that reach the model are always the words the user can see.\n\n---\n\n## 🚨 Hazard 2 — resolution is not the price dial here. LENGTH is.\n\nEvery other model in Slates trains the habit that lower resolution means lower cost. 2.5 breaks it,\nbecause the thing that moves the bill is **length**, and 2.5's length ceiling is double 2.0's.\n\nWorked, at the shipped rates:\n\n| Generation | Credits |\n|---|---|\n| 2.5 · 480p · 5s · faceless | 26 |\n| 2.5 · 720p · 5s · faceless | 58 |\n| 2.5 · 1080p · 5s · faceless | 103 |\n| 2.5 · 720p · 30s · faceless | 347 |\n| 2.5 · 720p · 30s · AI-face route | **489** |\n| 2.5 · 720p · 30s · consented real-face route | **710** |\n| 2.5 · 1080p · 30s · faceless | **614** |\n| 2.5 · 1080p · 30s · consented real-face route | **1,749** |\n| *(for scale)* 2.0 · 1080p · 15s · AI-face route | 411 |\n\n**A 30-second 720p clip can cost more than a 15-second 1080p one** — and a base licence starts\nwith 1,000 credits. Someone who reads \"720p\" as \"cheap\" and asks for a 30-second take on the\nreal-face route has spent 71% of their welcome grant on one clip; **on the real-face route a single\n30-second 1080p take is more than the whole grant.**\n\n**Discipline:**\n\n- **Always quote with `slates_estimate_generation_cost` before a take over ~10 seconds,** and say\n  the number out loud before generating.\n- **Find the shot at short LENGTH, not at low resolution.** Length is what moves the price, so cut\n  seconds while you are still exploring — 4–8s — and stay at the resolution you actually want.\n  **A 480p pass does not de-risk a 720p or 1080p render.** Generation is stochastic: the higher-\n  resolution run is a different take, not the same shot rendered better. So a 480p draft that looks\n  right buys you no guarantee, and one that looks wrong may have been fine at 720p — you paid 26\n  credits to learn nothing, when 58 would have bought a real candidate.\n- **Length is a creative decision, not a default.** 30 seconds is available; it is rarely the right\n  answer for a single shot. Multi-shot storyboards inside one 30s generation are what the length is\n  actually for.\n- Read `slates-cost-discipline` — all of it applies, more sharply here.\n\n---\n\n## Timestamps — the one grammar change\n\n<!-- @inject:seedance-25-timestamps -->\n**2.0 does not respond to timestamps and answers only to shot numbers. 2.5 responds to\ninteger-second timestamps.** That is ByteDance's own first line under \"Differences from Seedance\n2.0\", and it is why a 30-second take is usable at all: the length is only worth buying if you can\nsay *when* things happen inside it.\n\nBoth formats are valid on 2.5, and you can mix them — `Shot N` blocks for a storyboard whose\npacing you are happy to leave to the model, timestamps when a beat has to land at a moment.\n\n**Three ways to control time, all first-party:**\n\n| Form | Write it like |\n|---|---|\n| **Interval** | `0-3 seconds… 3-7 seconds… 7-15 seconds` or `[1s-4s]… [4s-8s]… [8s-12s]` |\n| **Time point** | *\"Quick left sideways transition at the 5-second mark.\"* |\n| **Relative** | *\"After 3 seconds, everyone around him shakes their head.\"* · *\"The frame freezes for 1 second after he presses the shutter.\"* |\n\n**The rules that come with them:**\n\n- **One second is the smallest unit.** Integers only — no `2.5s`, no frames.\n- **No gaps in the timeline.** `0-3s… 5-6s…` leaves 3-5s unspecified and the model fills it however\n  it likes. Intervals must abut: `0-3s`, `3-7s`, `7-15s`.\n- **Budget the plot to the seconds.** Too little content in a range and the model improvises to\n  fill it; too much and you get extra cuts or dropped beats. This is the actual craft of a 30s take.\n- **Never time-code a high-frequency action.** *\"Shake your head three times per second\"* is\n  explicitly called out as a misuse — timestamps schedule beats, they don't choreograph frames.\n- **Transitions want both halves:** the moment AND the method — *\"At the 5-second mark, the camera\n  transitions leftward with a left wipe into a natural dissolve.\"*\n- **Timestamps work on an EDIT too**, and that is where they earn the most: they scope a change in\n  time as well as in content — *\"Change the man's action from drinking coffee to mopping the floor\n  from 4-6 seconds in Video 1, and leave the rest of the content unchanged.\"* Without a range, a\n  whole-clip instruction is applied to the whole clip.\n\nDo **not** carry this back to 2.0, and do not carry Veo's `[00:00-00:02]` bracket syntax into\neither — 2.0 ignores time entirely, and the cross-model syntax swap is its own known failure.\n<!-- @end:seedance-25-timestamps -->\n\n---\n\n## What the extra reference budget is actually for\n\n30 image references (up from 9) does **not** mean \"attach 30 images\". Every rule in\n`slates-prompting-seedance` about references still holds — 2–4 strong references beat both\nextremes, and one reference per role.\n\n**Where 2.5 moves the ceiling, per ByteDance's own input recommendations:**\n\n| | Stable | Works, but expect re-rolls |\n|---|---|---|\n| Subjects bound by IMAGE reference | 1–8 | 9–12 |\n| Subjects bound by VIDEO or AUDIO reference | 1–5 | 6–10 |\n| Reference clip length, per subject | 5–10s | longer drops stability |\n\n**Multi-view images of one subject are supported on 2.5** — a turnaround sheet can be a single\nreference image, where 2.0 wanted one authoritative rendering per subject. Past **5 subjects**,\ngo back to single-view images, one per view, rather than one image carrying several viewpoints.\n\nThe larger budget earns its keep in exactly two places:\n\n- **A long multi-shot take** where different shots need different subjects and locations bound —\n  the budget is spread across the storyboard, not stacked on one frame.\n- **Video and audio references alongside images**, which is where 2.5's 10 + 10 matters far more\n  than the image count.\n\n### Audio-only references — the genuinely new input\n\n2.0 required an image or video alongside any audio reference. **2.5 accepts audio on its own.**\nThat makes one recipe possible that was not before: drive a scene's timing, voice or ambience from\na recording with no visual anchor at all — a voice line, a music bed, a room tone — and let the\nmodel build the picture to it. Cite it the same way as any other reference\n(`Reference the timbre in <Audio_N> to generate…`), and remember that audio references carry **no\nbilling dimension** on any Seedance route: audio is included.\n\n### Video references\n\nUp to 10 clips, ≤30s combined (2.0: 3 clips, ≤15s). A reference VIDEO switches the cost key to\n`seedance-2.5*-vref-{res}-{T}s`, where **T = Σ input seconds + output seconds** — the sum is across\n**every** clip attached, not just the longest. Three 6-second references on a 12-second output bills\n30 seconds, not 12 and not 18. Quote before confirming.\n\n**The cap is a refusal, not a trim.** Attach an eleventh clip, or push past 30 combined seconds, and\nthe composition is rejected before anything uploads. That asymmetry is deliberate: reference images\nwarn-and-trim because dropping one doesn't change the price, and a dropped reference VIDEO would be\none you were quoted for and the model never saw.\n\n### Mixing all three in one call\n\n50 files total (30 image + 10 video + 10 audio) is a shared budget. Everything is cited positionally\nby type — `image 1`, `video 2`, `audio 1` — in attachment order, so reordering the attachments\nrenumbers the citations. Write the prompt against those numbers:\n\n```\nMarcus (image 1) performs the motion from video 1 in the workshop from image 2,\nspeaking the line in audio 1. Preserve his identity, appearance and outfit.\n```\n\nFrames and reference media stay mutually exclusive, in every combination — the reference endpoint\nhas no first/last-frame parameters at all, so this is a shape mismatch rather than a preference.\n\n---\n\n## Seedance 2.5 Edit (`slates_edit_video`, `model: 'seedance-2.5-edit'`)\n\nIts own picker row and its own op call, deliberately: the task type is **the model you chose**,\nnever something inferred from your sentence.\n\n**Why route here at all:** it is the **only edit engine in Slates that accepts a clip longer than 15\nseconds** (4–30s, versus Kling O3 Edit's 3–15s and Omni Flash Edit's 3–10s). For a clip inside the\nothers' range, choose on fidelity instead — Omni Flash Edit won the prompt-only head-to-head, and\nKling O3 Edit is the one that takes element and style reference images.\n\n**How it behaves:**\n\n- **Output length follows the SOURCE clip**, and the bill is the ceiled source length. The provider\n  requires an automatic duration on this task type, so there is no length knob — the clip you attach\n  is the quote. The returned clip can differ from the source by up to ~0.3s, which only compresses\n  transition frames; a clip that 2.5 itself generated comes back at exactly its input length.\n- **Source clips under 20 seconds edit more reliably.** 4–30s is what the task type accepts;\n  ByteDance's own recommendation is to stay inside 20 for quality. A 28-second source is legal and\n  will need more attempts.\n- **The aspect ratio follows the source clip too.** No ratio control; the frame is the clip's frame.\n- **480p, 720p or 1080p output**, native audio — and an edit bills the video-reference tier ×2,\n  so 1080p on this row is the most expensive second in the app. Quote it.\n- **Prompt and source clip only** on this op. The MODEL takes reference images on an edit\n  (ByteDance recommends 1–5 — *\"replace the man in dark clothing in @Video 1 with @Image 2\"*);\n  **Slates has not wired that path**, so today an edit that must lock an identity from a photo\n  goes to Kling O3 Edit. Constraint of our build, not of the model — worth revisiting.\n- **An edit bills roughly DOUBLE a plain 2.5 generation of the same length**, because every provider\n  charges an edit on input + output seconds. Read the confirm gate's number; do not reason from the\n  generation rate.\n- **Set `seedanceFace: true` when a character's face is visible in the clip.** The faceless provider\n  blocks faces outright — this is not a price optimisation, it is whether the job runs at all.\n- **There is no consented-real-face route for editing.** Real-person footage that the AI-face route\n  rejects has to go to Kling O3 Edit.\n\n**Prompting an edit** — the same discipline as every other edit engine: **describe only what\nchanges.** The source already carries its composition, motion, timing and performance; re-describing\nthem fights the model. Use Seedance's own edit grammar from `slates-prompting-seedance`\n(*\"Strictly edit `<Video_1>`, and modify `<Original_Characteristic>` to `<New_Characteristic>`\"*) and\n**never** write *\"reference video 1\"* in an edit — the official guide is explicit that this phrasing\ngets the request reclassified as a reference task, which is the same landmine as Hazard 1.\n\nTwo things sharpen an edit prompt, both first-party:\n\n- **Say it as A → B, not as an outcome.** *\"Change the man's action from drinking coffee to mopping\n  the floor\"* beats *\"the man mops the floor\"* — naming what it currently is tells the model what\n  to overwrite.\n- **Timestamp a partial edit** — the edit task type reads the same integer-second timestamps the\n  generation path does. Rules and forms are in § Timestamps above; this is the single most useful\n  thing they buy.\n\n**Audio is editable too, and it is the least obvious use of this row.** The same op rewrites what\nis heard while the picture stays put: change a spoken line, change the accent, translate the\ndialogue and re-fit the lip movement, strip or replace the BGM or a sound effect. *\"Only edit the\nman's dialogue in Video 1: change it to 'Don't come over here,' in an American accent\"* is an edit,\nnot a lip-sync job. Bill it like any other edit — on the source clip's length.\n\n---\n\n## Faces, and what does NOT change\n\nThe three-tier face routing is identical to 2.0 — faceless → default route, an AI character's face →\n`seedanceFace: true` (the relaxed provider, a real cost premium), a real person's photo → the\nconsent-gated premium route after a `[REAL_FACE_DETECTED]` rejection, with `realFaceConsent: true`\nset **only** after the user explicitly confirms they hold the rights to the likeness. The full rules,\nincluding why the real-vs-AI call is the provider's and not yours, are in\n`slates-prompting-seedance`.\n\nAlso unchanged, and worth restating because 2.5's length makes each one more expensive to get wrong:\n\n- **One primary camera move per shot.**\n- **No lens / aperture / film-stock vocabulary.** That is image-model syntax and a Seedance\n  anti-pattern.\n- **No `negativePrompt` field** — constraints go inline, and 2.5 acts on negative phrasing in\n  exactly two dimensions: subtitles (*\"no subtitles\"*) and audio (*\"no BGM; environmental and\n  action sounds only\"*, *\"no audio\"*). Everywhere else, describe what you want, not what you don't.\n- **Legible in-shot text still belongs in a baked start frame**, not in the video prompt.\n",
-    "slates-prompting-seedance": "---\nname: slates-prompting-seedance\ndescription: How to prompt Seedance 2.0 (ByteDance video model). Read before calling slates_generate_video with model seedance-2. Seedance 2.0 structures multi-beat prompts as a \"Shot 1 / Shot 2 / Shot 3\" storyboard against an 8-slot advanced formula — never per-second time stamps, which 2.0 does not respond to (Seedance 2.5 does; see slates-prompting-seedance-2-5). Its syntax differs from Kling, Veo and the image models; don't cross-pollinate (in particular, no lens / aperture / film-stock vocabulary).\n---\n\n# Seedance 2.0 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Seedance 2.0.** Not copywriting — an ENGINEERING instruction to a spatial layer and a temporal layer: who, in what scene, doing what, how the camera moves, and in what order. Multi-beat work is a `Shot 1 / Shot 2 / Shot 3` storyboard.\n\n**The five levers**\n1. **Bind every subject to its reference** — `<Subject_1>@<Image_1>` — and keep the descriptions identical across shots. Unbound subjects are where twins come from.\n2. **Shot sizes and camera MOVES, not lens data** — `medium close-up`, `slow push in`, `handheld follow`, `whip pan`. One primary move per shot.\n3. **Externalise emotion as physical action.** Not \"she is nervous\": `she turns the ring on her finger twice, then stops`.\n4. **Use the image-quality slot vocabulary** for quality — `HD`, `rich details`, `cinematic texture`, `natural colors`, `soft lighting`. That is the officially sanctioned way to ask.\n5. **Constraints go INLINE**, led by the official templates — `keep it subtitle-free`, `do not generate a watermark`, `avoid jitter and bent limbs`, `avoid temporal flicker`.\n\n**Examples**\n- `Shot 1: medium shot, <Subject_1>@<Image_1> steps out of the freight lift into a wet loading bay, slow push in. Shot 2: close-up, she turns the ring on her finger twice and stops, handheld. Rich details, cinematic texture, natural colors. Keep it subtitle-free.`\n- `Single continuous take. Wide shot of a fishing skiff crossing a grey swell, camera tracks from the starboard rail. Spray hits the lens once. Soft lighting, natural colors, film-grain texture. Avoid jitter and bent limbs.`\n\n**Hard constraint:** NO timestamps — 2.0 ignores them and answers only to shot numbers (2.5 acts on them). No lens, aperture, film stock or camera body: that is image-model vocabulary and a Seedance anti-pattern. There is no negativePrompt field.\n<!-- @card:end -->\n\nByteDance's video model — first-party via **BytePlus ModelArk** (credits only, no BYOK). Audio always generated alongside the video. Single model `seedance-2` across the full resolution ladder (480p / 720p / 1080p / native 4K — 4K video is Pro-only, default 1080p), 4–15s, first+last frame, and up to 9 reference images / 3 videos / 3 audio clips.\n\n> **How to read this file.**\n> **[official :NNNN]** — ByteDance's own BytePlus ModelArk prompting guide, line `NNNN` of the archived doc dump (`research/seedance-2-modelark-docs.md`). Receipt-grade; treat as law.\n> **[community]** — third-party guides and our own field notes. Useful, but an `[official]` block always wins.\n> **[slates]** — how the Slates app composes or bills this; not ByteDance doctrine.\n>\n> The split is load-bearing. A community-sourced \"narrative timing beats\" doctrine shipped in this file for months teaching the **exact inverse** of ByteDance's published guidance. Never merge the two registers again.\n\n---\n\n# Part 1 — Official ByteDance doctrine\n\n## What Seedance actually is `[official :1450-1452]`\n\nSeedance 2.0 is a multimodal AI director. It reads text, images, video and audio **simultaneously** and internally decomposes them into two dimensions:\n\n- the **spatial layer** — what is in the frame\n- the **temporal layer** — how things change over time\n\nSo a good prompt is **not \"copywriting-style description\" but an \"engineering-style instruction\"**: who, in what scene, doing what action, how the camera moves, and in what chronological order events occur — delivered respectively to the spatial layer and the temporal layer.\n\n## The advanced formula — 8 slots `[official :1455]`\n\n```text\nprecise subject + action details + scene/environment + lighting & color tone\n+ camera movement + visual style + image quality + constraints\n```\n\n⚠️ There is **no official \"6-step formula.\"** `Subject + Action + Environment + Camera + Style + Constraints` is community branding with no ByteDance source, and it silently drops the **lighting & color tone** and **image quality** slots. Use the 8 slots above.\n\n## Task-type sentence patterns `[official :1389-1425]`\n\nSeedance classifies your request from the phrasing. Use the pattern that matches the task:\n\n| Task | Pattern |\n|---|---|\n| **Image reference** | ``Reference `<Subject_N>` in `<Image_N>` to generate…`` |\n| **Video reference** | ``Reference `<Action / Camera_movement / Style / Sound_effect>` in `<Video_N>` to generate…`` |\n| **Audio reference** | ``Reference the timbre in `<Audio_N>` to generate…`` |\n| **Video edit — modify** | ``Strictly edit `<Video_N>`, and modify `<Original_Characteristic>` in it to `<New_Characteristic>``` |\n| **Video edit — add** | ``<Element_Features>` + `<Timing>` + `<Location>`` |\n| **Video edit — delete** | Name what to delete; for anything that must stay, say so explicitly |\n| **Video extend** | ``Extend `<Video_N>` forward/backward to generate…`` |\n| **Combined** | ``Reference `[Dimension]` of `<Image/Video_N>`, strictly edit `<Video_X>`, `[Specific_Edits]``` |\n\n### ⚠️ Edit / extend phrasing landmine `[official :1431]`\n\n> *\"For edit / extend video tasks, directly use `<Video_N>` to refer to the video. **Do not use \"reference `<Video_N>`\"**, to avoid being incorrectly identified as a reference task.\"*\n\nThis is easy to trip: Slates has an edit lane<!-- slates-only --> (`slates_generate_video` with `videoReferenceAssetId`, plus the Seedance edit/relocate routes)<!-- /slates-only -->. Writing *\"reference video 1 and change the jacket to red\"* gets classified as a **reference** task — the model generates a brand-new clip inspired by the source instead of editing it. Write *\"Strictly edit video 1, and modify the blue jacket to red.\"*\n\n## Shot structure — \"Shot 1 / Shot 2 / Shot 3\" `[official :1563-1598]`\n\n> *\"Use shot order, write a simple 'Shot 1 / Shot 2 / Shot 3' storyboard for each segment of the video, and then merge them into a complete prompt.\"*\n\n**❌ Never second-stamp.** No `0:00–0:03`, no \"At 4 seconds\", no per-segment durations.\n\n> *\"Do not impose strict limits on the duration of each segment; prioritize allowing the model to naturally generate the pacing based on the plot.\"* `[:1580]`\n>\n> *\"The model's support for precise timing (such as 0–3 seconds) is **unstable**, and forcibly limiting duration may lead to **abnormal generation results**.\"* `[:1586]`\n\nOrder shots by when events occur — primary first, secondary later. Let the plot set the pacing.\n\n**Per-shot internal order** `[official :1590-1598]` — organize each shot in exactly this sequence:\n\n1. **Camera movement or shot transition** — \"slowly push in from a wide shot\", \"fixed camera position\", \"cut to…\"\n2. **Subject actions and expressions** — the key actions and expression changes of the core character/object\n3. **Position or spatial change** — where the subject is, and how that relationship shifts\n4. **Audio** — sound effects, voices, background music for that shot\n\n**One primary camera move per shot** — see Camera below. `[official :1648]`\n\n## Subject binding — names + image indexes `[official :1488-1556]`\n\nEvery time a subject appears, it must be **explicitly referred to**. Two supported forms:\n\n- **Undefined subjects** — bind inline every mention: `<Subject_N>@<Image_N>`. Official example: **`Zhang San@Image 1`**. `[:1540]`\n- **Pre-declared subjects** — define once, then reuse the same label verbatim: *\"Define the tall man in **Video 1** as **police officer**, and define the other short man as **thief**\"*, then say \"police officer\" every time after. `[:1514]`\n\n**One subject spread across several assets** — bind them together: *\"Define `[…]` in **Image 1** and `[…]` in **Image 2** as `<Subject N>`.\"* `[:1514]`\n\n⚠️ **An Asset ID must never substitute for `<Image/Video_N>`.** `[:1546]` *\"the model cannot directly associate the Asset ID with the reference content.\"* Always cite by index.\n\nAlso official: keep descriptions concise, avoid redundancy, avoid semantic conflicts (contradictory traits for one subject), and prefer expressing spatial relationships through reference images rather than dense text. `[:1550-1556]`\n\n**`[slates]`** — the app composes this for you. `composeReferences()` cites each canonical character or environment reference inline as `Name (image N)` in the exact order it sends them, which is ByteDance's own duplicate-character format (*\"Zhang San (corresponding to image 1)\"* `[:1976]`). You never hand-write role labels or index numbers.\n\n## Action description `[official :1602-1621]`\n\n- **Body-part specificity + quantified degree.** Name hands, legs, head, shoulders, back — and supplement **range, speed, and force**. *\"slowly raise a hand\", \"quickly turn the head\", \"push hard off the ground\", \"slightly lower the head.\"*\n- **Prioritize slow, gentle, continuous small movements.** Avoid high-burst, large-dynamic actions — sprinting, big jumps, violent rolls. *\"walk slowly\", \"gently raise a hand\", \"sit down naturally with the motion.\"* **This is the official basis for the folk rule that \"fast\" degrades quality** — it is not a banned token, it is a class of motion the model handles badly.\n- **Supplement transitions between actions.** Specify inertia and continuity between consecutive beats so movement reads coherent: *\"use the inertia of turning around to naturally raise a hand\", \"naturally transition from a pause into raising a hand.\"*\n\n## Externalize emotion `[official :1623-1636]`\n\nReplace abstract emotion words (\"very sad\", \"extremely angry\") with **specific physical detail**. This is the highest-leverage single habit in the official guide:\n\n| Abstract | Externalized as actions and details |\n|---|---|\n| **Sadness** | head lowering, shoulders trembling slightly, eyes reddening, fingers unconsciously clutching the corner of clothing, tears welling but not falling |\n| **Joy** | corners of the mouth rising uncontrollably, brows and eyes relaxing, steps becoming light, unconsciously humming a tune |\n| **Nervousness / anxiety** | frequently checking the watch, fingers constantly tapping the tabletop, rapid breathing, eyes darting away |\n| **Anger** | both fists clenched, jawline tense, chest heaving, eyes sharp, squeezing words out through gritted teeth |\n| **Relief** | letting out a long breath, tense shoulders completely relaxing, a faint smile appearing, looking up toward the distance |\n\n## Camera `[official :1643-1648]`\n\n> *\"The model has a **strong understanding of camera movement terms**, so you can **directly use standard camera movement terminology**, such as 'medium shot, close-up, wide shot, slow push-in, smooth lateral tracking, fixed shot.'\"*\n\nThis is an **open vocabulary, not a fixed list** — and it explicitly includes **shot size** (close-up / medium / wide / long shot), which is as much a camera instruction as the move itself.\n\n> ⚠️ *\"Try to specify only 1 type of camera movement in a single shot. Do not require push, pull, pan, and move at the same time, as this will increase image instability.\"* `[:1648]`\n\n## Image quality, style, and constraints `[official :1656-1679]`\n\nThese three slots \"define creative boundaries for the model, unify image quality and artistic tone, and avoid visual flaws and random deviations.\"\n\n**1. Image quality** — define clarity, texture detail, and lighting quality. Official vocabulary: `HD` · `rich details` · `cinematic texture` · `natural colors` · `soft lighting`.\n\n> ⚠️ This is a **real slot with real vocabulary** — do not confuse it with Stable-Diffusion-era quality incantations. `8K` / `masterpiece` / `trending on artstation` remain banned slop tokens (see Part 3); *\"cinematic texture, rich details, natural colors\"* is the officially sanctioned way to ask for the same thing.\n\n**2. Style** — the overall art style and visual tone: `cyberpunk cool blue-purple tone` · `retro film` · `fresh Japanese style`.\n\n**3. Constraint words** — *\"Constraint words are very important. They can effectively avoid visual flaws, deformities, breakdowns, and unreasonable elements.\"* Official templates, verbatim:\n\n- **No subtitles** — \"keep it subtitle-free\" / \"avoid generating any text or subtitles\"\n- **No logo** — \"do not generate a logo\"\n- **No watermark** — \"do not generate a watermark\"\n\nSeedance has **no `negativePrompt` field** — constraints go inline in this slot. See Part 3 for the wider inline-negative kit.\n\n## 🔴 Duplicated characters — the twin problem `[official :1948-1994]`\n\n**Symptom:** in frames with **many characters**, where **three-view / multi-view character images** are supplied as references, two identical characters appear in the same generated frame.\n\n**Root causes** `[:1954-1959]`:\n1. Character subjects are not clearly defined in the prompt, so the model cannot distinguish roles.\n2. *\"When character **three-view / multi-view images** are used as reference assets, it is easy to confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"*\n\n**Official fixes, in their order** `[:1971-1994]` — ByteDance is explicit that these *reduce probability*, not eliminate it:\n\n1. **Bind each character to its image explicitly**, in a consistent format. Official example: *\"Zhang San (corresponding to image 1) throws the green passbook toward Li Si (corresponding to image 2), who is standing.\"*\n2. **Append the global constraint verbatim** at the end of the prompt `[:1982]`:\n   > *\"Throughout the video, characters with completely identical appearance, clothing, and accessories are prohibited. Do not generate duplicate avatars or a twin effect. Keep only a single corresponding character in the same frame, and do not reproduce repeated copies of characters.\"*\n3. **Optimize reference assets** `[:1988]` — *\"For character reference images, prioritize independent single-person photos. Three-view or multi-view assets are not recommended.\"*\n4. **Simplify the prompt** — do not paste a whole script; redundant copy confuses the model.\n\n**Scope this honestly.** This is troubleshooting for the twin problem in **multi-character frames**, not a blanket verdict on identity sheets. Practical rule for Slates:\n\n- **Multi-character Seedance shot** → bind every character to its image, append the anti-twin constraint, and prefer single-person / dominant-portrait references over multi-view sheets.\n- **Single-character shot** → the standard character-sheet flow is fine.\n\n**Too many reference people** `[official :2048-2052]` — past **4 reference people**, output stability drops (wrong headcount, duplicates). Official workaround: group the cast into images of ≤4 people each, generate those stills first, then drive the video from them.\n\n## Worked examples `[official :1689-1745]`\n\nThese are ByteDance's own end-to-end cases. Note the shape: an asset-binding preamble, then `Shot N` blocks in event order, then a trailing style + stability paragraph. No time stamps anywhere — 2.0 does not respond to them at all, which is version-scoped and reverses on 2.5.\n\n**Example 1 — dormitory emotional short drama (dialogue-focused).** Assets: `@Image 1` half-body photo of the female lead · `@Image 2` dormitory scene reference · `@Video 1` camera-movement reference · `@Audio 1` indoor ambience.\n\n> Use the girl in @Image 1 as the main character, use @Image 2 as the dormitory scene style reference, and refer to the camera movement in @Video 1.\n>\n> **Shot 1**: At dusk, **girl @Image 1** walks briskly to the **dormitory entrance @Image 2**. The camera follows steadily in a medium shot. Warm yellow sunlight spills into the hallway from the window. She pauses at the doorway, takes a deep breath, and looks slightly nervous.\n>\n> **Shot 2**: **Girl @Image 1** pushes the door open and enters the dormitory. The camera cuts to an indoor medium shot. Her roommates look up at her while organizing their books. One of them smiles and asks {How did the exam go? Did you pass?}. The camera slowly cuts between half-body close-ups of several people.\n>\n> **Shot 3**: **Girl @Image 1** first lowers her head with a dejected expression. The camera gives her a close-up. Then she raises her head, unable to hold back a smile, laughs out loud, and says {I was kidding}. Her roommates start chasing and play-fighting with her. The camera slowly pulls back and freezes on a wide shot of the dormitory filled with laughter.\n>\n> The entire video should have a high-definition cinematic documentary style, with warm tones and soft lighting. The character's face remains stable without deformation; movements are natural and smooth, with no stutter or flicker. The ambient sound blends naturally with @Audio 1.\n\n**Example 2 — ancient-style cliff confrontation (action/atmosphere-focused).** Assets: `@Image 1` female lead in red · `@Image 2` assassin in black · `@Image 3` cliff and bamboo forest · `@Video 1` martial-arts camera reference · `@Audio 1` drum beats.\n\n> Use the woman in red from @Image 1 as the female lead, use the woman in black from @Image 2 as the opponent, use the cliff and bamboo forest environment in @Image 3 as the scene reference, refer to the overall camera movement and action rhythm in @Video 1, and synchronize the background sound effects with @Audio 1.\n>\n> **Shot 1**: At dusk, the camera slowly pushes in from a side medium shot of **woman in red @Image 1**. She stands at the edge of the cliff and lifts a wine flask to drink. Her sleeves and robe hem sway gently in the mountain wind. The camera circles halfway around her, moving from the front to her back. In the distance, a figure in black is faintly visible in the bamboo forest.\n>\n> **Shot 2**: The camera zooms and fades into a long shot. From a drone perspective, it overlooks the entire cliff and bamboo forest. The two characters stand at opposite ends of the cliff. The mountain wind lifts their robe hems and dust, and the rhythm slightly accelerates with the drum beats.\n>\n> **Shot 3**: The camera cuts back to a ground-level close shot. The two slowly draw their swords and face off. **Woman in red @Image 1** shifts from a careless expression to a cold gaze. **Woman in black @Image 2** looks determined, and the sword tip trembles slightly. The camera steadily follows the two as they circle each other, finally freezing on a close-up of the instant before the two swords meet.\n>\n> The overall visual style should feel like a cinematic wuxia world in misty rain, with cool tones, low saturation, a film-grain texture, and rich light-and-shadow layers. The characters' faces and body proportions remain stable without deformation. Movements are continuous and natural, not stiff, with no clipping or stutter.\n\n## Other official notes\n\n- **On-screen text** `[official :1758]` — Seedance can render common text (ad slogans, subtitles, speech bubbles) and will auto-match style/colour from context, or take an explicit colour / style / timing / position. Prefer **common characters**; avoid rare glyphs and special symbols. (For *guaranteed* legible text, the start-frame route in Part 3 is still safer.)\n- **Extension degrades quality** `[official :2004-2024]` — using a generated video as the input for extension compounds degradation, with mottled colour blocks in face regions. Limit repeated continuations; prefer HD assets as input.\n- **Special effects that miss** `[official :2031-2044]` — when a described effect comes out wrong (a countdown that scrolls randomly), define it with a **reference video** instead of words: *\"the way the number '2999' appears should reference video 1.\"*\n\n---\n\n# Part 2 — Slates-specific `[slates]`\n\n## Reference media — caps and transport\n\nReference-to-video accepts up to **9 reference images, 3 reference videos, 3 audio clips** `[official :275-281]`. Text+audio-only and audio-only inputs are not supported.\n\n**Mutually exclusive:** first-frame/last-frame mode CANNOT be combined with reference images. The error reads `\"first/last frame content cannot be mixed with reference media content.\"` Pick one or the other. *(Official note `[:284]`: you can approximate first/last frames via prompt wording inside a multimodal call, but if the frames must be exact, use the dedicated first/last-frame route.)* The same rule covers reference VIDEO and AUDIO: they ride the reference endpoint, which has no frame parameters at all.\n\n### All three modalities go in ONE call\n\nThe caps are a shared budget, not three separate features: **12 files total on 2.0** (9 image + 3 video + 3 audio), **15 seconds of reference video combined**, **15 seconds of audio combined**. On 2.0 an audio reference needs at least one image or video alongside it; 2.5 accepts audio on its own.\n\nCite each by type and index, in the order they were attached — `image 1`, `video 1`, `audio 1`. The index is positional: reorder the attachments and the numbers move with them.\n\n```\nMarcus (image 1) performs the motion from video 1, in the workshop from image 2,\nspeaking the line in audio 1. Preserve his identity, appearance and outfit.\n```\n<!-- slates-only -->\n**Attaching a clip is NOT the same as editing it.** \"Add as reference\" puts it in the composer alongside everything else and wipes nothing; \"Edit with AI\" makes the clip the canvas and clears the tray for a fresh instruction. Two different jobs, two different menu entries — never infer one from the other.\n\n**Over the cap is REFUSED, never trimmed.** A reference video is priced into the quote before it is sent, so a clip silently dropped after the quote would be a clip you paid for and the model never saw. Remove one and retry.\n<!-- /slates-only -->\n\n### Motion transfer & lip-sync recipes (reference video / audio)\n\nThese aren't separate Seedance features — they're prompting strategies over reference media.<!-- slates-only --> The Slates tools (`slates_generate_motion_transfer` / `slates_generate_lip_sync` with the seedance engine) compose them for you. When driving them by hand through `slates_generate_video`:<!-- /slates-only -->\n\n- **Motion transfer:** subject image as a reference + the driving clip<!-- slates-only --> via `videoReferenceAssetId`<!-- /slates-only --> (2–15s) + `The character from image 1 performs the exact motion, choreography, and camera movement from video 1. Preserve the character's identity, appearance, and outfit.`\n- **Lip-sync / dialogue:** write the line in the prompt — `The person in video 1 says: \"…\"` — with audio generation on (always on in Slates). A **video** source's own voice is cloned natively; an **audio** reference (≤15s) drives speech from an existing recording: `…speaks the dialogue from audio 1 with accurate lip sync.`\n- **Voice + face from one clip (the talking-head recipe):** ONE unedited 2–15s clip of the person speaking (clear voice, no music, no cuts) as the video reference + prompt with the new script → their likeness AND voice deliver the new line.\n<!-- slates-only -->\n- **Billing:** a reference VIDEO switches the cost key to `seedance-2*-vref-{res}-{T}s` where T = clip seconds + output seconds — quote before confirming. Audio references are free (audio is included on every route).\n<!-- /slates-only -->\n\n<!-- slates-only -->\n## Faces — set `seedanceFace` for AI-character faces\n\nSeedance routes through **three tiers** depending on the face in the reference, exposed as the \"Face in Reference\" toggle plus the real-face params on `slates_generate_video`:\n\n- **Faceless / object / environment refs → default route (cheapest).** Leave `seedanceFace` off.\n- **An AI-character's FACE in a reference → `seedanceFace: true`.** The default route's baseline moderation rejects or degrades faces, so this reroutes to the face-capable provider. It costs **~45% more** — the cost key becomes `seedance-2-face-{res}-{N}s`, so the pre-flight quote already reflects it. Announce the face-route price, not the faceless one.\n- **A REAL person's photo (the user themselves, an actor) → the consent-gated premium route.** If a `seedanceFace` gen fails with `[REAL_FACE_DETECTED]`, the provider classified the reference as a real person: confirm with the user that (a) they hold the rights/consent to the likeness and (b) they accept the higher price (cost key `seedance-2-realface-{res}-{N}s`, roughly 2× the AI-face rate — quote via `slates_estimate_generation_cost`), then retry with `seedanceRealFace: true` + `realFaceConsent: true`. Never set `realFaceConsent` without the user's explicit confirmation.\n\nRules:\n- **The real-vs-AI call is the PROVIDER'S, not yours.** ByteDance's classifier is probabilistic — some real photos pass the standard face route (billed at the cheap rate; fine), others get rejected with `[REAL_FACE_DETECTED]` (auto-refunded). Don't preemptively route to the real-face tier just because a photo looks real; try `seedanceFace: true` first and escalate only on the marked rejection. Public figures / celebrities fail on every route.\n- It's about the **reference, not the output.** If your character identity or generated portrait shows a face, turn it on. A product shot with no person stays off.\n- Don't toggle it on \"just in case\" — a faceless gen on the face route burns ~45% extra for nothing.\n<!-- /slates-only -->\n\n## Reference rules (the verified ones)\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### For Seedance specifically\n\n- **Describe the ACTION, never the reference's content.** With refs attached, prompt only what is *happening* — motion, change, camera. Never re-describe what's in the reference, and never say \"still / scene / from a movie / from the image.\" The model already sees the refs; narrating them wastes tokens and induces drift. Injection is stochastic — if a roll misses, **re-roll, don't re-engineer** (and a slow gen is not a failed one<!-- slates-only --> — see slates-cost-discipline<!-- /slates-only -->).\n- **Seedance's own idiom for rule 2 is `Reference <Subject_N> in <Image_N>`** `[official :1389]` — `Image_N` indexes the order the refs are attached, so the name plus the index carries the role. The full binding grammar is in Part 1 (Subject binding).\n- **Rule 3 has an official ceiling here.** The trend is MORE references (video and audio into Seedance), all addressed by name — but for **multi-character frames** see the twin-problem section above: bind every character to its image, append the anti-twin constraint, and prefer single-person references. Past 4 reference people, stability drops `[official :2048-2052]`.\n- **Rule 8 holds even though Seedance can render common text natively** `[official :1758]`. A baked NB2 start frame is still the reliable route for text that must be legible.\n- **Rule 5 pairs with the first/last-frame exclusion** — frames and reference images are mutually exclusive on this model (see Reference media above), so an environment you must match exactly costs you the frame lane.\n\n<!-- slates-only -->\n## Pre-flight: references arrive inline, refer by code\n\nWhen you call `slates_generate_video` with reference asset IDs (firstFrameAssetId, lastFrameAssetId, ingredientAssetIds), the first call returns those references **inline as image content blocks** alongside a cost estimate and `requires_confirm: true`. **Look at the references** — if they suggest a different framing, lighting, or motion than your current prompt captures, revise the prompt before re-calling with `confirm=true`.\n\nWhen talking to the user about the gen, refer to each reference by its short code: `IMG-A12 — Beach Sunset`. The user sees that code as a badge on the gallery thumbnail, so they can match what you're saying to what they're looking at.\n\n- ✅ \"I'm using **IMG-A12** as the first frame and **IMG-A15** as the last frame — the camera move is going to be a slow dolly forward through the gap.\"\n- ❌ \"I'm using the first beach image and the last one...\" (which? They have four.)\n<!-- /slates-only -->\n\n---\n\n# Part 3 — Community field notes `[community]`\n\nThird-party guides and Slates field experience. Useful heuristics — but if one of these ever appears to contradict Part 1, **Part 1 wins**.\n\n## Length\n\n**Sweet spot 60-150 words** for a single shot (not 150-300 — that's the upper bound). Multi-shot storyboards run longer; official Example 1 above is ~230 words across three shots.\n\n## Pin the subject in the first 20-30 words\n\nThe opening sentence is the **identity anchor**. If the subject isn't locked early, the model hallucinates new subjects mid-clip. (Compatible with Part 1: the binding preamble comes before `Shot 1`.)\n\n```\nA matte black earbud case sits on a polished obsidian surface...\n```\n\n## Lighting is a top quality lever\n\nLighting has an outsized impact on output quality — which is why it has its own slot in the official 8-slot formula. Describe it before or alongside the subject.\n\n```\nA cool-white diagonal beam from upper left, dust particles drifting through.\nSoft golden hour lighting from low west angle.\nDramatic rim light against dark background.\n```\n\n## Camera and subject motion — separate sentences\n\nMixing them is a common cause of glitchy / shaky output.\n\n❌ \"The camera speed ramps as the earbud rises.\"\n✅ \"The earbud rises smoothly. The camera tracks upward.\"\n\n## Slow-motion works; \"fast\" is a known bad token\n\nSpeed ramps and slow-motion are supported in natural language, and `fast` is widely reported as a quality-degrading keyword. **The official version of this rule is stronger and better founded** — prioritize slow, gentle, continuous small movements and avoid high-burst action (Part 1, Action description `[:1611-1615]`). Prompt the motion class, not the adjective.\n\n```\nthe lid opens in slow-motion · the blade whips through the air\n```\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ — same contract as the anti-list in slates-prompting-nano-banana-2.\n     Extracted by src/prompts/banned-tokens.ts into the slates_generate_video op\n     description and matched against submitted prompts. The RECOMMENDED vocabulary\n     below sits OUTSIDE the markers on purpose — it is backticked too. -->\n<!-- /slates-only -->\n**Slop tokens to avoid:** `epic`, `amazing`, `beautiful`, `lots of movement`, `8K`, `masterpiece`, `trending on artstation`.\n<!-- @banned:end -->\nThese are quality *incantations* — the officially sanctioned way to ask for quality is the image-quality slot vocabulary in Part 1 (`HD`, `rich details`, `cinematic texture`, `natural colors`, `soft lighting`).\n\n## Style block at the end\n\nOne primary anchor + 2-3 supporting details, as the trailing paragraph (both official examples do exactly this). End with `Single continuous take` if you want one shot with no cuts. **Never** write `no cut` or `seamless transition` — not in the training vocabulary.\n\n## ⚠️ Don't cross-pollinate image-model syntax\n\nNamed **lenses, apertures, film stocks, and camera bodies** — `85mm f/1.4`, `Kodak Portra 400`, `ARRI Alexa 65`, `shot on Sony A7S3` — are an **image-model lever** (correct and encouraged in `slates-prompting-nano-banana-2`) and a **Seedance anti-pattern**. ByteDance's guide uses shot sizes, camera moves, pacing words, and the image-quality/style vocabulary throughout, and never once mentions fps, shutter angle, f-stop, or lens millimetres.\n\nIf you are carrying a look over from an NB2 start frame, translate it: `85mm f/1.4, Portra 400` → `close-up, shallow depth of field, warm natural colors, cinematic texture, film-grain texture`.\n\n## Negative prompting — inline only\n\nSeedance has **no `negativePrompt` field**. Put negatives in the constraints slot, led by the three official templates (Part 1):\n\n```\nkeep it subtitle-free · do not generate a logo · do not generate a watermark\navoid jitter and bent limbs\navoid temporal flicker\navoid identity drift\nno distortion, no stretching\n```\n\nAlso fine: positive reframing (\"empty street\" not \"no cars\").\n\n## Image-to-video / first-frame guidance\n\n**Describe motion, not image.** The model already sees the visual; tokens spent re-describing appearance are wasted.\n\nStability phrases that help:\n- `preserve composition and colors`\n- `maintain exact appearance from reference image`\n- `consistent character throughout, no deformation or drift`\n\n**Cap I2V prompts under 60 words** when possible. Over 100 words frequently triggers silent generation failure.\n\n## Common failure modes + fixes\n\n| Failure | Fix |\n|---|---|\n| Hallucinated subject mid-clip | First 20-30 words = identity anchor |\n| Bent limbs / extra fingers | `avoid jitter and bent limbs` in Constraints |\n| Identity drift across multi-shot | Re-name the bound subject in **every** `Shot N` block `[official :1537]` |\n| Two identical characters in one frame | The twin fix in Part 1 — bind each character to its image + append the global anti-twin constraint |\n| Silent generation failure on I2V | Cut prompt under 100 words, single primary camera move |\n| Speech / motion conflict | Limit dialogue to one line per action shot |\n| Erratic/random pacing | You second-stamped. Remove all time markers and use `Shot N` `[official :1586]` |\n\n## Sources\n\n**Official (authoritative):**\n- BytePlus ModelArk — Seedance 2.0 prompting guide, archived at `research/seedance-2-modelark-docs.md` (all `:NNNN` refs above)\n\n**Community (secondary):**\n- [fal.ai — How to Use Seedance 2.0](https://fal.ai/learn/tools/how-to-use-seedance-2-0)\n- [apiyi.com — Seedance 2.0 Prompt Guide](https://help.apiyi.com/en/seedance-2-0-prompt-guide-video-generation-camera-style-tips-en.html)\n- [atlabs.ai — Ultimate Seedance 2.0 Prompting Guide](https://www.atlabs.ai/blog/the-ultimate-seedance-2.0-prompting-guide-47-prompts-2026)\n",
+    "slates-prompting-seedance-2-5": "---\nname: slates-prompting-seedance-2-5\ndescription: How to prompt Seedance 2.5 and Seedance 2.5 Edit. Read before calling slates_generate_video with model seedance-2.5, or slates_edit_video with model seedance-2.5-edit. 2.5 is a SECOND SEAT next to 2.0, not an upgrade — it buys 30-second takes, 30 image references, audio-only references and INTEGER-SECOND TIMESTAMPS, and it gives up native 4K and costs more than 2.0 at every resolution they share. Timestamps are the one grammar difference that matters: 2.0 ignores them and answers only to shot numbers, 2.5 acts on them. Otherwise it shares 2.0's grammar (read slates-prompting-seedance for subject binding, camera and constraint vocabulary); this file covers what is different, plus the two hazards unique to 2.5 — the prompt-intent task classifier and the cost trap that comes with 30-second takes.\n---\n\n# Seedance 2.5 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Seedance 2.5.** Shares 2.0's grammar exactly (subject binding, camera vocabulary, externalised emotion, inline constraints — read `slates-prompting-seedance` for those). Two things are different, and both matter.\n\n**The five levers**\n1. **Timestamps work here** — integer seconds, and the model acts on them: `[0-4] she reads the letter. [4-9] she folds it and looks up.` 2.0 ignores exactly this syntax.\n2. **Length is the reason to be here** — takes up to 30 seconds, where 2.0 stops at 15. Write the beats as `[0-6]`, `[6-12]`, `[12-18]`; do not hope for them.\n3. **Up to 30 image references**, and a multi-view image can serve as ONE subject reference (up to 5 subjects). 2.0 cannot do either.\n4. **Audio-only references are accepted** without an image or video alongside — the only Seedance seat that takes one.\n5. **Keep the 2.0 discipline**: one camera move per beat (`slow track right`, `handheld follow`), physical action instead of stated emotion, and quality asked for in the image-quality slot vocabulary — `rich details`, `natural colors`, `cinematic texture`, `soft lighting`.\n\n**Examples**\n- `[0-6] Wide shot, <Subject_1>@<Image_1> crosses an empty car park toward a idling van, slow track right. [6-12] Medium, she stops as the driver's window comes down. [12-18] Close-up, she looks off past the lens and does not answer. Rich details, natural colors. Keep it subtitle-free.`\n- `[0-10] A single continuous handheld follow behind a courier climbing a fire escape, rain. [10-20] She reaches the landing, turns, and the city opens behind her. Cinematic texture, soft lighting.`\n\n**Hard constraint:** it is the EXPENSIVE seat and it has NO 4K — 480p/720p/1080p only, and dearer than 2.0 at every resolution they share. It is a second seat, never an upgrade. Long takes multiply cost linearly: quote a 30-second take before you fire it.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** (2.5 reclassifies the task and fails a fresh generation on these):\n- `edit`, `extend`, `continue the video`, `same video but` — they make the provider read a fresh generation as an edit\n- `85mm`, `f/1.4`, `Portra 400` and any other lens, aperture, film-stock or camera-body token — image-model vocabulary, a Seedance anti-pattern on both seats\n<!-- @banned:end -->\n\n**Read `slates-prompting-seedance` first.** The prompt GRAMMAR is the same model family: the\n8-slot advanced formula, subject binding by `<Subject_N>@<Image_N>`, camera vocabulary,\nexternalised emotion, inline constraint words, the anti-twin fix. None of it is restated here.\nThis file is only what 2.5 changes — and the biggest change is that **2.5 acts on timestamps\nwhere 2.0 ignores them.**\n\n---\n\n## The one fact that decides whether you use it at all\n\n**Seedance 2.5 is the EXPENSIVE seat, not the cheap one — and it has no 4K.**\n\nIt runs at 480p, 720p or 1080p (1080p landed on all three routes on 2026-08-24), and at every\nresolution the two seats share it costs MORE than 2.0 — 720p $0.231/s against $0.15/s, **54% more**.\nSo 2.5 does not replace 2.0; it sits beside it, and you pay for what it buys:\n\n| | Seedance 2.0 | Seedance 2.5 |\n|---|---|---|\n| Resolution | 480p / 720p / 1080p / **native 4K** | 480p / 720p / 1080p — **no 4K** |\n| Price at 720p (faceless) | **$0.15/s** | $0.231/s |\n| Length | 4–15s | **4–30s in one take** |\n| Reference budget | 15 (9 image + 3 video + 3 audio) | **50 (30 image + 10 video + 10 audio)** |\n| Combined reference video/audio | ≤15s | **≤30s** |\n| Audio-only reference | ✗ (needs an image or video alongside) | **✓** |\n| **Timestamps in the prompt** | **✗ — ignored; shot numbers only** | **✓ — integer seconds, acted on** |\n| Multi-view image as ONE subject reference | ✗ (not recommended) | **✓ (up to 5 subjects)** |\n| Video edit as its own task type | ✗ | **✓ (`seedance-2.5-edit`)** |\n| Default video model | **yes** | no |\n\n**Route to 2.5 when the shot needs LENGTH, MANY REFERENCES, or an audio-only reference.\nRoute to 2.0 when resolution matters at all** — which, for anything a client will see full-screen,\nis most of the time.\n\n---\n\n## 🚨 Hazard 1 — the prompt-intent task classifier\n\nThis is the one that costs money and time, and it has no equivalent on 2.0.\n\n**Seedance 2.5 sorts every request into one of five task types** — text-to-video,\nreference-to-video, first/last-frame, **video edit**, **video extend** — from the reference roles\nattached **plus the intent of your sentence**. Each type then has its own parameter constraints,\nand a violation comes back **asynchronously**: the task queues, credits are reserved, and only then\ndoes it fail.\n\nThe trigger words are ordinary English:\n\n| Reclassified as | Words that do it (ByteDance's own list) |\n|---|---|\n| **video edit** | `edit video` · `add` · `insert` · `remove` · `delete` · `modify` · `replace` · `change to` |\n| **video extend** | `extend forward` · `extend backward` · `continue` · `continue from` · `extend the story` |\n\nSo a perfectly legitimate reference-to-video prompt — *\"a wide shot of the workshop, **remove** the\ntripod from frame\"* — gets classified as an edit and fails on constraints it never set.\n\n**What to do:**\n\n1. **If you mean to edit an existing clip, say so with the MODEL, not the sentence.** Call\n   `slates_edit_video` with `model: 'seedance-2.5-edit'`. That routes to a dedicated\n   task-typed endpoint and the classifier never has to guess.\n2. **If you mean a fresh shot, describe the finished frame rather than an instruction to change\n   one.** Not *\"remove the tripod\"* → *\"the workshop bench, clear and uncluttered\"*. Not\n   *\"add rain\"* → *\"heavy rain falling through the streetlight\"*. This is better prompting anyway:\n   the model renders what you describe, it does not take edits to an imagined draft.\n3. The trigger only fires when **references are attached**. A plain text-to-video prompt is safe\n   however it is worded.\n\n**Slates will warn you, and it will never rewrite your prompt.** When a 2.5 reference generation's\nprompt contains one of these words, the composer shows a warning that NAMES the words and the agent\nroute returns the same string. Silently editing the user's sentence to dodge a provider classifier\nis forbidden — the words that reach the model are always the words the user can see.\n\n---\n\n## 🚨 Hazard 2 — resolution is not the price dial here. LENGTH is.\n\nEvery other model in Slates trains the habit that lower resolution means lower cost. 2.5 breaks it,\nbecause the thing that moves the bill is **length**, and 2.5's length ceiling is double 2.0's.\n\nWorked, at the shipped rates:\n\n| Generation | Credits |\n|---|---|\n| 2.5 · 480p · 5s · faceless | 26 |\n| 2.5 · 720p · 5s · faceless | 58 |\n| 2.5 · 1080p · 5s · faceless | 103 |\n| 2.5 · 720p · 30s · faceless | 347 |\n| 2.5 · 720p · 30s · AI-face route | **489** |\n| 2.5 · 720p · 30s · consented real-face route | **710** |\n| 2.5 · 1080p · 30s · faceless | **614** |\n| 2.5 · 1080p · 30s · consented real-face route | **1,749** |\n| *(for scale)* 2.0 · 1080p · 15s · AI-face route | 411 |\n\n**A 30-second 720p clip can cost more than a 15-second 1080p one** — and a base licence starts\nwith 1,000 credits. Someone who reads \"720p\" as \"cheap\" and asks for a 30-second take on the\nreal-face route has spent 71% of their welcome grant on one clip; **on the real-face route a single\n30-second 1080p take is more than the whole grant.**\n\n**Discipline:**\n\n- **Always quote with `slates_estimate_generation_cost` before a take over ~10 seconds,** and say\n  the number out loud before generating.\n- **Find the shot at short LENGTH, not at low resolution.** Length is what moves the price, so cut\n  seconds while you are still exploring — 4–8s — and stay at the resolution you actually want.\n  **A 480p pass does not de-risk a 720p or 1080p render.** Generation is stochastic: the higher-\n  resolution run is a different take, not the same shot rendered better. So a 480p draft that looks\n  right buys you no guarantee, and one that looks wrong may have been fine at 720p — you paid 26\n  credits to learn nothing, when 58 would have bought a real candidate.\n- **Length is a creative decision, not a default.** 30 seconds is available; it is rarely the right\n  answer for a single shot. Multi-shot storyboards inside one 30s generation are what the length is\n  actually for.\n- Read `slates-cost-discipline` — all of it applies, more sharply here.\n\n---\n\n## Timestamps — the one grammar change\n\n<!-- @inject:seedance-25-timestamps -->\n**2.0 does not respond to timestamps and answers only to shot numbers. 2.5 responds to\ninteger-second timestamps.** That is ByteDance's own first line under \"Differences from Seedance\n2.0\", and it is why a 30-second take is usable at all: the length is only worth buying if you can\nsay *when* things happen inside it.\n\nBoth formats are valid on 2.5, and you can mix them — `Shot N` blocks for a storyboard whose\npacing you are happy to leave to the model, timestamps when a beat has to land at a moment.\n\n**Three ways to control time, all first-party:**\n\n| Form | Write it like |\n|---|---|\n| **Interval** | `0-3 seconds… 3-7 seconds… 7-15 seconds` or `[1s-4s]… [4s-8s]… [8s-12s]` |\n| **Time point** | *\"Quick left sideways transition at the 5-second mark.\"* |\n| **Relative** | *\"After 3 seconds, everyone around him shakes their head.\"* · *\"The frame freezes for 1 second after he presses the shutter.\"* |\n\n**The rules that come with them:**\n\n- **One second is the smallest unit.** Integers only — no `2.5s`, no frames.\n- **No gaps in the timeline.** `0-3s… 5-6s…` leaves 3-5s unspecified and the model fills it however\n  it likes. Intervals must abut: `0-3s`, `3-7s`, `7-15s`.\n- **Budget the plot to the seconds.** Too little content in a range and the model improvises to\n  fill it; too much and you get extra cuts or dropped beats. This is the actual craft of a 30s take.\n- **Never time-code a high-frequency action.** *\"Shake your head three times per second\"* is\n  explicitly called out as a misuse — timestamps schedule beats, they don't choreograph frames.\n- **Transitions want both halves:** the moment AND the method — *\"At the 5-second mark, the camera\n  transitions leftward with a left wipe into a natural dissolve.\"*\n- **Timestamps work on an EDIT too**, and that is where they earn the most: they scope a change in\n  time as well as in content — *\"Change the man's action from drinking coffee to mopping the floor\n  from 4-6 seconds in Video 1, and leave the rest of the content unchanged.\"* Without a range, a\n  whole-clip instruction is applied to the whole clip.\n\nDo **not** carry this back to 2.0, and do not carry Veo's `[00:00-00:02]` bracket syntax into\neither — 2.0 ignores time entirely, and the cross-model syntax swap is its own known failure.\n<!-- @end:seedance-25-timestamps -->\n\n---\n\n## What the extra reference budget is actually for\n\n30 image references (up from 9) does **not** mean \"attach 30 images\". Every rule in\n`slates-prompting-seedance` about references still holds — 2–4 strong references beat both\nextremes, and one reference per role.\n\n**Where 2.5 moves the ceiling, per ByteDance's own input recommendations:**\n\n| | Stable | Works, but expect re-rolls |\n|---|---|---|\n| Subjects bound by IMAGE reference | 1–8 | 9–12 |\n| Subjects bound by VIDEO or AUDIO reference | 1–5 | 6–10 |\n| Reference clip length, per subject | 5–10s | longer drops stability |\n\n**Multi-view images of one subject are supported on 2.5** — a turnaround sheet can be a single\nreference image, where 2.0 wanted one authoritative rendering per subject. Past **5 subjects**,\ngo back to single-view images, one per view, rather than one image carrying several viewpoints.\n\nThe larger budget earns its keep in exactly two places:\n\n- **A long multi-shot take** where different shots need different subjects and locations bound —\n  the budget is spread across the storyboard, not stacked on one frame.\n- **Video and audio references alongside images**, which is where 2.5's 10 + 10 matters far more\n  than the image count.\n\n### Audio-only references — the genuinely new input\n\n2.0 required an image or video alongside any audio reference. **2.5 accepts audio on its own.**\nThat makes one recipe possible that was not before: drive a scene's timing, voice or ambience from\na recording with no visual anchor at all — a voice line, a music bed, a room tone — and let the\nmodel build the picture to it. Cite it the same way as any other reference\n(`Reference the timbre in <Audio_N> to generate…`), and remember that audio references carry **no\nbilling dimension** on any Seedance route: audio is included.\n\n### Video references\n\nUp to 10 clips, ≤30s combined (2.0: 3 clips, ≤15s). A reference VIDEO switches the cost key to\n`seedance-2.5*-vref-{res}-{T}s`, where **T = Σ input seconds + output seconds** — the sum is across\n**every** clip attached, not just the longest. Three 6-second references on a 12-second output bills\n30 seconds, not 12 and not 18. Quote before confirming.\n\n**The cap is a refusal, not a trim.** Attach an eleventh clip, or push past 30 combined seconds, and\nthe composition is rejected before anything uploads. That asymmetry is deliberate: reference images\nwarn-and-trim because dropping one doesn't change the price, and a dropped reference VIDEO would be\none you were quoted for and the model never saw.\n\n### Mixing all three in one call\n\n50 files total (30 image + 10 video + 10 audio) is a shared budget. Everything is cited positionally\nby type — `image 1`, `video 2`, `audio 1` — in attachment order, so reordering the attachments\nrenumbers the citations. Write the prompt against those numbers:\n\n```\nMarcus (image 1) performs the motion from video 1 in the workshop from image 2,\nusing the voice timbre from audio 1. Preserve his identity, appearance and outfit.\n```\n\n🚨 **SAY WHAT AN AUDIO REFERENCE IS FOR.** It can mean music, dialogue, voice, tone or timbre — five roles on one attachment — so an unroled clip falls back to **dialogue**: the model re-transcribes it and speaks ITS words. A real take came back as *\"a map called Slates\"* for *\"an app called Slates\"*. Name it as the voice timbre and the clip carries the voice while the prompt carries the words. ByteDance's own sentence: *\"Image 1 depicts the protagonist John and uses the voice timbre from Audio 1.\"* Bind each speaker in a sentence, never by attachment order — position carries nothing.\n\nFrames and reference media stay mutually exclusive, in every combination — the reference endpoint\nhas no first/last-frame parameters at all, so this is a shape mismatch rather than a preference.\n\n---\n\n## Seedance 2.5 Edit (`slates_edit_video`, `model: 'seedance-2.5-edit'`)\n\nIts own picker row and its own op call, deliberately: the task type is **the model you chose**,\nnever something inferred from your sentence.\n\n**Why route here at all:** it is the **only edit engine in Slates that accepts a clip longer than 15\nseconds** (4–30s, versus Kling O3 Edit's 3–15s and Omni Flash Edit's 3–10s). For a clip inside the\nothers' range, choose on fidelity instead — Omni Flash Edit won the prompt-only head-to-head, and\nKling O3 Edit is the one that takes element and style reference images.\n\n**How it behaves:**\n\n- **Output length follows the SOURCE clip**, and the bill is the ceiled source length. The provider\n  requires an automatic duration on this task type, so there is no length knob — the clip you attach\n  is the quote. The returned clip can differ from the source by up to ~0.3s, which only compresses\n  transition frames; a clip that 2.5 itself generated comes back at exactly its input length.\n- **Source clips under 20 seconds edit more reliably.** 4–30s is what the task type accepts;\n  ByteDance's own recommendation is to stay inside 20 for quality. A 28-second source is legal and\n  will need more attempts.\n- **The aspect ratio follows the source clip too.** No ratio control; the frame is the clip's frame.\n- **480p, 720p or 1080p output**, native audio — and an edit bills the video-reference tier ×2,\n  so 1080p on this row is the most expensive second in the app. Quote it.\n- **Prompt and source clip only** on this op. The MODEL takes reference images on an edit\n  (ByteDance recommends 1–5 — *\"replace the man in dark clothing in @Video 1 with @Image 2\"*);\n  **Slates has not wired that path**, so today an edit that must lock an identity from a photo\n  goes to Kling O3 Edit. Constraint of our build, not of the model — worth revisiting.\n- **An edit bills roughly DOUBLE a plain 2.5 generation of the same length**, because every provider\n  charges an edit on input + output seconds. Read the confirm gate's number; do not reason from the\n  generation rate.\n- **Set `seedanceFace: true` when a character's face is visible in the clip.** The faceless provider\n  blocks faces outright — this is not a price optimisation, it is whether the job runs at all.\n- **There is no consented-real-face route for editing.** Real-person footage that the AI-face route\n  rejects has to go to Kling O3 Edit.\n\n**Prompting an edit** — the same discipline as every other edit engine: **describe only what\nchanges.** The source already carries its composition, motion, timing and performance; re-describing\nthem fights the model. Use Seedance's own edit grammar from `slates-prompting-seedance`\n(*\"Strictly edit `<Video_1>`, and modify `<Original_Characteristic>` to `<New_Characteristic>`\"*) and\n**never** write *\"reference video 1\"* in an edit — the official guide is explicit that this phrasing\ngets the request reclassified as a reference task, which is the same landmine as Hazard 1.\n\nTwo things sharpen an edit prompt, both first-party:\n\n- **Say it as A → B, not as an outcome.** *\"Change the man's action from drinking coffee to mopping\n  the floor\"* beats *\"the man mops the floor\"* — naming what it currently is tells the model what\n  to overwrite.\n- **Timestamp a partial edit** — the edit task type reads the same integer-second timestamps the\n  generation path does. Rules and forms are in § Timestamps above; this is the single most useful\n  thing they buy.\n\n**Audio is editable too, and it is the least obvious use of this row.** The same op rewrites what\nis heard while the picture stays put: change a spoken line, change the accent, translate the\ndialogue and re-fit the lip movement, strip or replace the BGM or a sound effect. *\"Only edit the\nman's dialogue in Video 1: change it to 'Don't come over here,' in an American accent\"* is an edit,\nnot a lip-sync job. Bill it like any other edit — on the source clip's length.\n\n---\n\n## Faces, and what does NOT change\n\nThe three-tier face routing is identical to 2.0 — faceless → default route, an AI character's face →\n`seedanceFace: true` (the relaxed provider, a real cost premium), a real person's photo → the\nconsent-gated premium route after a `[REAL_FACE_DETECTED]` rejection, with `realFaceConsent: true`\nset **only** after the user explicitly confirms they hold the rights to the likeness. The full rules,\nincluding why the real-vs-AI call is the provider's and not yours, are in\n`slates-prompting-seedance`.\n\nAlso unchanged, and worth restating because 2.5's length makes each one more expensive to get wrong:\n\n- **One primary camera move per shot.**\n- **No lens / aperture / film-stock vocabulary.** That is image-model syntax and a Seedance\n  anti-pattern.\n- **No `negativePrompt` field** — constraints go inline, and 2.5 acts on negative phrasing in\n  exactly two dimensions: subtitles (*\"no subtitles\"*) and audio (*\"no BGM; environmental and\n  action sounds only\"*, *\"no audio\"*). Everywhere else, describe what you want, not what you don't.\n- **Legible in-shot text still belongs in a baked start frame**, not in the video prompt.\n",
+    "slates-prompting-seedance": "---\nname: slates-prompting-seedance\ndescription: How to prompt Seedance 2.0 (ByteDance video model). Read before calling slates_generate_video with model seedance-2. Seedance 2.0 structures multi-beat prompts as a \"Shot 1 / Shot 2 / Shot 3\" storyboard against an 8-slot advanced formula — never per-second time stamps, which 2.0 does not respond to (Seedance 2.5 does; see slates-prompting-seedance-2-5). Its syntax differs from Kling, Veo and the image models; don't cross-pollinate (in particular, no lens / aperture / film-stock vocabulary).\n---\n\n# Seedance 2.0 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Seedance 2.0.** Not copywriting — an ENGINEERING instruction to a spatial layer and a temporal layer: who, in what scene, doing what, how the camera moves, and in what order. Multi-beat work is a `Shot 1 / Shot 2 / Shot 3` storyboard.\n\n**The five levers**\n1. **Bind every subject to its reference** — `<Subject_1>@<Image_1>` — and keep the descriptions identical across shots. Unbound subjects are where twins come from.\n2. **Shot sizes and camera MOVES, not lens data** — `medium close-up`, `slow push in`, `handheld follow`, `whip pan`. One primary move per shot.\n3. **Externalise emotion as physical action.** Not \"she is nervous\": `she turns the ring on her finger twice, then stops`.\n4. **Use the image-quality slot vocabulary** for quality — `HD`, `rich details`, `cinematic texture`, `natural colors`, `soft lighting`. That is the officially sanctioned way to ask.\n5. **Constraints go INLINE**, led by the official templates — `keep it subtitle-free`, `do not generate a watermark`, `avoid jitter and bent limbs`, `avoid temporal flicker`.\n\n**Examples**\n- `Shot 1: medium shot, <Subject_1>@<Image_1> steps out of the freight lift into a wet loading bay, slow push in. Shot 2: close-up, she turns the ring on her finger twice and stops, handheld. Rich details, cinematic texture, natural colors. Keep it subtitle-free.`\n- `Single continuous take. Wide shot of a fishing skiff crossing a grey swell, camera tracks from the starboard rail. Spray hits the lens once. Soft lighting, natural colors, film-grain texture. Avoid jitter and bent limbs.`\n\n**Hard constraint:** NO timestamps — 2.0 ignores them and answers only to shot numbers (2.5 acts on them). No lens, aperture, film stock or camera body: that is image-model vocabulary and a Seedance anti-pattern. There is no negativePrompt field.\n<!-- @card:end -->\n\nByteDance's video model — first-party via **BytePlus ModelArk** (credits only, no BYOK). Audio always generated alongside the video. Single model `seedance-2` across the full resolution ladder (480p / 720p / 1080p / native 4K — 4K video is Pro-only, default 1080p), 4–15s, first+last frame, and up to 9 reference images / 3 videos / 3 audio clips.\n\n> **How to read this file.**\n> **[official :NNNN]** — ByteDance's own BytePlus ModelArk prompting guide, line `NNNN` of the archived doc dump (`research/seedance-2-modelark-docs.md`). Receipt-grade; treat as law.\n> **[community]** — third-party guides and our own field notes. Useful, but an `[official]` block always wins.\n> **[slates]** — how the Slates app composes or bills this; not ByteDance doctrine.\n>\n> The split is load-bearing. A community-sourced \"narrative timing beats\" doctrine shipped in this file for months teaching the **exact inverse** of ByteDance's published guidance. Never merge the two registers again.\n\n---\n\n# Part 1 — Official ByteDance doctrine\n\n## What Seedance actually is `[official :1450-1452]`\n\nSeedance 2.0 is a multimodal AI director. It reads text, images, video and audio **simultaneously** and internally decomposes them into two dimensions:\n\n- the **spatial layer** — what is in the frame\n- the **temporal layer** — how things change over time\n\nSo a good prompt is **not \"copywriting-style description\" but an \"engineering-style instruction\"**: who, in what scene, doing what action, how the camera moves, and in what chronological order events occur — delivered respectively to the spatial layer and the temporal layer.\n\n## The advanced formula — 8 slots `[official :1455]`\n\n```text\nprecise subject + action details + scene/environment + lighting & color tone\n+ camera movement + visual style + image quality + constraints\n```\n\n⚠️ There is **no official \"6-step formula.\"** `Subject + Action + Environment + Camera + Style + Constraints` is community branding with no ByteDance source, and it silently drops the **lighting & color tone** and **image quality** slots. Use the 8 slots above.\n\n## Task-type sentence patterns `[official :1389-1425]`\n\nSeedance classifies your request from the phrasing. Use the pattern that matches the task:\n\n| Task | Pattern |\n|---|---|\n| **Image reference** | ``Reference `<Subject_N>` in `<Image_N>` to generate…`` |\n| **Video reference** | ``Reference `<Action / Camera_movement / Style / Sound_effect>` in `<Video_N>` to generate…`` |\n| **Audio reference** | ``Reference the timbre in `<Audio_N>` to generate…`` |\n| **Video edit — modify** | ``Strictly edit `<Video_N>`, and modify `<Original_Characteristic>` in it to `<New_Characteristic>``` |\n| **Video edit — add** | ``<Element_Features>` + `<Timing>` + `<Location>`` |\n| **Video edit — delete** | Name what to delete; for anything that must stay, say so explicitly |\n| **Video extend** | ``Extend `<Video_N>` forward/backward to generate…`` |\n| **Combined** | ``Reference `[Dimension]` of `<Image/Video_N>`, strictly edit `<Video_X>`, `[Specific_Edits]``` |\n\n### ⚠️ Edit / extend phrasing landmine `[official :1431]`\n\n> *\"For edit / extend video tasks, directly use `<Video_N>` to refer to the video. **Do not use \"reference `<Video_N>`\"**, to avoid being incorrectly identified as a reference task.\"*\n\nThis is easy to trip: Slates has an edit lane<!-- slates-only --> (`slates_generate_video` with `videoReferenceAssetId`, plus the Seedance edit/relocate routes)<!-- /slates-only -->. Writing *\"reference video 1 and change the jacket to red\"* gets classified as a **reference** task — the model generates a brand-new clip inspired by the source instead of editing it. Write *\"Strictly edit video 1, and modify the blue jacket to red.\"*\n\n## Shot structure — \"Shot 1 / Shot 2 / Shot 3\" `[official :1563-1598]`\n\n> *\"Use shot order, write a simple 'Shot 1 / Shot 2 / Shot 3' storyboard for each segment of the video, and then merge them into a complete prompt.\"*\n\n**❌ Never second-stamp.** No `0:00–0:03`, no \"At 4 seconds\", no per-segment durations.\n\n> *\"Do not impose strict limits on the duration of each segment; prioritize allowing the model to naturally generate the pacing based on the plot.\"* `[:1580]`\n>\n> *\"The model's support for precise timing (such as 0–3 seconds) is **unstable**, and forcibly limiting duration may lead to **abnormal generation results**.\"* `[:1586]`\n\nOrder shots by when events occur — primary first, secondary later. Let the plot set the pacing.\n\n**Per-shot internal order** `[official :1590-1598]` — organize each shot in exactly this sequence:\n\n1. **Camera movement or shot transition** — \"slowly push in from a wide shot\", \"fixed camera position\", \"cut to…\"\n2. **Subject actions and expressions** — the key actions and expression changes of the core character/object\n3. **Position or spatial change** — where the subject is, and how that relationship shifts\n4. **Audio** — sound effects, voices, background music for that shot\n\n**One primary camera move per shot** — see Camera below. `[official :1648]`\n\n## Subject binding — names + image indexes `[official :1488-1556]`\n\nEvery time a subject appears, it must be **explicitly referred to**. Two supported forms:\n\n- **Undefined subjects** — bind inline every mention: `<Subject_N>@<Image_N>`. Official example: **`Zhang San@Image 1`**. `[:1540]`\n- **Pre-declared subjects** — define once, then reuse the same label verbatim: *\"Define the tall man in **Video 1** as **police officer**, and define the other short man as **thief**\"*, then say \"police officer\" every time after. `[:1514]`\n\n**One subject spread across several assets** — bind them together: *\"Define `[…]` in **Image 1** and `[…]` in **Image 2** as `<Subject N>`.\"* `[:1514]`\n\n⚠️ **An Asset ID must never substitute for `<Image/Video_N>`.** `[:1546]` *\"the model cannot directly associate the Asset ID with the reference content.\"* Always cite by index.\n\nAlso official: keep descriptions concise, avoid redundancy, avoid semantic conflicts (contradictory traits for one subject), and prefer expressing spatial relationships through reference images rather than dense text. `[:1550-1556]`\n\n**`[slates]`** — the app composes this for you. `composeReferences()` cites each canonical character or environment reference inline as `Name (image N)` in the exact order it sends them, which is ByteDance's own duplicate-character format (*\"Zhang San (corresponding to image 1)\"* `[:1976]`). You never hand-write role labels or index numbers.\n\n## Action description `[official :1602-1621]`\n\n- **Body-part specificity + quantified degree.** Name hands, legs, head, shoulders, back — and supplement **range, speed, and force**. *\"slowly raise a hand\", \"quickly turn the head\", \"push hard off the ground\", \"slightly lower the head.\"*\n- **Prioritize slow, gentle, continuous small movements.** Avoid high-burst, large-dynamic actions — sprinting, big jumps, violent rolls. *\"walk slowly\", \"gently raise a hand\", \"sit down naturally with the motion.\"* **This is the official basis for the folk rule that \"fast\" degrades quality** — it is not a banned token, it is a class of motion the model handles badly.\n- **Supplement transitions between actions.** Specify inertia and continuity between consecutive beats so movement reads coherent: *\"use the inertia of turning around to naturally raise a hand\", \"naturally transition from a pause into raising a hand.\"*\n\n## Externalize emotion `[official :1623-1636]`\n\nReplace abstract emotion words (\"very sad\", \"extremely angry\") with **specific physical detail**. This is the highest-leverage single habit in the official guide:\n\n| Abstract | Externalized as actions and details |\n|---|---|\n| **Sadness** | head lowering, shoulders trembling slightly, eyes reddening, fingers unconsciously clutching the corner of clothing, tears welling but not falling |\n| **Joy** | corners of the mouth rising uncontrollably, brows and eyes relaxing, steps becoming light, unconsciously humming a tune |\n| **Nervousness / anxiety** | frequently checking the watch, fingers constantly tapping the tabletop, rapid breathing, eyes darting away |\n| **Anger** | both fists clenched, jawline tense, chest heaving, eyes sharp, squeezing words out through gritted teeth |\n| **Relief** | letting out a long breath, tense shoulders completely relaxing, a faint smile appearing, looking up toward the distance |\n\n## Camera `[official :1643-1648]`\n\n> *\"The model has a **strong understanding of camera movement terms**, so you can **directly use standard camera movement terminology**, such as 'medium shot, close-up, wide shot, slow push-in, smooth lateral tracking, fixed shot.'\"*\n\nThis is an **open vocabulary, not a fixed list** — and it explicitly includes **shot size** (close-up / medium / wide / long shot), which is as much a camera instruction as the move itself.\n\n> ⚠️ *\"Try to specify only 1 type of camera movement in a single shot. Do not require push, pull, pan, and move at the same time, as this will increase image instability.\"* `[:1648]`\n\n## Image quality, style, and constraints `[official :1656-1679]`\n\nThese three slots \"define creative boundaries for the model, unify image quality and artistic tone, and avoid visual flaws and random deviations.\"\n\n**1. Image quality** — define clarity, texture detail, and lighting quality. Official vocabulary: `HD` · `rich details` · `cinematic texture` · `natural colors` · `soft lighting`.\n\n> ⚠️ This is a **real slot with real vocabulary** — do not confuse it with Stable-Diffusion-era quality incantations. `8K` / `masterpiece` / `trending on artstation` remain banned slop tokens (see Part 3); *\"cinematic texture, rich details, natural colors\"* is the officially sanctioned way to ask for the same thing.\n\n**2. Style** — the overall art style and visual tone: `cyberpunk cool blue-purple tone` · `retro film` · `fresh Japanese style`.\n\n**3. Constraint words** — *\"Constraint words are very important. They can effectively avoid visual flaws, deformities, breakdowns, and unreasonable elements.\"* Official templates, verbatim:\n\n- **No subtitles** — \"keep it subtitle-free\" / \"avoid generating any text or subtitles\"\n- **No logo** — \"do not generate a logo\"\n- **No watermark** — \"do not generate a watermark\"\n\nSeedance has **no `negativePrompt` field** — constraints go inline in this slot. See Part 3 for the wider inline-negative kit.\n\n## 🔴 Duplicated characters — the twin problem `[official :1948-1994]`\n\n**Symptom:** in frames with **many characters**, where **three-view / multi-view character images** are supplied as references, two identical characters appear in the same generated frame.\n\n**Root causes** `[:1954-1959]`:\n1. Character subjects are not clearly defined in the prompt, so the model cannot distinguish roles.\n2. *\"When character **three-view / multi-view images** are used as reference assets, it is easy to confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"*\n\n**Official fixes, in their order** `[:1971-1994]` — ByteDance is explicit that these *reduce probability*, not eliminate it:\n\n1. **Bind each character to its image explicitly**, in a consistent format. Official example: *\"Zhang San (corresponding to image 1) throws the green passbook toward Li Si (corresponding to image 2), who is standing.\"*\n2. **Append the global constraint verbatim** at the end of the prompt `[:1982]`:\n   > *\"Throughout the video, characters with completely identical appearance, clothing, and accessories are prohibited. Do not generate duplicate avatars or a twin effect. Keep only a single corresponding character in the same frame, and do not reproduce repeated copies of characters.\"*\n3. **Optimize reference assets** `[:1988]` — *\"For character reference images, prioritize independent single-person photos. Three-view or multi-view assets are not recommended.\"*\n4. **Simplify the prompt** — do not paste a whole script; redundant copy confuses the model.\n\n**Scope this honestly.** This is troubleshooting for the twin problem in **multi-character frames**, not a blanket verdict on identity sheets. Practical rule for Slates:\n\n- **Multi-character Seedance shot** → bind every character to its image, append the anti-twin constraint, and prefer single-person / dominant-portrait references over multi-view sheets.\n- **Single-character shot** → the standard character-sheet flow is fine.\n\n**Too many reference people** `[official :2048-2052]` — past **4 reference people**, output stability drops (wrong headcount, duplicates). Official workaround: group the cast into images of ≤4 people each, generate those stills first, then drive the video from them.\n\n## Worked examples `[official :1689-1745]`\n\nThese are ByteDance's own end-to-end cases. Note the shape: an asset-binding preamble, then `Shot N` blocks in event order, then a trailing style + stability paragraph. No time stamps anywhere — 2.0 does not respond to them at all, which is version-scoped and reverses on 2.5.\n\n**Example 1 — dormitory emotional short drama (dialogue-focused).** Assets: `@Image 1` half-body photo of the female lead · `@Image 2` dormitory scene reference · `@Video 1` camera-movement reference · `@Audio 1` indoor ambience.\n\n> Use the girl in @Image 1 as the main character, use @Image 2 as the dormitory scene style reference, and refer to the camera movement in @Video 1.\n>\n> **Shot 1**: At dusk, **girl @Image 1** walks briskly to the **dormitory entrance @Image 2**. The camera follows steadily in a medium shot. Warm yellow sunlight spills into the hallway from the window. She pauses at the doorway, takes a deep breath, and looks slightly nervous.\n>\n> **Shot 2**: **Girl @Image 1** pushes the door open and enters the dormitory. The camera cuts to an indoor medium shot. Her roommates look up at her while organizing their books. One of them smiles and asks {How did the exam go? Did you pass?}. The camera slowly cuts between half-body close-ups of several people.\n>\n> **Shot 3**: **Girl @Image 1** first lowers her head with a dejected expression. The camera gives her a close-up. Then she raises her head, unable to hold back a smile, laughs out loud, and says {I was kidding}. Her roommates start chasing and play-fighting with her. The camera slowly pulls back and freezes on a wide shot of the dormitory filled with laughter.\n>\n> The entire video should have a high-definition cinematic documentary style, with warm tones and soft lighting. The character's face remains stable without deformation; movements are natural and smooth, with no stutter or flicker. The ambient sound blends naturally with @Audio 1.\n\n**Example 2 — ancient-style cliff confrontation (action/atmosphere-focused).** Assets: `@Image 1` female lead in red · `@Image 2` assassin in black · `@Image 3` cliff and bamboo forest · `@Video 1` martial-arts camera reference · `@Audio 1` drum beats.\n\n> Use the woman in red from @Image 1 as the female lead, use the woman in black from @Image 2 as the opponent, use the cliff and bamboo forest environment in @Image 3 as the scene reference, refer to the overall camera movement and action rhythm in @Video 1, and synchronize the background sound effects with @Audio 1.\n>\n> **Shot 1**: At dusk, the camera slowly pushes in from a side medium shot of **woman in red @Image 1**. She stands at the edge of the cliff and lifts a wine flask to drink. Her sleeves and robe hem sway gently in the mountain wind. The camera circles halfway around her, moving from the front to her back. In the distance, a figure in black is faintly visible in the bamboo forest.\n>\n> **Shot 2**: The camera zooms and fades into a long shot. From a drone perspective, it overlooks the entire cliff and bamboo forest. The two characters stand at opposite ends of the cliff. The mountain wind lifts their robe hems and dust, and the rhythm slightly accelerates with the drum beats.\n>\n> **Shot 3**: The camera cuts back to a ground-level close shot. The two slowly draw their swords and face off. **Woman in red @Image 1** shifts from a careless expression to a cold gaze. **Woman in black @Image 2** looks determined, and the sword tip trembles slightly. The camera steadily follows the two as they circle each other, finally freezing on a close-up of the instant before the two swords meet.\n>\n> The overall visual style should feel like a cinematic wuxia world in misty rain, with cool tones, low saturation, a film-grain texture, and rich light-and-shadow layers. The characters' faces and body proportions remain stable without deformation. Movements are continuous and natural, not stiff, with no clipping or stutter.\n\n## Other official notes\n\n- **On-screen text** `[official :1758]` — Seedance can render common text (ad slogans, subtitles, speech bubbles) and will auto-match style/colour from context, or take an explicit colour / style / timing / position. Prefer **common characters**; avoid rare glyphs and special symbols. (For *guaranteed* legible text, the start-frame route in Part 3 is still safer.)\n- **Extension degrades quality** `[official :2004-2024]` — using a generated video as the input for extension compounds degradation, with mottled colour blocks in face regions. Limit repeated continuations; prefer HD assets as input.\n- **Special effects that miss** `[official :2031-2044]` — when a described effect comes out wrong (a countdown that scrolls randomly), define it with a **reference video** instead of words: *\"the way the number '2999' appears should reference video 1.\"*\n\n---\n\n# Part 2 — Slates-specific `[slates]`\n\n## Reference media — caps and transport\n\nReference-to-video accepts up to **9 reference images, 3 reference videos, 3 audio clips** `[official :275-281]`. Text+audio-only and audio-only inputs are not supported.\n\n**Mutually exclusive:** first-frame/last-frame mode CANNOT be combined with reference images. The error reads `\"first/last frame content cannot be mixed with reference media content.\"` Pick one or the other. *(Official note `[:284]`: you can approximate first/last frames via prompt wording inside a multimodal call, but if the frames must be exact, use the dedicated first/last-frame route.)* The same rule covers reference VIDEO and AUDIO: they ride the reference endpoint, which has no frame parameters at all.\n\n### All three modalities go in ONE call\n\nThe caps are a shared budget, not three separate features: **12 files total on 2.0** (9 image + 3 video + 3 audio), **15 seconds of reference video combined**, **15 seconds of audio combined**. On 2.0 an audio reference needs at least one image or video alongside it; 2.5 accepts audio on its own.\n\nCite each by type and index, in the order they were attached — `image 1`, `video 1`, `audio 1`. The index is positional: reorder the attachments and the numbers move with them.\n\n```\nMarcus (image 1) performs the motion from video 1, in the workshop from image 2,\nusing the voice timbre from audio 1. Preserve his identity, appearance and outfit.\n```\n\n🚨 **SAY WHAT AN AUDIO REFERENCE IS FOR.** It can mean music, dialogue, voice, tone or timbre — five roles on one attachment — so an unroled clip falls back to **dialogue**: the model re-transcribes it and speaks ITS words. A real take came back as *\"a map called Slates\"* for *\"an app called Slates\"*. Name it as the voice timbre and the clip carries the voice while the prompt carries the words. ByteDance's own sentence: *\"Image 1 depicts the protagonist John and uses the voice timbre from Audio 1.\"* Bind each speaker in a sentence, never by attachment order — position carries nothing.\n<!-- slates-only -->\n**Attaching a clip is NOT the same as editing it.** \"Add as reference\" puts it in the composer alongside everything else and wipes nothing; \"Edit with AI\" makes the clip the canvas and clears the tray for a fresh instruction. Two different jobs, two different menu entries — never infer one from the other.\n\n**Over the cap is REFUSED, never trimmed.** A reference video is priced into the quote before it is sent, so a clip silently dropped after the quote would be a clip you paid for and the model never saw. Remove one and retry.\n<!-- /slates-only -->\n\n### Motion transfer & lip-sync recipes (reference video / audio)\n\nThese aren't separate Seedance features — they're prompting strategies over reference media.<!-- slates-only --> The Slates tools (`slates_generate_motion_transfer` / `slates_generate_lip_sync` with the seedance engine) compose them for you. When driving them by hand through `slates_generate_video`:<!-- /slates-only -->\n\n- **Motion transfer:** subject image as a reference + the driving clip<!-- slates-only --> via `videoReferenceAssetId`<!-- /slates-only --> (2–15s) + `The character from image 1 performs the exact motion, choreography, and camera movement from video 1. Preserve the character's identity, appearance, and outfit.`\n- **Lip-sync / dialogue:** write the line in the prompt — `The person in video 1 says: \"…\"` — with audio generation on (always on in Slates). A **video** source's own voice is cloned natively; an **audio** reference (≤15s) drives speech from an existing recording: `…speaks the dialogue from audio 1 with accurate lip sync.`\n- **Voice + face from one clip (the talking-head recipe):** ONE unedited 2–15s clip of the person speaking (clear voice, no music, no cuts) as the video reference + prompt with the new script → their likeness AND voice deliver the new line.\n<!-- slates-only -->\n- **Billing:** a reference VIDEO switches the cost key to `seedance-2*-vref-{res}-{T}s` where T = clip seconds + output seconds — quote before confirming. Audio references are free (audio is included on every route).\n<!-- /slates-only -->\n\n<!-- slates-only -->\n## Faces — set `seedanceFace` for AI-character faces\n\nSeedance routes through **three tiers** depending on the face in the reference, exposed as the \"Face in Reference\" toggle plus the real-face params on `slates_generate_video`:\n\n- **Faceless / object / environment refs → default route (cheapest).** Leave `seedanceFace` off.\n- **An AI-character's FACE in a reference → `seedanceFace: true`.** The default route's baseline moderation rejects or degrades faces, so this reroutes to the face-capable provider. It costs **~45% more** — the cost key becomes `seedance-2-face-{res}-{N}s`, so the pre-flight quote already reflects it. Announce the face-route price, not the faceless one.\n- **A REAL person's photo (the user themselves, an actor) → the consent-gated premium route.** If a `seedanceFace` gen fails with `[REAL_FACE_DETECTED]`, the provider classified the reference as a real person: confirm with the user that (a) they hold the rights/consent to the likeness and (b) they accept the higher price (cost key `seedance-2-realface-{res}-{N}s`, roughly 2× the AI-face rate — quote via `slates_estimate_generation_cost`), then retry with `seedanceRealFace: true` + `realFaceConsent: true`. Never set `realFaceConsent` without the user's explicit confirmation.\n\nRules:\n- **The real-vs-AI call is the PROVIDER'S, not yours.** ByteDance's classifier is probabilistic — some real photos pass the standard face route (billed at the cheap rate; fine), others get rejected with `[REAL_FACE_DETECTED]` (auto-refunded). Don't preemptively route to the real-face tier just because a photo looks real; try `seedanceFace: true` first and escalate only on the marked rejection. Public figures / celebrities fail on every route.\n- It's about the **reference, not the output.** If your character identity or generated portrait shows a face, turn it on. A product shot with no person stays off.\n- Don't toggle it on \"just in case\" — a faceless gen on the face route burns ~45% extra for nothing.\n<!-- /slates-only -->\n\n## Reference rules (the verified ones)\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### For Seedance specifically\n\n- **Describe the ACTION, never the reference's content.** With refs attached, prompt only what is *happening* — motion, change, camera. Never re-describe what's in the reference, and never say \"still / scene / from a movie / from the image.\" The model already sees the refs; narrating them wastes tokens and induces drift. Injection is stochastic — if a roll misses, **re-roll, don't re-engineer** (and a slow gen is not a failed one<!-- slates-only --> — see slates-cost-discipline<!-- /slates-only -->).\n- **Seedance's own idiom for rule 2 is `Reference <Subject_N> in <Image_N>`** `[official :1389]` — `Image_N` indexes the order the refs are attached, so the name plus the index carries the role. The full binding grammar is in Part 1 (Subject binding).\n- **Rule 3 has an official ceiling here.** The trend is MORE references (video and audio into Seedance), all addressed by name — but for **multi-character frames** see the twin-problem section above: bind every character to its image, append the anti-twin constraint, and prefer single-person references. Past 4 reference people, stability drops `[official :2048-2052]`.\n- **Rule 8 holds even though Seedance can render common text natively** `[official :1758]`. A baked NB2 start frame is still the reliable route for text that must be legible.\n- **Rule 5 pairs with the first/last-frame exclusion** — frames and reference images are mutually exclusive on this model (see Reference media above), so an environment you must match exactly costs you the frame lane.\n\n<!-- slates-only -->\n## Pre-flight: references arrive inline, refer by code\n\nWhen you call `slates_generate_video` with reference asset IDs (firstFrameAssetId, lastFrameAssetId, ingredientAssetIds), the first call returns those references **inline as image content blocks** alongside a cost estimate and `requires_confirm: true`. **Look at the references** — if they suggest a different framing, lighting, or motion than your current prompt captures, revise the prompt before re-calling with `confirm=true`.\n\nWhen talking to the user about the gen, refer to each reference by its short code: `IMG-A12 — Beach Sunset`. The user sees that code as a badge on the gallery thumbnail, so they can match what you're saying to what they're looking at.\n\n- ✅ \"I'm using **IMG-A12** as the first frame and **IMG-A15** as the last frame — the camera move is going to be a slow dolly forward through the gap.\"\n- ❌ \"I'm using the first beach image and the last one...\" (which? They have four.)\n<!-- /slates-only -->\n\n---\n\n# Part 3 — Community field notes `[community]`\n\nThird-party guides and Slates field experience. Useful heuristics — but if one of these ever appears to contradict Part 1, **Part 1 wins**.\n\n## Length\n\n**Sweet spot 60-150 words** for a single shot (not 150-300 — that's the upper bound). Multi-shot storyboards run longer; official Example 1 above is ~230 words across three shots.\n\n## Pin the subject in the first 20-30 words\n\nThe opening sentence is the **identity anchor**. If the subject isn't locked early, the model hallucinates new subjects mid-clip. (Compatible with Part 1: the binding preamble comes before `Shot 1`.)\n\n```\nA matte black earbud case sits on a polished obsidian surface...\n```\n\n## Lighting is a top quality lever\n\nLighting has an outsized impact on output quality — which is why it has its own slot in the official 8-slot formula. Describe it before or alongside the subject.\n\n```\nA cool-white diagonal beam from upper left, dust particles drifting through.\nSoft golden hour lighting from low west angle.\nDramatic rim light against dark background.\n```\n\n## Camera and subject motion — separate sentences\n\nMixing them is a common cause of glitchy / shaky output.\n\n❌ \"The camera speed ramps as the earbud rises.\"\n✅ \"The earbud rises smoothly. The camera tracks upward.\"\n\n## Slow-motion works; \"fast\" is a known bad token\n\nSpeed ramps and slow-motion are supported in natural language, and `fast` is widely reported as a quality-degrading keyword. **The official version of this rule is stronger and better founded** — prioritize slow, gentle, continuous small movements and avoid high-burst action (Part 1, Action description `[:1611-1615]`). Prompt the motion class, not the adjective.\n\n```\nthe lid opens in slow-motion · the blade whips through the air\n```\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ — same contract as the anti-list in slates-prompting-nano-banana-2.\n     Extracted by src/prompts/banned-tokens.ts into the slates_generate_video op\n     description and matched against submitted prompts. The RECOMMENDED vocabulary\n     below sits OUTSIDE the markers on purpose — it is backticked too. -->\n<!-- /slates-only -->\n**Slop tokens to avoid:** `epic`, `amazing`, `beautiful`, `lots of movement`, `8K`, `masterpiece`, `trending on artstation`.\n<!-- @banned:end -->\nThese are quality *incantations* — the officially sanctioned way to ask for quality is the image-quality slot vocabulary in Part 1 (`HD`, `rich details`, `cinematic texture`, `natural colors`, `soft lighting`).\n\n## Style block at the end\n\nOne primary anchor + 2-3 supporting details, as the trailing paragraph (both official examples do exactly this). End with `Single continuous take` if you want one shot with no cuts. **Never** write `no cut` or `seamless transition` — not in the training vocabulary.\n\n## ⚠️ Don't cross-pollinate image-model syntax\n\nNamed **lenses, apertures, film stocks, and camera bodies** — `85mm f/1.4`, `Kodak Portra 400`, `ARRI Alexa 65`, `shot on Sony A7S3` — are an **image-model lever** (correct and encouraged in `slates-prompting-nano-banana-2`) and a **Seedance anti-pattern**. ByteDance's guide uses shot sizes, camera moves, pacing words, and the image-quality/style vocabulary throughout, and never once mentions fps, shutter angle, f-stop, or lens millimetres.\n\nIf you are carrying a look over from an NB2 start frame, translate it: `85mm f/1.4, Portra 400` → `close-up, shallow depth of field, warm natural colors, cinematic texture, film-grain texture`.\n\n## Negative prompting — inline only\n\nSeedance has **no `negativePrompt` field**. Put negatives in the constraints slot, led by the three official templates (Part 1):\n\n```\nkeep it subtitle-free · do not generate a logo · do not generate a watermark\navoid jitter and bent limbs\navoid temporal flicker\navoid identity drift\nno distortion, no stretching\n```\n\nAlso fine: positive reframing (\"empty street\" not \"no cars\").\n\n## Image-to-video / first-frame guidance\n\n**Describe motion, not image.** The model already sees the visual; tokens spent re-describing appearance are wasted.\n\nStability phrases that help:\n- `preserve composition and colors`\n- `maintain exact appearance from reference image`\n- `consistent character throughout, no deformation or drift`\n\n**Cap I2V prompts under 60 words** when possible. Over 100 words frequently triggers silent generation failure.\n\n## Common failure modes + fixes\n\n| Failure | Fix |\n|---|---|\n| Hallucinated subject mid-clip | First 20-30 words = identity anchor |\n| Bent limbs / extra fingers | `avoid jitter and bent limbs` in Constraints |\n| Identity drift across multi-shot | Re-name the bound subject in **every** `Shot N` block `[official :1537]` |\n| Two identical characters in one frame | The twin fix in Part 1 — bind each character to its image + append the global anti-twin constraint |\n| Silent generation failure on I2V | Cut prompt under 100 words, single primary camera move |\n| Speech / motion conflict | Limit dialogue to one line per action shot |\n| Erratic/random pacing | You second-stamped. Remove all time markers and use `Shot N` `[official :1586]` |\n\n## Sources\n\n**Official (authoritative):**\n- BytePlus ModelArk — Seedance 2.0 prompting guide, archived at `research/seedance-2-modelark-docs.md` (all `:NNNN` refs above)\n\n**Community (secondary):**\n- [fal.ai — How to Use Seedance 2.0](https://fal.ai/learn/tools/how-to-use-seedance-2-0)\n- [apiyi.com — Seedance 2.0 Prompt Guide](https://help.apiyi.com/en/seedance-2-0-prompt-guide-video-generation-camera-style-tips-en.html)\n- [atlabs.ai — Ultimate Seedance 2.0 Prompting Guide](https://www.atlabs.ai/blog/the-ultimate-seedance-2.0-prompting-guide-47-prompts-2026)\n",
     "slates-prompting-seedream-5-lite": "---\nname: slates-prompting-seedream-5-lite\ndescription: How to prompt Seedream 5 Lite (ByteDance image model — the cheap volume option in Slates). Read before calling slates_generate_image with model seedream-5-lite, or slates_edit_image with editModel seedream-5-lite. Seedream front-loads attention, likes 30-100 focused words, and takes quoted strings for in-image text.\n---\n\n# Seedream 5 Lite — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Seedream 5 Lite.** The cheap volume seat: flat-priced at every resolution, which makes it the right default for storyboard passes, variant grids and look-dev. Structure, most important first: `Subject + Style + Composition + Lighting/Atmosphere + Technical`.\n\n**The five levers**\n1. **Lead with the subject.** Earlier words weigh more; close with the camera and technical detail.\n2. **Keep it 30-100 words.** Unlike models that reward verbosity, this one gets confused by very long prompts. Focused beats exhaustive.\n3. **Name the composition** — `symmetrical composition`, `rule of thirds`, `foreground detail with blurred background`, `overhead perspective`.\n4. **Name the light as a named condition** — `golden hour`, `dramatic side lighting`, `soft diffused light`, `moody low-key`, `bright high-key`.\n5. **Quote in-image text.** It takes quoted strings for posters and layouts, which is half of why it is the drafting seat.\n\n**Examples**\n- `Professional headshot of a female CEO, short blonde hair, confident expression, navy suit, neutral office background. Studio lighting, shallow depth of field, high-end corporate photography, shot on 85mm.`\n- `A rain-soaked night market stall, cinematic, rule of thirds with the vendor camera-right, foreground steam blurred, moody low-key lighting with practical neon, shot on 35mm.`\n\n**Hard constraint:** it is the DRAFTING seat, not the hero seat. Explore here, then re-run the winner on Nano Banana 2 or FLUX.2 Max for the locked shot.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use:**\n- `masterpiece`, `best quality`, `highly detailed`, `8k`, `award-winning` — quality incantations, not description\n- a prompt past about 100 words: this model gets confused by very long prompts, and focused beats exhaustive\n<!-- @banned:end -->\n\nByteDance's Seedream image model, Lite tier, routed via fal.ai. In Slates: `slates_generate_image` with `model: seedream-5-lite` (REQUIRES projectId — no headless path). **Flat-priced regardless of resolution** — the cheapest image model in Slates, which makes it the right default for high-volume drafting, storyboard exploration, and variant grids. Call `slates_estimate_generation_cost` for the current number; never quote prices from memory. Less censored than Nano Banana 2.\n\n**When to pick it:** lots of frames cheap (storyboard passes, 3-4 variant exploration), posters/layouts with text, quick look-dev. Step up to NB2 or FLUX.2 Max for the locked hero shot.\n\n## Core structure — five components, most important first\n\n```\nSubject + Style + Composition + Lighting/Atmosphere + Technical parameters\n```\n\nSeedream weights concepts mentioned **earlier in the prompt** more heavily. Lead with the subject; close with camera/technical details.\n\n**Length sweet spot: 30-100 words.** Unlike models that reward verbosity, Seedream gets confused by very long prompts. Focused beats exhaustive.\n\n## Style, composition, lighting vocabulary it responds to\n\n- **Style:** portrait photography, macro photography, cinematic, photorealistic, minimalist, oil painting, watercolor, digital art\n- **Composition:** symmetrical composition, rule of thirds, foreground detail with blurred background, wide-angle view, overhead perspective, medium shot, close-up\n- **Lighting:** golden hour lighting, dramatic side lighting, soft diffused light, moody low-key lighting, bright high-key lighting\n- **Technical:** shot on 85mm lens, shallow depth of field, high resolution\n\n## Worked examples\n\n**Portrait:**\n> \"Professional headshot of a female CEO with short blonde hair, confident expression, wearing a navy blue suit, neutral office background, studio lighting, shallow depth of field, high-end corporate photography style\"\n\n**Product:**\n> \"Modern smartphone floating in space, dark background with subtle blue gradient, product photography, studio lighting highlighting the glossy screen, ultra-detailed, commercial quality, photorealistic rendering\"\n\n## In-image text: double-quote it\n\nPut the exact string in double quotation marks — Seedream treats quoted text as render-this-verbatim:\n\n```\nA minimalist poster with the headline \"SUMMER SALE\" in bold sans-serif, centered\n```\n\nSeedream is one of the stronger models for layout-heavy work (posters, mockups, diagrams): call out the layout explicitly — \"centered headline, subtitle beneath, clean margins.\"\n\n## Edits: change one thing, lock the rest\n\nVia `slates_edit_image` with `editModel: seedream-5-lite`. Seedream edits respond well to instructions that name the change AND the preserved elements:\n\n```\nChange the bag to brown leather. Keep the person's face, pose, and the room unchanged.\n```\n\nNote: Seedream edits in Slates ignore extra `referenceAssetIds` — that path is Nano Banana 2 only.\n\n## Common failure modes + fixes\n\n| Failure | Fix |\n|---|---|\n| Subject inconsistent / mutates | Put the subject description first; break complex subjects into clear components |\n| Style drift | Reinforce the aesthetic with 2-3 related terms (\"cinematic, photorealistic, shallow depth of field\") |\n| Compositional confusion | Use photography terms (\"medium shot,\" \"overhead view\"); simplify the scene |\n| Garbled text | Double-quote the exact string; keep it short; state placement |\n| Mushy long-prompt output | Cut to under 100 words — Seedream rewards focus, not volume |\n\n## Iterate cheap, lock expensive\n\nFlat pricing makes Seedream the iterate-fast model: run the 3-strike loop here (draft → evaluate inline → one specific delta → regenerate), and only re-render the winning composition on a pricier model if the project's hero shot demands it. Cost rules live in `slates-cost-discipline` — the batch-authorization pattern applies when generating variant grids.\n\n## Sources\n\n- [fal.ai — Seedream Prompt Guide](https://fal.ai/learn/devs/seedream-v4-5-prompt-guide)\n- [BytePlus ModelArk — Seedream Prompt Guide](https://docs.byteplus.com/en/docs/ModelArk/1829186)\n",
     "slates-prompting-veo-3": "---\nname: slates-prompting-veo-3\ndescription: How to prompt Veo 3.1 (Google). Read before calling slates_generate_video with veo-3.1-fast or veo-3.1-standard. Veo is a NICHE pick, never the default (route per slates-model-selection — Kling is the general default, Seedance the premium tier) — reach for it only when native synchronized audio must generate WITH the video in one gen. 16:9 or 9:16, 4/6/8s. Different cinematography formula than Seedance/Kling. (no subtitles) is mandatory after every dialogue line.\n---\n\n# Veo 3.1 — prompting\n\n<!-- @card:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Everything between the @card markers is extracted by\n     src/prompts/craft-cards.ts and returned on every cost estimate for this\n     model, so it is the ONE piece of positive craft guidance the agent cannot\n     skip. Measured 2026-08-30: a fact inlined where it cannot be skipped moved\n     compliance 0/8 to 30/32; the same guidance behind a fetch moved nothing.\n     Keep it under 2,400 characters (the build fails above that) and keep the\n     rationale, the receipts and the worked examples in the body below. -->\n<!-- /slates-only -->\n**Card — Veo 3.1.** Google's formula, in order: `[Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]`. Sweet spot 50-150 words; the official benchmark is about 50.\n\n**The five levers**\n1. **Open with the cinematography** — `Medium shot`, `low angle`, `aerial`, `dolly in`, `rack focus`, `vertigo effect`. Veo reads the first clause as the camera.\n2. **Texture words counter the AI-plastic look** — `fine skin pores`, `visible fabric weave`, `subtle contrast, no gloss or sharpening`. Name materials concretely: `charcoal cotton hoodie`, `matte concrete`, `silk lapel`.\n3. **Weight verbs stop floaty motion** — `trudges`, `drops heavily`, and a ground contact: `boots crunch on gravel`.\n4. **Terse voice direction only** — `says in a weary voice`, `whispers`, `mutters`. Veo is far less responsive to long voice blocks than Kling.\n5. **Always include an ambience line.** Without one the mix feels dead. `Soft office ambience.` `Wind on the open ridge.` And SFX always carries a cause: `SFX: thunder cracks in the distance`, never `SFX: thunder`.\n\n**Examples**\n- `Medium shot, a tired founder rubbing her temples in front of a bulky monitor in a cluttered office late at night. Harsh fluorescent overheads and the green glow of the screen. Fine skin pores, visible fabric weave on a charcoal cotton hoodie. Soft office ambience, a fan hum. Retro, slightly grainy.`\n- `Low angle, a farrier trudges across a wet yard carrying a shoeing box, boots crunching on gravel. Overcast north light, matte concrete and wet steel. Wind and distant livestock. He says in a weary voice, \"One more and we're done.\" (no subtitles).`\n\n**Hard constraint:** `(no subtitles)` after EVERY dialogue line you do not want burned in as text. Negatives are NOUNS, not instructions — `wall, frame`, never `no walls`. And do not cross syntaxes: Seedance's `single continuous take` suppresses Veo's cuts, and Veo timestamps in a Seedance prompt cause drift.\n<!-- @card:end -->\n\n<!-- @banned:start -->\n<!-- slates-only -->\n<!-- MACHINE-READ. Every `backticked` token between the @banned markers is\n     extracted by src/prompts/banned-tokens.ts and returned on this model's cost\n     estimate, and every submitted prompt is matched against it. Keep entries\n     backticked and prose outside the backticks. -->\n<!-- /slates-only -->\n**Never use** (each one is spelled out above):\n- `single continuous take` — Seedance's phrase; in a Veo prompt it suppresses the cuts you asked for\n- `no subtitles` is REQUIRED after dialogue, but instruction-shaped negatives are not: `no walls`, `no man-made structures`, `don't show` — negatives are NOUNS here\n- `SFX: thunder` and any label-only effect — every effect carries a cause and a distance\n<!-- @banned:end -->\n\nGoogle DeepMind's video model. Two tiers: `veo-3.1-fast` (cheaper, quick) and `veo-3.1-standard` (higher quality). 4k variants exist for both (4K video requires Slates Pro).\n\n**Native single-shot duration: 4, 6, or 8 seconds** — and **8s only** at 1080p or 4K, or whenever you attach reference images (that endpoint is 8s-fixed). 4s and 6s exist at 720p, text-to-video or single-start-frame only. Longer durations require chaining clips via Extend / last-frame reuse — quality degrades if naively requested past 8s in a single generation. Aspect ratio: **16:9 or 9:16** on the route Slates uses. `slates_generate_video` REFUSES anything outside these before submit and names the legal set — nothing is silently ignored or downgraded.\n\nNative synchronized audio at 48kHz: dialogue, SFX, ambient — generated WITH video, not added after.\n\n## Official Google formula\n\n```\n[Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]\n```\n\nSweet spot length: 50-150 words. Cloud's official benchmark is ~50 words.\n\nVerbatim official benchmark:\n> \"Medium shot, a tired corporate worker, rubbing his temples in exhaustion, in front of a bulky 1980s computer in a cluttered office late at night. The scene is lit by the harsh fluorescent overhead lights and the green glow of the monochrome monitor. Retro aesthetic, shot as if on 1980s color film, slightly grainy.\"\n\n## Cinematography vocabulary (Vertex AI docs)\n\n**Lenses:** wide-angle, telephoto, fisheye, anamorphic, 35mm, 85mm, shallow/deep depth of field\n\n**Lighting:** Rembrandt lighting, volumetric lighting, backlighting, golden hour glow, lens flare, rack focus, **vertigo effect** (dolly zoom)\n\n**Camera moves:** dolly (in/out), truck (left/right), pan, tilt, crane, aerial/drone, handheld, whip pan, arc shot, zoom\n\n## Texture-realism phrases (counter the AI-plastic look)\n\n```\nfine skin pores · visible fabric weave · subtle contrast, no gloss or sharpening\n```\n\nSpecify materials concretely: `charcoal cotton hoodie`, `matte concrete`, `silk lapel`. Generic \"smooth, beautiful\" rendering is the failure mode you're avoiding.\n\n## Dialogue — `(no subtitles)` is mandatory\n\nEvery dialogue line you don't want burned in as text overlay needs `(no subtitles)`. Verbatim from the founder talking-head benchmark:\n\n```\nThe founder says, \"This update cuts setup time in half, helping teams get started faster.\" (no subtitles).\n```\n\nWithout this, Veo will overlay subtitle text on top of your generation.\n\n## Voice direction — keep it terse\n\nVeo is less responsive to long voice-direction blocks than Kling. Use brief modifiers:\n\n```\nsays in a weary voice\nwhispers\nshouts\nmutters\n```\n\nMulti-character: handles 2-3 speakers natively. Past 3, sync degrades — use first-frame/last-frame chaining for 4+.\n\n## SFX with cause\n\n```\n✅ SFX: thunder cracks in the distance\n❌ SFX: thunder\n```\n\nAlways specify direction or distance.\n\n## Ambient is mandatory\n\nAlways include an ambience line per scene. Without it, the audio mix feels dead.\n\n```\nSoft office ambience.\nWind on the open ridge.\nDistant city hum.\n```\n\n## First-frame + last-frame workflow (Veo's strength)\n\n1. Generate start frame (Gemini 2.5 Flash Image is the recommended pair — Slates' Nano Banana 2 works)\n2. Generate end frame\n3. Animate with both frames as anchors\n\n**Motion-Lock hack:** Keep ~60% of the same background pixels between start and end frames. Prevents latent drift across the clip.\n\nVerbatim arc-shot example:\n> \"The camera performs a smooth 180-degree arc shot, starting with the front-facing view of the singer and circling around her to seamlessly end on the POV shot from behind her on stage. The singer sings 'when you look me in the eyes, I can see a million stars.'\"\n\n## Ingredients-to-Video (multiple references)\n\nVerbatim example:\n> \"Using the provided images for the detective, the woman, and the office setting, create a medium shot of the detective behind his desk. He looks up at the woman and says in a weary voice, 'Of all the offices in this town, you had to walk into mine.'\"\n\n## Reference discipline (character / environment refs)\n\n<!-- @inject:references-read-literally -->\n> **The general law: the model reads a reference literally.**\n> A reference image is not a suggestion. Whatever is baked into it — lighting, medium, texture, symmetry, competing identities — is read as a **property of the subject** and reproduced downstream. A baked rim light tints every shot made from that sheet. A sheet that looks like a 3D game render gets animated like game footage. Two competing renderings of one face get averaged into a third face.\n\nEvery reference rule below is a corollary of that one sentence, which is why \"prep the reference\" beats \"prompt around the reference\" every time:\n\n- **Flat, plain identity refs** — because scene lighting in the sheet becomes scene lighting in the output (Slates' own receipt: a studio-lit sheet produced a subject that looked green-screen-pasted in front of mountains).\n- **One authoritative rendering per subject** — because the model cannot tell which panel is the real one. ByteDance documents this failure directly: multi-view character assets \"confuse the model's character recognition, causing it to generate duplicate characters of the same appearance.\"\n- **No 3D-game-render look in a reference** — the model recognizes the render mood and inherits its motion character, so the *animation* comes out looking like game footage. This is not a taste rule; it is the same literal-reading mechanism applied to the temporal layer.\n- **Break perfect symmetry** — mirrored faces and dead-square framing read as synthetic, and the model preserves that reading rather than correcting it.\n\n**What this means in practice:** when output is wrong in a way that tracks the *subject* rather than the *scene* — the lighting is wrong the same way in every shot, the face drifts, the material looks synthetic everywhere — fix the reference, not the prompt. Prompting around a baked-in property is the expensive way to lose.\n<!-- @end:references-read-literally -->\n\n<!-- @inject:reference-rules-core -->\nIdentity = a few flat-lit neutral angles; one reference per role, named inline; 2-4 refs not 12; describe environments instead of feeding a grid.\n\n1. **2-4 strong references beat both extremes.** Not 1 (warps toward itself), not 12 (averages worse). Start with 2-3 focused refs — each one adds context AND another variable to balance.\n2. **One reference per ROLE, named in the prompt** — identity / style-grade / environment. The model does **not** infer a reference's role from its position in the list; the inline name carries it. Same-role competitors drift (two \"identity\" refs of different people blend into a third face). Slates composes the naming for you from your `@mentions` / `#tags` — you never hand-write role labels.\n3. **One identity sheet per character, named inline.** A character's identity is a single asset (dominant portrait + body panels), so attach that one asset rather than a pile of views: **fewer competing renderings of a face is better, because the model cannot tell which one is authoritative and averages them.** Slates cites it as `Marcus (image 1)`. **Do NOT hand-write a \"Reference Image Instructions\" block or role essays** (\"use for identity, ignore the outfit, render a neutral expression\") — that drags the sheet's studio lighting and wardrobe into a scene that asked for neither. The prompt leads; the user's words own wardrobe, expression, lighting, and action.\n4. **Flat-light identity refs.** Prep identity references with flat, even, shadowless lighting on a plain neutral background. A studio-lit or scene-lit character sheet bleeds its lighting into every generation — the failure looks like the subject was green-screen-pasted in front of the location. Reference prep beats prompting here.\n5. **Environment: describe it, don't feed a grid.** Default to describing the location in words and let the model build a space that fits the shot. Reserve an environment reference for a mandatory exact-match, and then use ONE clean establishing image with natural ambient light that reads as the location's real light — never a multi-panel grid fed whole.\n6. **Grids: explore, don't input.** Use grids to explore compositions cheaply, then pick a cell. Never feed a grid back in as a reference — the cells share a split detail budget and were generated jointly, so their flaws propagate.\n7. **Reuse the same refs across every shot** in a sequence. Lock a set and keep it; swapping references mid-sequence causes drift, because the model adapts each reference to the current prompt rather than copying it.\n8. **Legible in-shot text → bake it into a still start frame, never trust text-to-video.** Have an image model render the text, then animate from that locked frame. Video models smear type.\n9. **Working from existing media — describe ONLY what changes.** The source already carries its composition, motion, timing, and performance; re-describing them fights the model. Narrate the delta. (Video lane: restyle your own clip while keeping the performance; delayed-VFX on \"video one\"; marker-object insertion; video-as-reference for a series.)\n10. **Style transforms happen in natural language.** By default the source's artistic medium and visual style are inherited. To change it, add a plain-text instruction (\"anime → real person\"). There are no preset pickers, and there is no style slider.\n<!-- @end:reference-rules-core -->\n\n### For Veo specifically\n\n- **Veo's idiom for rule 2 is plain-English role naming in the sentence itself** — *\"Using the provided images for the detective, the woman, and the office setting, create a medium shot of…\"* (see Ingredients-to-Video above). The role rides in the noun phrase, not in a separate label block.\n- **Rule 8 has a second reason to matter here:** Veo bakes subtitle text into the frame unless every dialogue line carries `(no subtitles)`. Text you did not ask for is the failure mode, not just text you did.\n\n## Negative prompting — nouns, not instructions\n\nVeo has a `negativePrompt` field. **Verbatim Vertex AI rule:**\n> \"Describe unwanted elements as nouns rather than instructions. Use 'wall, frame' instead of 'no walls' or 'don't show walls.'\"\n\nInline: positive reframing in the body too.\n- ✅ `\"a desolate landscape with no buildings or roads\"`\n- ❌ `\"no man-made structures\"`\n\n## Common failure modes + fixes\n\n| Failure | Fix |\n|---|---|\n| Subject identity shifts mid-clip | Front-load identity at prompt start; use material cues (`charcoal canvas`, `cotton`, `silk`) to stabilize |\n| Floaty / weightless motion | Weight verbs (`trudges`, `drops heavily`), ground contact (`boots crunch on gravel`) |\n| AI-plastic look | `fine skin pores`, `visible fabric weave`, `subtle contrast` |\n| Subtitles baked into video | `(no subtitles)` after every dialogue line |\n| Rushed dialogue | Lines fit one natural breath in 8s |\n| Mismatched ambience | Always include an ambience line |\n| Warped geometry | `photorealistic stability` |\n\n## Timestamp shot syntax (for chained / multi-beat scenes)\n\nVeo accepts `[00:00-00:02]` brackets for timed sequences within an 8s clip. **Do NOT cross syntaxes** — Veo timestamps in a Seedance prompt cause subject drift; Seedance \"single continuous take\" in a Veo prompt suppresses cuts.\n\nVerbatim multi-beat:\n> \"[00:00-00:02] Medium shot from behind a young female explorer with a leather satchel and messy brown hair in a ponytail, as she pushes aside a large jungle vine to reveal a hidden path.\n> [00:02-00:04] Reverse shot of the explorer's freckled face, her expression filled with awe as she gazes upon ancient, moss-covered ruins. SFX: The rustle of dense leaves, distant exotic bird calls.\n> [00:04-00:06] Tracking shot following the explorer as she steps into the clearing and runs her hand over the intricate carvings on a crumbling stone wall.\n> [00:06-00:08] Wide, high-angle crane shot, revealing the lone explorer standing small in the center of the vast, forgotten temple complex, half-swallowed by the jungle. SFX: A swelling, gentle orchestral score begins to play.\"\n\n## Benchmark prompt — founder talking head (full)\n\n> \"Camera locked at eye level, medium close-up on a 35mm lens: a startup founder in his late 30s with short black hair and light stubble, wearing a charcoal cotton hoodie, speaking directly to camera, leaning slightly forward as he speaks, lifting one hand to emphasize a point, then relaxing back to neutral, in a quiet office during late afternoon, with blurred monitors glowing faintly in the background, lit by soft daylight from a side window with gentle fill on the opposite side and natural falloff across his face. Style: fine skin pores, visible fabric weave, subtle contrast, no gloss or sharpening. Audio: The founder says, 'This update cuts setup time in half, helping teams get started faster.' (no subtitles). Soft office ambience.\"\n\n## Pre-flight: references arrive inline, refer by code\n\nWhen you call `slates_generate_video` with `firstFrameAssetId` / `lastFrameAssetId` / `ingredientAssetIds`, the first call returns those references **inline as image content blocks** alongside a cost estimate and `requires_confirm: true`. Veo's strongest move is first-frame + last-frame; the pre-flight is where you confirm the two frames actually anchor the motion you wrote. Revise the prompt before `confirm=true` if needed.\n\nWhen talking to the user about the gen, refer to each reference by its short code: `IMG-A12 — Founder Headshot`. The user sees that code as a badge on the gallery thumbnail, so they can match what you're saying to what they're looking at.\n\n- ✅ \"I'm anchoring on **IMG-A12** as the open shot and **IMG-A18** as the close — the 180° arc lands on her looking offscreen left.\"\n- ❌ \"I'm using two of the founder shots...\" (which two? They have six.)\n\n## Sources\n\n- [Google Cloud — Ultimate Prompting Guide for Veo 3.1](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1)\n- [Google DeepMind — Veo Prompt Guide](https://deepmind.google/models/veo/prompt-guide/)\n- [Google Cloud Docs — Vertex AI Video Generation Prompt Guide](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/video-gen-prompt-guide)\n- [Atlas Cloud — Veo 3.1 Master Guide](https://www.atlascloud.ai/blog/guides/google-veo-3-1-guide-master-image-to-video-ai-with-native-sound-and-4k-realism)\n- [Invideo — Veo 3.1 Prompt Guide](https://invideo.io/blog/google-veo-prompt-guide/)\n",
     "slates-restyle-from-blocking": "---\nname: slates-restyle-from-blocking\ndescription: Render one blocking pass as several different visual worlds — live action, 2.5D painted, 2D ink, toybox — matching cut for cut. Use when a client needs style options, when someone wants to see the same edit in another look, or when an approved edit needs a new treatment without re-blocking.\n---\n\n# Restyle — one edit, many worlds\n\nThe commercial payoff of the whole previs workflow, and the reason a blocking file is an asset rather than a step.\n\n## The idea\n\nEvery prompt has two halves:\n\n- **Structure** — cuts, camera, timing, who is where. Lives in the blocking clip. **Never changes.**\n- **Style** — what any of it looks like. Lives in the references and the prompt text. **Changes freely.**\n\nHold the structure, swap the style, and the same edit comes back as live action, painted 2.5D, ink on paper or a toybox — **matching frame for frame across all of them.** Cuts land on the same frames, the car drifts at the same moment, the same head turns at the same beat.\n\nFor anyone pitching work: three visual worlds in a day, off one edit the client has already approved. The foundation is not up for renegotiation, so the conversation is only about look.\n\n## Before you restyle\n\nYou need a blocking clip whose structure you are happy with, and a finished prompt for at least one style (per `slates-blocking-to-prompt`). The first style is the expensive one; every later style is an edit of its text.\n\n## What stays fixed\n\nCopy these across every style **verbatim**. Changing them is what desynchronises the outputs:\n\n- The blocking reference's own contract — that it is the master for all movement, the placement-only clause, the tie-break clause, the disambiguation clause\n- The shot count and every timestamp\n- Every shot's camera position, angle, framing and cut point\n- Screen direction and seating\n- The `HOLD FOR THE FULL TIMELINE` block\n- `videoReferenceAssetIds` and `videoReferenceSecondsEach`\n\nLead each style's prompt with a lock so the style layer cannot leak into the structure:\n\n> VIDEO LOCK — the dominant rule of this prompt: the reference defines 100% of the motion, editing and object choreography. The text below defines only look, materials, locations and effects layered onto that motion. Wherever the text and the video could be read differently about motion, the video decides.\n\n## What changes\n\n| Layer | What you swap |\n|---|---|\n| Rendering style | photoreal · painted 2.5D · 2D ink · miniature/toybox |\n| Characters | different sheets entirely — a couple, grandparents, a robot and a cat |\n| Locations | the same four beats set in a different world |\n| Time of day / weather | night after rain · golden hour · hard noon |\n| Lighting and colour | per style |\n| Audio | SFX-only, or scored, or lip-synced dialogue |\n\nCharacters can change species and still land, because the blocking only supplies where a body is and how it moves.\n\n## Dummy mapping — the mechanism that makes it work\n\nEach style needs its own explicit mapping from grey proxy to real object. The proxy is a slot; the style fills it:\n\n> DUMMY MAPPING: the front-LEFT sphere-head dummy (with its grey arm at the shifter and grey leg at the pedals) is THE GRANDPA; the front-RIGHT sphere-head dummy is THE GRANDMA; a front-seat dummy together with its loose blocks is that ONE whole person. Blocks on the rear bench are the luggage. The low-poly flying model in SHOT 18 is THE HELICOPTER. The two vehicles behind the hero car in SHOT 19 are THE POLICE CARS.\n\nSame clause per style, different right-hand side. And restate the placement-only rule in style terms:\n\n> The source defines only placement and motion, never appearance: every placeholder becomes the real object its position implies — spheres are always people, cabin blocks are always cases and bags, fully drawn.\n\n## Location continuity\n\nIf the piece travels, name the places and pin each shot to one. Reusing labels across styles keeps the four prompts diffable:\n\n```\nLOCATION CONTINUITY — one journey through four fixed places; each looks\nidentical in every shot where it appears:\n  LOC-A <opening>   LOC-B <middle>   LOC-C <turn>   LOC-D <finale>\n```\n\nThen tag every beat: `SHOT 9 — 7.79-9.33s — LOCKED, LOC-B: <description>`.\n\n**On a piece that visits many places, make the map absolute and countable** — otherwise the model reuses a room it liked and you get the same interior three times:\n\n> The location map is absolute — SEVEN locations, each appearing EXACTLY ONCE, in this exact order: 1) yard 00:00-00:03.3 … 7) rooftop 00:20-00:30. No location ever appears twice, and the three interiors are three COMPLETELY DIFFERENT rooms — different walls, furniture, people and light — never the same room repeated.\n\n## Style references\n\nA style reference is **not a keyframe**, and saying so prevents the model reproducing its composition as a shot:\n\n> STYLE MASTER — defines the painting and rendering style only: hand-painted look with visible brushstrokes, sculpted painterly volumes, textured matte surfaces, dramatic coloured rim light, deep moody shadows. NOT a keyframe, NOT a location to reproduce, NOT a frame that ever appears in the film. Its own subject, framing and composition are never seen in any shot.\n\nA style can also be **text-only** — no reference image at all. Ink and toybox looks usually specify better in words than they match from a still.\n\n## Keep performance inside the existing shots\n\nStyle changes tempt the model to earn new coverage. Refuse it:\n\n> ACTING — inside the existing shots only: performance is visible only at the size and distance the reference already gives it, only where the source already shows a face; everywhere else it reads through posture and hands alone. The performance NEVER earns a new shot, a new angle or a closer framing.\n\n## Text-free worlds\n\nStylised worlds are where invented signage and garbled lettering appear. One clause kills it:\n\n> TEXT-FREE WORLD: every sign is a blank painted shape, every gauge face carries tick marks only, every licence plate is a blank plate.\n\n## Running it\n\nGenerate each style as its own `slates_generate_video` call against the **same** `videoReferenceAssetIds`. Keep them in one project so they sit side by side; name assets by style so the comparison reads at a glance.\n\nQuote the whole set before firing — `slates_estimate_generation_cost` per style — and confirm. Four styles is four generations, not one.\n\n🚨 Never fire a batch of style variants without showing the user the prompts and the total cost first.\n\n## Checklist per style\n\n- [ ] Same blocking asset, same `videoReferenceSecondsEach`\n- [ ] VIDEO LOCK leads the prompt\n- [ ] Every timestamp and shot count identical to style 1\n- [ ] Dummy mapping written for this style's cast\n- [ ] Style reference declared as style-only, or none used\n- [ ] Location labels reused; on a travelling piece the map is absolute and countable\n- [ ] Acting-inside-existing-shots clause present\n- [ ] HOLD block copied verbatim\n- [ ] Cost quoted and confirmed\n\n## Related\n\n`slates-previs-blocking` · `slates-blocking-to-prompt` · `slates-style-prompting` (style vocabulary per model) · `slates-cost-discipline` (batch quoting)\n",
     "slates-shot-variety": "---\nname: slates-shot-variety\ndescription: Use when planning a set of shots, before firing a batch, when the variety counts on a shot list show one bucket dominating, or when a finished piece reads as flat, samey or repetitive despite each shot being individually fine.\n---\n\n# Shot variety — why it all starts to look the same\n\nThe failure this skill exists for is mechanical, not artistic: **every shot is fine and the sequence is dead.** It happens because a model asked ten times in a row, from ten similar prompts, returns ten similar framings — and because the person writing those prompts is thinking about *what is in the shot* rather than *how this shot differs from the one before it*.\n\nSlates counts the sameness for you. It does not judge it. **`slates_list_shots` returns the distribution with every listing** — shot sizes, camera moves, models, durations, and any bucket repeating three or more times in a row — so the numbers are in front of you before you spend anything. Reading them is the job; deciding what to do is yours.\n\n## Read the table as a COLUMN, not as rows\n\nThis is the whole technique, and it is one sentence:\n\n> Sort the list by shot size and count. Sort it by camera move and count. **If push-in is the plurality, or every row says wide, the batch is wrong before a single credit is spent.**\n\nA shot list read row by row always looks fine, because each row was written to be good on its own. The sameness only shows up in the column. That is why the counts ship in the op result rather than in your head.\n\n## What to vary, in order of how much it matters\n\n1. **Shot size.** The single biggest lever, and the one that collapses first. A sequence of seven mediums reads as a slideshow no matter what is in them. Wide → close is a cut; medium → medium is a dissolve nobody asked for.\n2. **Camera move.** Second-biggest, and the one models default to: ask for \"cinematic\" and you get a slow push-in, every time. If five of seven cuts push in, four of them should not.\n3. **Duration.** Rhythm is not decoration. Seven identical 8-second cuts is a metronome. A 3-second cut lands differently *because* the one before it ran twelve.\n4. **Distance between subjects and lens.** A long lens on a close-up and a wide lens on a close-up are different shots, and the model will render the difference.\n5. **Location and cast.** If three cuts happen in the same room with the same person, that is a scene — fine. If *nine* do, the piece has one idea.\n\n## When repetition is deliberate — and it often is\n\nDo not treat the counts as a defect list. Repetition is a technique with two real uses:\n\n- **A repeated frame IS the joke, or the point.** Three identical wides with one thing changed each time is a gag structure, and varying them would destroy it.\n- **A locked-off frame is a choice.** Static, static, static, then a move — the move only lands because the first three did not have one.\n\nThe check catches a bucket dominating. It cannot tell whether you meant it. If you did, say so and move on; the counts do not block anything and never will.\n\n## Choosing the chop: one long take or several short ones\n\nThis is the decision that owns the rhythm, and Slates puts the price next to it: `slates_split_shot` and `slates_merge_shots` re-cut a board, and the row's duration and quote move as you do it.\n\n- **Merge** when the words run continuously and the picture has no reason to change. One 16-second take on a model that holds up is cheaper to *make* than two 8-second cuts and reads calmer.\n- **Split** when the words keep going and the picture should not. This is the strongest move in the format: one spoken line running unbroken while the visual hard-cuts mid-clause to a new world. Split at a word boundary mid-sentence and both rows carry the same sentence — Slates marks the second as continuing the first, so the script still reads as one line.\n- **Split** also when a line will not fit its cut. Slates flags only lines that cannot be read at *any* plausible pace (above the fastest read in a corpus of 71 real ads), so a flag is never a matter of taste — the chop is genuinely wrong. Splitting the line across two cuts or merging into a longer one both fix it.\n\n## Multi-shot generations count as their cuts, not as one\n\nA model that puts three cuts inside one generation contributes **three** rows to the distribution. That is deliberate: counting generations would score a three-cut clip as a single wide shot and miss exactly the sequences this check exists to catch. Money is counted per generation; rhythm is counted per cut, and the header says which is which.\n\n## What this cannot tell you\n\nIt measures buckets, not taste. Seven varied shot sizes can still be seven boring shots, and a piece that scores perfectly can still be flat. It catches the one mechanical failure — everything starting to look the same — and nothing else.\n\nIt also only sees what is filled in. A board where nobody wrote `shotSize` reports `other` for every cut and tells you nothing. That is correct rather than a gap: inferring a shot size from a prompt would be Slates guessing at your work, and the counts are trustworthy precisely because they are arithmetic on what you actually wrote.\n",
     "slates-storyboard-from-script": "---\nname: slates-storyboard-from-script\ndescription: Turn a script or treatment into a Slates storyboard with scenes and frames. Use when the user has a script, treatment, shot list, or scene-by-scene description and wants to materialize it as a Slates storyboard, optionally generating frame images per shot.\n---\n\n# Storyboard from script — Slates workflow\n\nThe user has a script, treatment, or shot list. You're turning it into a Slates storyboard with scene → frame structure, optionally generating images for each frame.\n\n## Workflow\n\n### 1. Parse the script\nRead the user's script. Decide:\n\n- **Scene count** — usually 1 scene per location/setting change. Don't fragment into one-frame scenes.\n- **Frames per scene** — match the shot list. Default is 3-6 frames per scene unless the script specifies more.\n- **Shot labels** — pull them from the script (e.g., \"Wide\", \"Close-up\", \"Over-the-shoulder\").\n\nIf the user hasn't named the storyboard, suggest one based on the project tone.\n\n### 2. Materialize the structure first (no generation yet)\n- `slates_create_storyboard` with the chosen name.\n- For each scene: `slates_add_scene` with a descriptive name and order.\n- For each shot: `slates_create_shot` with a *visual-only* prompt, the model, the params, and whatever character / environment / style references the project already holds. **Don't generate yet.**\n\n🚨 **A Shot needs no image, and that is the point.** `slates_add_frame` requires an `assetId`, so before Shots existed there was nowhere to put a planned shot until it had been paid for — the plan lived in chat and the user had to trust your memory of it. A Shot is a row: named, listed, priced, forkable, and readable back COMPOSED with `slates_get_shot` before a single credit is spent. Write the plan as Shots, not as sentences you will have to re-type later.\n\n🚨 **A Shot files itself, so pass `sceneId` (or `frameId`) when you know where it goes.** Omit both and it lands in the scene the user has open, else the most recently updated storyboard's last scene, creating a storyboard named after the project if there is none. Nothing you save is ever unfiled — but naming the scene is how the board comes out in the order you wrote it.\n\n🚨 **Shots are addressed by CODE.** Every one comes back as `SHOT-A1`, `SHOT-A2` … per project, monotonic, never reused — the same vocabulary the gallery gives assets. Speak to the user in codes, and pass a code anywhere a `shotId` is taken. It is for pointing at a row THIS session, not for retrieval later: there is no shot search and no shot library, because a Shot is workspace state.\n\n### 2a. Write the SCRIPT COLUMNS as well as the prompt\n\n`slates_create_shot` takes the beat itself, not only the prompt: `line` (what is said, verbatim), `speaker` (a character id, a bare name, or `VO`), `delivery` (the parenthetical), `action` (what happens, screenplay-style), `prop`, `shotSize` and `camera`. All are optional, all are free text, and none of them is sent to a model — they are what the user READS and what the variety check COUNTS.\n\n⚠️ **In this version you state dialogue TWICE, and that is deliberate rather than an oversight.** `line` is the readable script and the input to the fit check; the model only receives what is in the Shot's *prompt*, so the spoken words still go inside that prompt verbatim, with their delivery, in the body on the beat. **Write both in the same call** — then they agree at authoring time and can only drift if a human edits one side.\n\nA speaker who matches no saved character is a working state, not an error: it renders as plain text and groups its lines. Do not create a placeholder character to avoid it.\n\nSurface the planned structure back to the user as a tight summary — `slates_list_shots` gives you the count and the total in one call:\n> Storyboard \"X\" • 4 scenes • 12 shots • 340 credits to fire them all\n>   Scene 1: Forest opening (3 shots)\n>   Scene 2: Confrontation (4 shots)\n>   ...\n\n**Surface a decision log alongside that summary.**\n\n<!-- @inject:decision-log -->\nWhen you surface the plan, include a short **decision log** — one line per decision *you* made that the user did not specify **and that no row already records**:\n\n```\nsource phrase or declared default → what you wrote → what it resolves\n\"in a diner\"        → warm, and the light is the reason           → why the anchor was chosen, not what it is\n(no time of day)    → late afternoon, low warm key                 → default; say the word and it changes\n```\n\n🚨 **Keep it to what is NOT already data — and almost everything now IS.** A Shot holds the references and their roles, the model, every param, the shot size, the camera, the prop, the action and the spoken line, and `slates_list_shots` reads the whole board back in order with its variety counts. Narrating any of those is retelling a row the user can open. **Write the Shot, and let the log carry only the judgement no field holds** — why this world, why this light, why this register.\n\n**Hard rule: never silently add weather, props, style, or camera movement.** Four of those are now FIELDS: put the value on the Shot (`prop`, `camera`, `shotSize`, `action`) so the user can read and change it, and put the *reason* in the log only when you invented it rather than being told it. The rule has not softened — it moved from narration into data, which is stronger, because a field can be corrected and a sentence in chat cannot.\n\n> ❌ **Do NOT turn this into a question gate.** Clarifying questions before optimizing directly fight the locked fast-path rule: *if intent is clear, generate immediately with sane defaults, don't ask questions; only ask for production intent, and batch every question into one message.* Log the decisions, then go. The log is an **output**, not an interrogation — surfaced alongside the plan, never as a separate ceremony, and never as a reason to wait.\n<!-- @end:decision-log -->\n\nTurning a script into *visual* frame prompts means resolving things the script left open — what the room looks like, where the light comes from, how the shot is framed. Those are your decisions, not the writer's; name them.\n\nAsk: **\"Generate frame images now? (y/N)\"**\n\n### 3. Generate frames if requested\n- `slates_generate_from_shots` with every image Shot's id and no `confirm` — it returns ONE itemised quote for the set plus the largest single item. Show that total, get an explicit OK, then re-call with `confirm: true`.\n- It fires the Shots one after another and BLOCKS until the last one lands, so it can outlast the HTTP timeout on a long set. If that happens the run is still going: poll `slates_get_shot` for each Shot's `generationIds` rather than re-firing, which double-spends.\n- Each result returns inline. Evaluate. If one is wrong, fix that Shot (`slates_update_shot`) and re-fire only it — never the set.\n- Bind the keeper to a frame with `slates_add_frame`, then `slates_update_shot` with `attachFrameId` so the recipe and the picture stay together.\n\n### 4. Hand back\n- Total frames generated, total credits spent, storyboard id.\n- Suggest next steps: review via `slates_get_storyboard_with_frames` (it returns every scene, every Shot in order, and the variety distribution), or take the frames to motion — fork each image Shot with `slates_duplicate_shot` (`model:` the video model), give the copy its frame with `slates_update_shot` (`firstFrameAssetId`), then fire the set with `slates_generate_from_shots`. Assemble with `slates_add_clip_to_timeline` in story order and `slates_export_video`. The full frames-to-film pipeline (batch cost authorization, model mixing) is `slates-one-prompt-film`.\n\n## Anti-patterns\n\n- **Don't** auto-generate without asking. Generation is the expensive step. Always confirm first.\n- **Don't** invent shot details the script doesn't mention. If the script says \"they argue,\" ask what the shot looks like, don't fabricate \"she clenches her fists in a wide shot.\"\n- **Don't** mix scene structure and frame generation in one pass — building the skeleton first lets the user catch errors before spending credits.\n- **Don't** write the plan into chat when a Shot can hold it. A Shot is the whole recipe AND the beat — line, speaker, action, prop, framing, references, model, params — and it is the only form of the plan the user can open, price, re-chop and re-fire without you.\n- **Don't** fire a set without reading its variety counts first. `slates_list_shots` returns them with every listing; if one shot size is the plurality or three cuts in a row share a camera move, fix the board before spending. `slates-shot-variety` is the craft behind that.\n",
     "slates-style-prompting": "---\nname: slates-style-prompting\ndescription: Use when the user asks for a visual style (\"make it anime\", \"painterly look\", \"like a Pixar film\"), or when a style has to hold across several shots. Covers how photoreal, anime, painterly and 3d-render are prompted DIFFERENTLY per model, and the style-routing recipe (reference-first, styled start-frame → i2v).\n---\n\n# Per-style prompting (photoreal · anime · painterly · 3d-render)\n\nThe style library (`slates_create_style` / the app's style ids) defines what each style IS. This guide is how to PROMPT each style per model. Derived from `research/style-prompting-research.md` (second-brain) — claims marked *(hypothesis)* are untested; don't present them to users as fact.\n\n## The four ground rules (all styles)\n\n1. **Reference beats adjectives.** A style reference image outperforms prose style instructions. If the user has an on-style image — or you can cheaply generate one — attach it and let the default `inherit` behavior match it. Prose styling is the fallback.\n2. **Style lives in ONE slot per model:**\n   - **Nano Banana 2** — narrative prose; the style is the opening framing of the sentence (\"A hand-drawn 2D anime cel illustration of…\"), never a comma tag.\n   - **Seedance 2.0** — the 8-part formula reserves \"visual style\" (slot 6) and \"image quality\" (slot 7). One clause each. Don't scatter style words through the action text.\n   - **Kling V3** — prose scene direction; style rides the lighting/style tail of Scene → Subject → Action → Camera → Lighting/Style. Tag soup underperforms badly.\n3. **Multi-shot consistency = the SAME style clause, byte-identical, in every shot's prompt** (plus shared references). Paraphrasing the style clause between shots invites drift.\n4. **The styled start-frame is the cheapest reliable style lever for video.** Compose the styled frame in NB2 (cheap), hand it to Seedance/Kling image-to-video, and describe only what CHANGES (motion). Never re-describe the style in the i2v prompt — the frame already encodes it.\n\nNever stack style buzzwords (\"ARRI ALEXA, 35mm, film grain, depth-of-field mastery…\"). One or two register tokens maximum — piles of specs dull the image.\n\n## Photoreal\n\n- **NB2:** never the literal word \"photorealistic\". Describe *a real photograph*: natural skin texture and imperfection, motivated lighting, one lens/film register (\"shot on a 50mm, soft window light\"). Photographic composition terms: wide-angle / macro / low-angle.\n- **Seedance:** put \"sharp focus, natural color, high detail\" in the image-quality slot and always include a lighting clause. Keep motion slow and coherent — fast/burst action is the #1 quality killer and reads most fake in photoreal.\n- **Kling:** the photoreal-PEOPLE lane — convincing acting, dialogue, lip-sync. It breaks on close-up hands, fine fluids, and crowds beyond ~5 faces: route those beats to Seedance or reframe.\n- **Faces on Seedance:** photoreal humans trigger the face-tier routing (AI face vs consented real face — see slates-prompting-seedance §Faces). Set the face flags honestly; never skip them to save credits.\n\n## Anime\n\n- **NB2:** open with the medium — \"A hand-drawn 2D anime cel illustration of…\" — then normal narrative Subject/Setting/Action. Clean line art, flat-shaded color, expressive eyes. NB2 has no negative prompt: phrase exclusions positively (\"flat cel shading with uniform focus\", not \"no depth of field\").\n- **Seedance:** visual-style slot = \"2D anime style, clean line art, flat cel shading\". The slow/coherent-motion preference still applies — burst sakuga actions are the same instability trap as in photoreal.\n- **Kling:** weakest anime lane (its strength is live-action-like acting); expect style drift on long prose-only shots. Prefer ground rule 4: NB2 anime start-frame → i2v with a motion-only prompt. *(hypothesis: refs hold Kling's anime better than prose — verify before promising.)*\n- Anime faces drift under multiple references faster than photoreal — the named-entity two-sheet doctrine applies unchanged.\n\n## Painterly\n\n- **NB2:** medium + technique in the style framing: \"digital concept-art painting, visible brushwork, painted edges\". At most ONE school/era register (\"classic gouache illustration\") — a register, not an artist-name pile.\n- **Video:** the least-supported style lane. Use ground rule 4 (painterly NB2 frame → i2v, motion-only prompt) and expect some cleanup of painterliness over the clip *(hypothesis — set user expectations, don't promise a perfectly painterly clip)*.\n- Camera language still applies — painterly ≠ static; \"slow push-in\" works the same.\n\n## 3D render\n\n- **NB2:** name the lineage register in the style framing: \"stylized 3D render, soft global illumination, subsurface skin\". Lighting vocabulary (GI, rim light) is unusually load-bearing for the 3D read.\n- **Seedance:** the physics/effects lane flatters 3D content — visual-style slot \"stylized 3D animation\", image-quality slot \"clean render, high detail\".\n- **Kling:** same start-frame preference as anime.\n- *(hypothesis)* An engine token (\"Unreal Engine 5 render\") may help NB2; if used, ONE token, style slot only — never on Seedance where spec-stuffing hurts.\n\n## Routing recipe (what to actually do)\n\n1. Style reference available → attach it, rely on inherit. Done.\n2. No reference, image request → styled NB2 prose per the section above.\n3. No reference, video request → NB2 styled start-frame first, then i2v with motion-only prompt. Direct styled text-to-video is the fallback when a start frame doesn't fit (e.g. dialogue-first Kling shots).\n4. Multi-shot run → byte-identical style clause per shot + shared references.\n",
-    "slates-ugc-influencer-ad": "---\r\nname: slates-ugc-influencer-ad\r\ndescription: Build a UGC-style AI-influencer ad in Slates — one synthetic person, phone-grade footage, one shot per generation, cut to a vertical direct-response ad. Read before making any talking-to-camera creator video, AI spokesperson, faceless-brand UGC ad, TikTok/Reels/Shorts ad with a person in it, or a testimonial-style spot. NOT for polished cinematic spots — that is a different register with an inverted rulebook, and mixing the two ruins both.\r\n---\r\n\r\n# UGC influencer ad — the full production doctrine\r\n\r\nOne synthetic person, talking to a phone, selling one thing. This is the highest-converting short-form ad shape on paid social right now, and almost every rule below exists because the obvious approach fails.\r\n\r\n**Read this file end to end before generating anything.** The phases are not independent: the plate rules decide whether the video rolls land, and the pacing arithmetic decides how long the script can be.\r\n\r\n---\r\n\r\n## 🚨 THE GOVERNING LAW: UGLY ON PURPOSE\r\n\r\n**Handheld, blown highlights, flat light, no colour grade, no music bed.** That is not a shortcut, it is the argument. The ad's implicit claim is *this is a real person filming herself*, and cinematic polish is exactly what makes AI legible as AI.\r\n\r\n**If a clip looks like a commercial, it failed.** Every competitor reaches for cinematic. The winners reach for boring.\r\n\r\n**Photoreal is a flaw budget, not a quality setting.** Everything that makes a frame read as real is something the model must be talked out of: even skin, symmetric light, a clean silhouette, a posed smile. Spend the prompt on imperfections. **Never spend it on quality words** — \"hyperrealistic\", \"8k\", \"masterpiece\" and \"ultra-detailed\" do nothing here and push toward the rendered look you are avoiding.\r\n\r\n---\r\n\r\n## The pipeline — nothing skips a stage\r\n\r\n```\r\nscript  →  character sheet  →  beat sheet  →  plates (one per beat)  →  video (one per beat)  →  edit\r\n         [ slates_generate_image ]            [ slates_generate_image ]  [ slates_generate_video ]\r\n```\r\n\r\nEvery video roll starts from an approved plate. Every plate starts from a beat. Every beat starts from a finished script. **A video roll with no plate behind it is the most expensive mistake available in this pipeline** — images are cheap and video is not.\r\n\r\nEstimate before every stage with `slates_estimate_generation_cost`, and read `slates-cost-discipline` before the first generation.\r\n\r\n---\r\n\r\n## PHASE 1 — THE SCRIPT\r\n\r\n**Start from a proven ad, not a blank page.** Find an ad in this format that is currently running, transcribe it, and put its script beside yours. Diagnose why theirs works before writing a word. **Clone the structure; swap the product and the specifics.** A paraphrase of a working script performs worse than the script.\r\n\r\n### Write the argument as one unbroken paragraph\r\n\r\nNo beats attached, no shot list, no timings. Get it to where **no sentence can be moved without breaking the one after it.** If two sentences can swap places, it is a list, and a list reads staccato on camera.\r\n\r\n### The four tests\r\n\r\n1. **Can any two sentences swap?** Then it is a list, not an argument.\r\n2. **Does the voiceover describe what is on screen?** It must not. **The picture proves the realism; the words argue the business case.** Two channels, two jobs.\r\n3. **Does one named person recognise themselves in the first sentence?** *\"Freelancers who haven't offered video to clients yet\"* works. *\"You're scared of AI\"* names nobody.\r\n4. **Does the close force a binary?** A question where both answers belong to the viewer and one of them is unacceptable.\r\n\r\n### Line-level rules, each earned by a failure\r\n\r\n- **Clarity beats cleverness, every time.** Write it so a ten-year-old follows it. A line the viewer has to decode is not a hook.\r\n- **Use the plain word.** \"Subscription\", not \"renting access\". \"Buy it once\", not \"perpetual license\".\r\n- **Never sell a feature — sell the end state.** Nobody cares what the tool does. They care about money kept, clients won, competitors beaten.\r\n- **Give the claim its context.** \"No subscription\" means nothing on its own — a subscription to *what*? Say what is for sale before saying why it is better.\r\n- **Fear of missing out must name the loss.** \"Your competitors are using this\" is nothing. Name what it costs them: the customers being taken, the winning ad someone else found first.\r\n- **No invented numbers.** A statistic you cannot source is a liability, and the model will happily produce one.\r\n- **No contrastive negation** — *\"this isn't X, it's Y\"* / *\"Not X. Y.\"* It is the single most recognisable AI-writing tell.\r\n\r\n### Runtime is arithmetic\r\n\r\n**Words ÷ 3.4 = seconds.** Budget the fast number. A 17-word line is a 5-second shot; a 24-word line is 7 seconds. Decide length by cutting words, never by asking for a slower read.\r\n\r\n⚠️ **A generous window manufactures a slow delivery.** Give a 5-word line a 4-second shot and the model stretches the words to fill it. That is what \"she talks too slowly\" actually is, and no adjective in the prompt fixes it.\r\n\r\n⚠️ **The model normalises non-standard grammar and you cannot argue with it.** \"One less client\" comes back as \"one fewer client\" every roll. Rewrite the line so the construction never appears rather than re-rolling against it.\r\n\r\n---\r\n\r\n## PHASE 2 — THE CHARACTER\r\n\r\n**One person carries the entire ad.** Build her once, lock her, and every plate afterwards references her.\r\n\r\n### Build order, and it is counter-intuitive\r\n\r\n1. **Generate one ordinary plate first** — any location, described positively: bone structure, colouring, eye spacing, lip shape, hair texture, build.\r\n2. **Crop the face out of that plate.** That crop is now the seed.\r\n3. **Generate the identity sheet from the crop.** Sheets seeded from a rendered image come back rendered; sheets seeded from a photographic plate come back photographic. **Realism carries through the reference chain.**\r\n\r\n### The sheet itself\r\n\r\nThree panels on one plate, 16:9: a large chest-up portrait at three-quarter angle on the left, a full-body A-pose centre, a full-body back view right with hair visible. The portrait is the largest panel and carries the face at maximum detail. **No second rendering of the face anywhere on the sheet.** Identical wardrobe and hair across all three panels. Plain neutral-grey background.\r\n\r\n⚠️ **Phrase a cropped panel as FRAMING, never as absence.** *\"The head not shown\"* is a content-policy refusal. *\"Cropped at the collarbone\"* states a camera fact and passes.\r\n\r\n⚠️ **Light the sheet with flat phone light and a raking side source, never studio light.** The sheet drives lighting on every plate built from it, so a studio-lit sheet contaminates the whole ad with a studio look.\r\n\r\n### Making skin read as photographic\r\n\r\n**Name the photographic source. Do not inventory flaws.** A list — *\"open pores, oily sheen, a blemish\"* — gets treated as tokens and produces plastic. This works:\r\n\r\n> These are photographs off someone's camera roll that were never opened in an editor. Daylight arrives from one side and rakes across her face, so the texture catches the light along the cheekbone and the bridge of the nose while the other side sits in soft shadow. Slight JPEG softness, visible sensor grain in the shadows, white balance left a little cool and uncorrected.\r\n\r\n**Flat light hides texture. Raking side light is what makes pores visible at all.**\r\n\r\n**Give her a name**, in the prompt and in the script. A name creates a person; \"an AI influencer\" creates nothing.\r\n\r\n### 🚨 References decide things prompts cannot\r\n\r\n| The REFERENCE decides | The PROMPT decides |\r\n|---|---|\r\n| face, body, build, skin tone, framing, crop | skin detail, light quality, expression, what the camera is, what else is in frame |\r\n\r\n**Text cannot override the left column.** Asking for fair skin and an ordinary build against a tanned, model-built reference returns tanned and model-built. **Do not argue with a reference — replace it.**\r\n\r\n⚠️ **Never drop the reference to win the left column.** Running the same beat text-only returns a stranger and breaks the one thing this format cannot lose.\r\n\r\n---\r\n\r\n## PHASE 3 — THE BEATS\r\n\r\nCut the finished script into scenes. **Cut mid-clause, never on a full stop.** A line that resolves lets the generator sit in one room; an unresolved clause forces the location change.\r\n\r\n**Each shot gets one complete phrase — roughly 12 to 18 words.** Cut the script into phrases first, then assign a scene to each phrase. Never cut a phrase to fit a scene.\r\n\r\n### 🚨 THE CAMERA IS THE VARIETY AXIS, NOT THE LOCATION\r\n\r\nEight plates in eight different rooms, all carrying *\"shot on a phone at arm's length\"*, produce eight shots that look identical. Rewrite the **camera position** per shot and the same eight locations become a real ad.\r\n\r\n**Six generators — assign a different one to each consecutive beat:**\r\n\r\n1. **Inside an object, looking out** — washing machine drum, fridge, microwave, locker, glovebox, mailbox, vending machine, oven. The object's mouth rings the frame out of focus.\r\n2. **On the floor, looking up** — sidewalk at ankle height, base of a pillar, stairwell landing, under a bench.\r\n3. **High, looking down** — wedged on a shop shelf, on top of a fuel pump, a ceiling corner, one level above.\r\n4. **Set on a surface she walks into frame of** — car roof, folding table, bus bench, tailgate.\r\n5. **Behind or through something** — a hot food counter, chain-link, a clothes rack, a rain-covered window.\r\n6. **Reflections that are not a bathroom mirror** — polished elevator steel, a dark shop window at night, a wing mirror, a puddle, a dead ATM screen.\r\n\r\n⚠️ **UGC does not mean selfie.** An arm's-length clause in a shared prompt tail produces one camera position N times *and* puts her extended arm through the corner of every frame. **The phone is propped, set down, wedged or left somewhere far more often than it is held.**\r\n\r\n⚠️ **Rooms of a house are a cop-out.** A place with a job beats a place to live: laundromat, parking garage, corner store, gas station, car wash, storage corridor, elevator, stairwell, loading dock.\r\n\r\n⚠️ **Changing location costs nothing.** Identity travels on **wardrobe + a fixed hair style + one accessory motif**, never on the room. Never let set continuity argue you into staying put.\r\n\r\n⚠️ **Fast cutting is close-weighted.** A wide shot needs time for the eye to find the subject; a close-up lands instantly. At 1.5–2.5 seconds a beat, weight the set toward tight.\r\n\r\n---\r\n\r\n## PHASE 4 — THE PLATES. One per beat.\r\n\r\n`slates_generate_image`, model `gpt-image-2`, `quality: high`, vertical `9:16`. **Fire concurrently — 8 slots run at once; serial calls waste the day.** Pass `projectId` so every plate lands in the gallery where you can see it.\r\n\r\n**GPT Image 2 is the rail for photoreal people.** Head-to-head against the alternatives it is the one that returns photographic rather than plastic skin. Routing detail: `slates-model-selection`.\r\n\r\n### 🚨 GPT Image 2 routes references through its EDIT endpoint\r\n\r\nTwo consequences, both of which waste a generation:\r\n\r\n1. **A prompt that opens by describing the reference gets the reference edited.** Open with *\"one reference image is a three-panel character identity sheet\"* and it returns a brand-new character identity sheet. The head block below is written the way it is precisely to prevent this.\r\n2. **One reference alone drags the scene into that reference's background.** A sheet-only roll puts her on the sheet's plain grey backdrop with no location at all. **Always pass two references:** the identity sheet, plus a second image supplying either the environment or the skin quality to match.\r\n\r\n### 🚨 THE PLATE IS FRAME ZERO\r\n\r\nThe plate is the video roll's first frame, so **whatever is going to happen has not happened yet.** If the beat is something breaking, the thing is intact in the plate. A plate showing the aftermath gives the model nothing to do and opens the shot on a result instead of an event.\r\n\r\n**Sole exception:** a background element that must be visible in frame one to stop the scroll.\r\n\r\n### Build every plate from three blocks\r\n\r\n**HEAD — identical every time. The anti-sheet clause is load-bearing.**\r\n\r\n> Reference image 1 is a character identity sheet showing one woman across several panels — the face in the large portrait panel is the authority for her identity. Use that exact woman. Reference image 2 is [the second reference and its job]. Produce a single photograph, not a sheet and not multiple panels.\r\n\r\n**MIDDLE — the only part that changes:** location, camera position, wardrobe, **expression, and hair state.**\r\n\r\n**TAIL — identical every time.**\r\n\r\n> Shot on a phone, slightly soft, mild wide-lens distortion, no colour grade. Camera-original skin with open pores across the nose and cheeks, an oily sheen on the forehead, faint creases under the eyes. These are photographs off someone's camera roll that were never opened in an editor. No retouching and no skin smoothing. Nobody is holding a camera and no phone, camera or device is visible anywhere in the picture. Clean unmarked frame with no text and no logo.\r\n\r\n### 🚨 Expression and hair state go in the MIDDLE, never the tail\r\n\r\nPutting *\"caught mid-sentence, mouth slightly open, eyes on the lens\"* in the shared tail produces the identical face in every location.\r\n\r\n**Hair STYLE stays fixed** — that plus wardrobe is what carries identity. **Hair STATE varies:** tucked behind an ear, blown across her face, tangled from lying down, scraped back.\r\n\r\nExpressions that worked: mid-sentence and locked on the lens · laughing with eyes half shut · concentrating on something, not the camera · squinting into daylight · flat and unimpressed, waiting · bored and blinking, about to speak and not.\r\n\r\n### 🚨 The plate decides whether a physical feat is possible at all\r\n\r\n**The model will mime an action it cannot resolve.** Three rolls failed the same way on 2026-08-24: a dumbbell whose bar stayed straight while her hands moved, a pan that floated free of her grip and folded against her chest, a steel pole that leaned instead of folding.\r\n\r\n**All three were plate problems.** The pan plate had her holding it one-handed at arm's length with the phone in her other hand — no second hand existed to fold it with, so the model invented one and the object drifted.\r\n\r\n- **If the act needs both hands, the plate must show both hands free.** That means the camera is propped, set down or wedged — never held at arm's length.\r\n- **Direct the resistance, not the result.** Knuckles whitening, tendons standing, a beat where it holds, then the give.\r\n- **State the finished shape.** *\"Folded in half so the two ends touch\"* is checkable. *\"Bends it\"* permits a lean.\r\n\r\n### 🚨 NEVER composite a character onto a finished scene. Generate the whole frame at once.\r\n\r\n**Receipt, 2026-08-24.** A spectacle plate was generated first (a street, correctly lit), then the character was added by passing that plate as a second reference. Every attempt came back looking green-screened: her face carried a soft frontal key that existed nowhere in the scene, she was sharper and cooler than the background, and no haze or flare crossed in front of her. **A finished scene handed in as a reference tells the model to paste, not to light.**\r\n\r\n**The fix is to describe one scene from scratch with only the identity reference attached.** Same subject, same location, same everything — the difference is that the model renders one frame in one pass instead of merging two. The first from-scratch attempt landed.\r\n\r\nUse a second reference for *design* (a creature's build, a style) — never for a whole composed frame the character has to be inserted into.\r\n\r\n### 🚨 Write the light as ONE physical system, or the character reads as pasted in\r\n\r\nThe single most common tell is a key light on the subject's face that has no source in the picture. Name all four of these explicitly:\r\n\r\n1. **One source, and where it is.** *\"The sun is low and directly down the avenue behind everything.\"*\r\n2. **What that does to HER specifically.** If she is backlit, say her face is in ambient bounce with no highlight and no catchlight, and put the hard rim on her hair, one shoulder and the collar.\r\n3. **Forbid the key that does not exist.** *\"There is no light in front of her: no frontal key, no fill, no soft source on her face.\"* Without this clause the model adds one every time.\r\n4. **Atmosphere in FRONT of her, not only behind.** Haze, dust and veiling flare must cross the foreground, so her edges are no sharper than the parked cars beside her. **Every shadow in frame runs parallel**, hers included.\r\n\r\n**Let the camera's exposure be part of the physics.** Backlit means the phone lifts for her face and the sky blows to white. Night means the fire blows out and the blacks lift. That single decision ties subject and background into one image better than any amount of description.\r\n\r\n⚠️ **Do not stack grain adjectives.** *\"Heavy luminance noise\"* plus *\"aggressive noise reduction\"* plus *\"milky lifted blacks\"* in one prompt produced a clip so noisy it read as broken. Modern phones are clean at night — name the light, not the sensor damage.\r\n\r\n### Other plate rules\r\n\r\n- **Never describe the phone as an object.** *\"She lifts the phone\"* renders a phone in her hand. Describe what the camera does.\r\n- **Write negatives as positives.** *Matte uneven skin*, not \"not smooth\". *Flat daylight*, not \"no studio lighting\". *Caught mid-step*, not \"not posed\". A \"no X\" list is either ignored or quietly summons X.\r\n- **Text on props renders cleanly and is worth using.** A label reading `YOUR PRODUCT` comes back legible.\r\n- **Content refusals are stochastic. Re-roll the identical prompt before rewriting it.** A plate can refuse once and pass on an unchanged second call. A refused job returns no image and costs nothing. Only redesign after two or three refusals.\r\n- **Generate a pool and pick.** A mediocre plate wastes a metered video roll, and video is the only expensive stage.\r\n\r\n---\r\n\r\n## PHASE 5 — THE VIDEO. One generation per shot.\r\n\r\n`slates_generate_video`, vertical `9:16`, the plate as the first frame. Model routing lives in `slates-model-selection` — read it rather than guessing, and check `slates_estimate_generation_cost` before firing.\r\n\r\n### 🚨 ONE SHOT PER GENERATION. Never a multi-scene take.\r\n\r\nMeasured directly: a six-scene roll returned **nothing usable**, a three-scene roll returned about one usable second, and **the same beats shot as ten separate single-scene generations all landed on the first attempt.**\r\n\r\nCost is linear per second, so splitting is free. What it buys:\r\n\r\n- **A bad shot costs one shot, not the whole ad.** In a long take one broken scene kills every good scene beside it and the retry re-buys all of them.\r\n- **Every shot re-rolls independently**, so a note lands on one shot.\r\n- **8 concurrent jobs means the whole ad shoots in one wave anyway.** A long take buys no wall-clock.\r\n- **No transition can fail**, because there are no transitions inside a generation. Cuts happen in the edit, where they are free and reversible.\r\n\r\nA short line padded into the minimum duration is trimmed in the edit for nothing. **That padding is a far cheaper problem than a dead long take.**\r\n\r\n### The plate is the first frame, always\r\n\r\nPassing the plate as the first frame locks the opening on a composition you already approved, so the model animates a frame you chose instead of inventing one. Pass the identity sheet alongside it as a reference.\r\n\r\n**The one exception** is a shot whose first frame must not exist yet — a reveal, an assembly, anything resolving *into* the plate. There, pass the plate as a reference only.\r\n\r\n⚠️ **Re-read the plate before writing its prompt.** A prompt describing a camera resting on a counter, against a plate that is plainly a held selfie, is a contradiction the model resolves badly.\r\n\r\n### 🚨 Eyeline decides whether a beat lip-syncs or becomes voiceover\r\n\r\n**A character not addressing the lens reads as B-roll, and the model lays the voice over silent footage.** This is not a frame-size problem — the identical shot lip-syncs correctly once the eyeline is locked.\r\n\r\n**Write where her eyes ARE for the whole duration:** *\"her eyes are locked on the lens for the entire shot and she talks straight into it.\"*\r\n\r\n⚠️ **A brief exception becomes the whole shot.** *\"Glances at her feet once and back up\"* produces five seconds of staring at her feet. *\"Never looks down at her hand\"* produces her looking down at her hand. The model takes the vivid clause and runs the beat on it. State the eyeline as the dominant instruction and add exceptions only if you must.\r\n\r\nMixed voiceover and on-camera is a legitimate choice in this format. It has to be chosen, not discovered.\r\n\r\n### 🚨 Name the rig, THEN the operator's body\r\n\r\n**\"Single continuous handheld shot\" tells the model what the camera IS. The operator's body tells it what the camera DOES.** Ship one without the other and it renders its default: a smooth slow push with almost no shake.\r\n\r\n**Write a person's body, never a camera path.** Not *\"the camera drifts along the tailgate\"* but *\"the phone is lying flat on the tailgate metal, held loosely by a hand at the corner, so the frame sags to one side and gets nudged back level once.\"* Arrives, sags, overshoots, corrects, gets bumped, catches up. **And the world must register it** — a move nothing reacts to is a move nobody made.\r\n\r\n**An unusual camera position is still handheld.** Inside a fridge, on a car roof, in the gutter — someone is holding the phone there. A locked-off version reads as a tripod and loses the whole register.\r\n\r\n### Micro-movement is directed, never assumed\r\n\r\nEvery beat carries an explicit list of small involuntary actions. A beat that only says what she says produces a mannequin talking.\r\n\r\n> Small things: she blinks twice, her weight shifts onto one hip, a strand of hair slips forward over her shoulder, and the framing drifts and corrects the way a held phone does.\r\n\r\nOthers that work: a container tipping as she takes its weight · her shoulder rising as she reaches · her hand sliding along a handrail · picking dry skin off a lip · blowing hair out of the corner of her mouth.\r\n\r\n🚨 **Small involuntary business only — never a big physical action at the lens.** Directing her to climb onto something threw her whole body across a close frame and wrecked it. Re-rolled with nothing but a hair flip, a weight shift and two fingers tapping, it landed. **Small business reads as captured; a big action reads as staged, and at close framing it also destroys the composition.**\r\n\r\n⚠️ **When a re-roll must hold the plate's pose, say so as a prohibition:** *\"She does not climb onto the truck, she does not sit down, and she does not lean any further toward the camera than she already is.\"* Pair it with *\"the camera never travels and nothing in the shot jolts it.\"*\r\n\r\n**For a physical feat, direct the struggle, not the event.** Something that gives way on contact reads as a trick; something that **resists for a beat and holds** before it gives reads as real — knuckles whitening, tendons standing out, the forearm tensing, the grip shaking.\r\n\r\n### Audio\r\n\r\n**Always on.** \"No music\" means no score; it never meant no audio. Say what you want present — her lip-synced dialogue, the sound the action makes, footsteps, room tone — then close with **\"No music. No score. No background track of any kind.\"**\r\n\r\nGenerating with audio disabled produces a mouth moving in silence and the roll is wasted.\r\n\r\n---\r\n\r\n## PHASE 6 — THE EDIT\r\n\r\nFree, local, and where the ad is actually made. Trim dead air off a beat that ran slow. Drop a re-rolled beat over the original. Burn captions — all-caps, centred low, one or two words at a time, cut on the stressed syllable.\r\n\r\n**A one-second overrun is an edit problem, not a re-roll.** Re-rolling a shot to fix a second of air is the expensive way round.\r\n\r\n### Verify by WATCHING, never by exit code or by sampling stills\r\n\r\nA job that exits cleanly can still be a refusal with no file, and a landed take can still have a beat that never cut. **Check the file has an audio stream and the right dimensions, then look at a frame from each beat.**\r\n\r\n---\r\n\r\n## Anti-patterns\r\n\r\n- **Don't** write cinematic. Polish is the tell.\r\n- **Don't** put quality words in a prompt. Spend it on imperfections.\r\n- **Don't** combine scenes into one generation to save money. It costs more and returns less.\r\n- **Don't** generate video without a plate behind it.\r\n- **Don't** generate text overlays in an image. Captions happen in the edit.\r\n- **Don't** rewrite a prompt after one content refusal. Re-roll it unchanged first.\r\n- **Don't** describe the reference at the top of a prompt. Describe the photograph you want.\r\n- **Don't** fire the next roll because the last one failed. Diagnose against this file first.\r\n",
-    "slates-vision-feedback-loop": "---\nname: slates-vision-feedback-loop\ndescription: Lower-level utility skill for any Slates workflow that needs to \"generate, look at the result, refine, regenerate.\" Defines the standard inline-vision pattern. Other Slates skills compose this. Use when generating images and you need to confirm they match the brief before moving on, or when the user asks to \"iterate\" on an image.\n---\n\n# Vision feedback loop — Slates utility skill\n\nSlates returns generated images inline as base64. You see the actual pixels. Use that — don't trust prompt-following blindly.\n\n## Asset codes are your shared vocabulary with the user\n\nEvery asset in Slates has a short stable code (e.g. `IMG-A12`, `VID-V3`, `AUD-S1`) and a label derived from its prompt (e.g. `Beach Sunset`). These are visible in the gallery as a corner badge on each thumbnail. **Always refer to assets by their code in chat** so the user can match what you're saying to a specific card in their gallery.\n\n- ✅ \"I'm using **IMG-A12 — Beach Sunset** as the first frame. The second-frame candidate **IMG-A15** has the right composition but warmer light — want me to use that one instead?\"\n- ❌ \"I'm using the beach sunset image...\" (user has four beach sunset variants — which one?)\n- ❌ \"I'm using asset `7a3f9e4b-...`\" (UUIDs aren't readable; user can't match to a badge)\n\nThe code is the FORMAL reference. The label is human texture. Use both: `IMG-A12 — Beach Sunset`.\n\n## Vision tools at your disposal\n\n- `slates_get_asset_image` — pull one image into context. Returns its code+label.\n- `slates_get_assets_batch` — pull up to 8 images in one call. Use when picking from a candidate set; cheaper than N individual fetches.\n- `slates_get_asset_video_frames` — extract N keyframes (default 3) from a video and inline them as JPEGs. You can't see video natively; this is how you \"look at\" a clip before refining its motion prompt.\n\n## Pre-flight is automatic on the gen tools\n\n`slates_generate_video`, `slates_generate_motion_transfer`, and `slates_generate_lip_sync` now show you their reference assets **inline** on the confirm response. You don't need to fetch them yourself — but you DO need to look at what comes back, revise the prompt if the references suggest a different motion/framing, and only then re-call with `confirm=true`.\n\n## 🔴 The still-gate — never animate a bad frame\n\n<!-- @inject:still-gate -->\n**A visible defect in the still is already a STOP.** Do not animate it. Fix the frame first, then move to motion — and go to motion only when the crop passes the still scan and you genuinely need movement to confirm an uncertain edge, reflection, or object.\n\nThis is a **cost** rule as much as a craft rule: a 1080p/10s premium video generation costs many multiples of an image re-roll, and video is where a defect stops being fixable. Anything wrong in the still gets worse in motion — soft geometry mushes, broken-but-plausible objects fall apart, oily textures start crawling. **Animating a known-bad frame is the single most expensive mistake in the pipeline.** Re-rolling the image is the cheap move; re-rolling the video is not.\n<!-- @end:still-gate -->\n\n## The pattern\n\n1. **Generate.** Call `slates_generate_image` with a prompt. The result is in your context as an image content block.\n2. **Evaluate on TWO axes — they are different questions:**\n   - **Brief-conformance** — what did the user actually want? Are the elements right? Composition? Lighting? Subject identity?\n   - **Defects** — run the slop rubric below. *A frame can match the brief perfectly and still be slop that mushes the moment it moves.* Checking only the first axis is how a bad frame reaches an expensive video call.\n3. **One of three outcomes:**\n   - **Right** → save it (bind to a frame, character slot, etc.) and move on.\n   - **Close, but adjustable** → refine with a specific delta, regenerate **once**.\n   - **Wrong direction** → ask the user before regenerating. Don't burn credits on prompt-thrashing.\n\n## The defect rubric — four slop tells\n\n| Tell | What it looks like | Why it matters downstream |\n|---|---|---|\n| **Light with no transitions** | Flat-black pits instead of a shadow ramp; light that stops rather than falls off | Transfers onto every character or object added into that plate later |\n| **Broken-but-plausible objects** | Crates, railings, hardware, mechanisms you can *almost* read but that don't resolve | Turn to mush in motion, and the model multiplies them |\n| **Local logic breaks** | An effect present in only part of the frame — rain scratching one corner, wet ground under one figure | The video model's physical logic breaks along with it |\n| **Oily textures** | Soapy, licked-smooth surfaces that have lost their material identity | Reflections crawl in motion; the plate can't hold continuity |\n\n### Per-model accents — check the one you actually used\n\n- **Nano Banana Pro** (`nano-banana-pro`) — ruler-straight symmetry, everything parallel and square, flat even light, pretty but staged/stock, textures reading as 3D render rather than photograph. **It hyperbolizes every edit**: ask for graffiti on one wall and the whole location gets tagged.\n- **GPT Image 2** (`gpt-image-2`) — microcontrast to the ceiling, hard halos on every edge, no depth or bokeh, white balance pulled warm until the frame yellows, plastic licked-smooth materials. Worst tell: **one sickly texture pattern laid over the entire frame**.\n\n> ⚠️ These are accents for **`nano-banana-pro`** and **`gpt-image-2`** specifically. `nano-banana-2` is a **different model** (Gemini 3.1 Flash Image vs NB Pro's Gemini 3 Pro Image) and we have **no evidence** about its accent. Do not inherit one — say nothing rather than warn about a failure mode you can't substantiate.\n\n## Where the fault lives — triage before you change anything\n\nWe say \"one specific delta per regeneration\" but that only helps once you know *which* variable to move. Diagnose first:\n\n| Visible pattern | Diagnosis | Fix |\n|---|---|---|\n| The defect exists in the source asset, or stays tied to the same feature when the direction changes | **Source asset** | Fix the sheet / plate, not the prompt |\n| Source is clean, and the defect changes when only the suspect motion clause changes | **Motion direction** | Fix the prompt |\n| Controls conflict, or the failure follows neither variable | **Inconclusive** | Narrow the test — change less, not more |\n\n**Review routes; it is not pass/fail.** Geography melts → fix the location. Identity drifts → fix the character sheet. Assets are sound but the action is wrong → fix the video direction. Wrong idea entirely → reopen the brief with the user.\n\n**Correct the earliest broken handoff.** Polishing a downstream symptom hides the source and guarantees it resurfaces in the next shot built from the same asset.\n\n## Baseline hygiene — isolate the variable you're testing\n\nWhen the **character** is the question, keep the location out of it: test on a plate that already holds its own geometry, depth, materials, and light. **A broken plate gives every character failure a second plausible cause**, and you will spend re-rolls deciding which one you're looking at. The same applies in reverse — test a plate empty before you populate it.\n\n## Refinement rules\n\n- **One specific delta per regeneration.** Don't change five things at once — you won't know what helped.\n- **Rewrite the FULL prompt on every iteration — never a diff, never a fragment.** Change one decision, then re-emit the whole prompt so every slot still agrees with every other slot. This composes with the rule above rather than replacing it: *one delta* governs **what changes**, *full rewrite* governs **how you re-emit it**. A patched fragment leaves the old slots stale and silently contradicting the new one.\n  - On **Seedance**, a re-emit must keep the `Shot N` structure intact — see `slates-prompting-seedance`.\n  - **Exception — Omni Flash Edit.** Long prompts documentedly destroy its fidelity. There the rule inverts: one short instruction plus *\"Keep everything else the same.\"*\n- **Anchor with references.** If the result drifted from the user's intent, attach the *previous best* generation as a reference image alongside the original brief.\n- **Use `slates_get_asset_image`** to pull a previously-generated image back into context if you need to compare against a fresh generation.\n- **Use `slates_edit_image`** for surgical tweaks instead of full regeneration when ~90% of the image is right — `sourceAssetId` = the asset, `prompt` = the change only. Edits preserve composition and identity; full regen rolls the dice. Recipe: `slates-edit-and-iterate`.\n\n## Cost discipline\n\n- Track total credits spent across the loop. Surface to the user every 3 iterations.\n- Stop after 3 failed iterations on the same prompt — escalate to the user with what you tried and what's not working. The slot machine never converges.\n- For a high-cost generation — anything past the confirm gate in `slates-cost-discipline` — confirm before *every* attempt, not just the first.\n\n## When to break the loop\n\n- The user said \"good enough\" or \"ship it.\" Stop iterating.\n- You've burned >5 generations on one frame. Hand back and ask.\n- The user changes brief mid-loop. Treat it as a new brief, not a continuation.\n\n## Voice when narrating to the user\n\nTight, observational, no editorializing.\n- ✅ \"Frame 2 has the wrong lighting direction — back-lit instead of side. Regenerating with side light.\"\n- ❌ \"I notice that the lighting in frame 2 isn't quite what we were going for. I'll go ahead and try again with a different approach.\"\n",
+    "slates-ugc-influencer-ad": "---\r\nname: slates-ugc-influencer-ad\r\ndescription: Build a UGC-style AI-influencer ad in Slates — one synthetic person, phone-grade footage, one shot per generation, cut to a vertical direct-response ad. Read before making any talking-to-camera creator video, AI spokesperson, faceless-brand UGC ad, TikTok/Reels/Shorts ad with a person in it, or a testimonial-style spot. NOT for polished cinematic spots — that is a different register with an inverted rulebook, and mixing the two ruins both.\r\n---\r\n\r\n# UGC influencer ad — the full production doctrine\r\n\r\nOne synthetic person, talking to a phone, selling one thing. This is the highest-converting short-form ad shape on paid social right now, and almost every rule below exists because the obvious approach fails.\r\n\r\n**Read this file end to end before generating anything.** The phases are not independent: the plate rules decide whether the video rolls land, and the pacing arithmetic decides how long the script can be.\r\n\r\n---\r\n\r\n## 🚨 THE GOVERNING LAW: UGLY ON PURPOSE\r\n\r\n**Handheld, blown highlights, flat light, no colour grade, no music bed.** That is not a shortcut, it is the argument. The ad's implicit claim is *this is a real person filming herself*, and cinematic polish is exactly what makes AI legible as AI.\r\n\r\n**If a clip looks like a commercial, it failed.** Every competitor reaches for cinematic. The winners reach for boring.\r\n\r\n**Photoreal is a flaw budget, not a quality setting.** Everything that makes a frame read as real is something the model must be talked out of: even skin, symmetric light, a clean silhouette, a posed smile. Spend the prompt on imperfections. **Never spend it on quality words** — \"hyperrealistic\", \"8k\", \"masterpiece\" and \"ultra-detailed\" do nothing here and push toward the rendered look you are avoiding.\r\n\r\n---\r\n\r\n## The pipeline — nothing skips a stage\r\n\r\n```\r\nscript  →  character sheet  →  beat sheet  →  plates (one per beat)  →  video (one per beat)  →  edit\r\n         [ slates_generate_image ]            [ slates_generate_image ]  [ slates_generate_video ]\r\n```\r\n\r\nEvery video roll starts from an approved plate. Every plate starts from a beat. Every beat starts from a finished script. **A video roll with no plate behind it is the most expensive mistake available in this pipeline** — images are cheap and video is not.\r\n\r\nEstimate before every stage with `slates_estimate_generation_cost`, and read `slates-cost-discipline` before the first generation.\r\n\r\n---\r\n\r\n## PHASE 1 — THE SCRIPT\r\n\r\n**Start from a proven ad, not a blank page.** Find an ad in this format that is currently running, transcribe it, and put its script beside yours. Diagnose why theirs works before writing a word. **Clone the structure; swap the product and the specifics.** A paraphrase of a working script performs worse than the script.\r\n\r\n### Write the argument as one unbroken paragraph\r\n\r\nNo beats attached, no shot list, no timings. Get it to where **no sentence can be moved without breaking the one after it.** If two sentences can swap places, it is a list, and a list reads staccato on camera.\r\n\r\n### The four tests\r\n\r\n1. **Can any two sentences swap?** Then it is a list, not an argument.\r\n2. **Does the voiceover describe what is on screen?** It must not. **The picture proves the realism; the words argue the business case.** Two channels, two jobs.\r\n3. **Does one named person recognise themselves in the first sentence?** *\"Freelancers who haven't offered video to clients yet\"* works. *\"You're scared of AI\"* names nobody.\r\n4. **Does the close force a binary?** A question where both answers belong to the viewer and one of them is unacceptable.\r\n\r\n### Line-level rules, each earned by a failure\r\n\r\n- **Clarity beats cleverness, every time.** Write it so a ten-year-old follows it. A line the viewer has to decode is not a hook.\r\n- **Use the plain word.** \"Subscription\", not \"renting access\". \"Buy it once\", not \"perpetual license\".\r\n- **Never sell a feature — sell the end state.** Nobody cares what the tool does. They care about money kept, clients won, competitors beaten.\r\n- **Give the claim its context.** \"No subscription\" means nothing on its own — a subscription to *what*? Say what is for sale before saying why it is better.\r\n- **Fear of missing out must name the loss.** \"Your competitors are using this\" is nothing. Name what it costs them: the customers being taken, the winning ad someone else found first.\r\n- **No invented numbers.** A statistic you cannot source is a liability, and the model will happily produce one.\r\n- **No contrastive negation** — *\"this isn't X, it's Y\"* / *\"Not X. Y.\"* It is the single most recognisable AI-writing tell.\r\n\r\n### Runtime is arithmetic\r\n\r\n**Words ÷ 3.4 = seconds.** Budget the fast number. A 17-word line is a 5-second shot; a 24-word line is 7 seconds. Decide length by cutting words, never by asking for a slower read.\r\n\r\n⚠️ **A generous window manufactures a slow delivery.** Give a 5-word line a 4-second shot and the model stretches the words to fill it. That is what \"she talks too slowly\" actually is, and no adjective in the prompt fixes it.\r\n\r\n⚠️ **The model normalises non-standard grammar and you cannot argue with it.** \"One less client\" comes back as \"one fewer client\" every roll. Rewrite the line so the construction never appears rather than re-rolling against it.\r\n\r\n---\r\n\r\n## PHASE 2 — THE CHARACTER\r\n\r\n**One person carries the entire ad.** Build her once, lock her, and every plate afterwards references her.\r\n\r\n### Build order, and it is counter-intuitive\r\n\r\n1. **Generate one ordinary plate first** — any location, described positively: bone structure, colouring, eye spacing, lip shape, hair texture, build.\r\n2. **Crop the face out of that plate.** That crop is now the seed.\r\n3. **Generate the identity sheet from the crop.** Sheets seeded from a rendered image come back rendered; sheets seeded from a photographic plate come back photographic. **Realism carries through the reference chain.**\r\n\r\n### The sheet itself\r\n\r\nThree panels on one plate, 16:9: a large chest-up portrait at three-quarter angle on the left, a full-body A-pose centre, a full-body back view right with hair visible. The portrait is the largest panel and carries the face at maximum detail. **No second rendering of the face anywhere on the sheet.** Identical wardrobe and hair across all three panels. Plain neutral-grey background.\r\n\r\n⚠️ **Phrase a cropped panel as FRAMING, never as absence.** *\"The head not shown\"* is a content-policy refusal. *\"Cropped at the collarbone\"* states a camera fact and passes.\r\n\r\n⚠️ **Light the sheet with flat phone light and a raking side source, never studio light.** The sheet drives lighting on every plate built from it, so a studio-lit sheet contaminates the whole ad with a studio look.\r\n\r\n### Making skin read as photographic\r\n\r\n**Name the photographic source. Do not inventory flaws.** A list — *\"open pores, oily sheen, a blemish\"* — gets treated as tokens and produces plastic. This works:\r\n\r\n> These are photographs off someone's camera roll that were never opened in an editor. Daylight arrives from one side and rakes across her face, so the texture catches the light along the cheekbone and the bridge of the nose while the other side sits in soft shadow. Slight JPEG softness, visible sensor grain in the shadows, white balance left a little cool and uncorrected.\r\n\r\n**Flat light hides texture. Raking side light is what makes pores visible at all.**\r\n\r\n**Give her a name**, in the prompt and in the script. A name creates a person; \"an AI influencer\" creates nothing.\r\n\r\n### 🚨 References decide things prompts cannot\r\n\r\n| The REFERENCE decides | The PROMPT decides |\r\n|---|---|\r\n| face, body, build, skin tone, framing, crop | skin detail, light quality, expression, what the camera is, what else is in frame |\r\n\r\n**Text cannot override the left column.** Asking for fair skin and an ordinary build against a tanned, model-built reference returns tanned and model-built. **Do not argue with a reference — replace it.**\r\n\r\n⚠️ **Never drop the reference to win the left column.** Running the same beat text-only returns a stranger and breaks the one thing this format cannot lose.\r\n\r\n---\r\n\r\n## PHASE 3 — THE BEATS\r\n\r\nCut the finished script into scenes. **Cut mid-clause, never on a full stop.** A line that resolves lets the generator sit in one room; an unresolved clause forces the location change.\r\n\r\n**Each shot gets one complete phrase — roughly 12 to 18 words.** Cut the script into phrases first, then assign a scene to each phrase. Never cut a phrase to fit a scene.\r\n\r\n### 🚨 THE CAMERA IS THE VARIETY AXIS, NOT THE LOCATION\r\n\r\nEight plates in eight different rooms, all carrying *\"shot on a phone at arm's length\"*, produce eight shots that look identical. Rewrite the **camera position** per shot and the same eight locations become a real ad.\r\n\r\n**Six generators — assign a different one to each consecutive beat:**\r\n\r\n1. **Inside an object, looking out** — washing machine drum, fridge, microwave, locker, glovebox, mailbox, vending machine, oven. The object's mouth rings the frame out of focus.\r\n2. **On the floor, looking up** — sidewalk at ankle height, base of a pillar, stairwell landing, under a bench.\r\n3. **High, looking down** — wedged on a shop shelf, on top of a fuel pump, a ceiling corner, one level above.\r\n4. **Set on a surface she walks into frame of** — car roof, folding table, bus bench, tailgate.\r\n5. **Behind or through something** — a hot food counter, chain-link, a clothes rack, a rain-covered window.\r\n6. **Reflections that are not a bathroom mirror** — polished elevator steel, a dark shop window at night, a wing mirror, a puddle, a dead ATM screen.\r\n\r\n⚠️ **UGC does not mean selfie.** An arm's-length clause in a shared prompt tail produces one camera position N times *and* puts her extended arm through the corner of every frame. **The phone is propped, set down, wedged or left somewhere far more often than it is held.**\r\n\r\n⚠️ **Rooms of a house are a cop-out.** A place with a job beats a place to live: laundromat, parking garage, corner store, gas station, car wash, storage corridor, elevator, stairwell, loading dock.\r\n\r\n⚠️ **Changing location costs nothing.** Identity travels on **wardrobe + a fixed hair style + one accessory motif**, never on the room. Never let set continuity argue you into staying put.\r\n\r\n⚠️ **Fast cutting is close-weighted.** A wide shot needs time for the eye to find the subject; a close-up lands instantly. At 1.5–2.5 seconds a beat, weight the set toward tight.\r\n\r\n---\r\n\r\n## PHASE 4 — THE PLATES. One per beat.\r\n\r\n`slates_generate_image`, model `gpt-image-2-5-sunburst`, `quality: max`, vertical `9:16`. **Fire concurrently — 8 slots run at once; serial calls waste the day.** Pass `projectId` so every plate lands in the gallery where you can see it.\r\n\r\n**Why that seat and that rung.** `max` because the 2026-08-24 photoreal result was measured at GPT Image 2's `high`, which is `max` here — the ladder was renamed, not repriced. **Sunburst** because these plates are finals and OpenAI routes quality to Sunburst: Flare's quality is only *comparable to* GPT Image 2, while Sunburst is documented higher. Same price, so the only cost is latency, and you are firing eight in parallel anyway. Use **Flare** while you are still exploring wardrobe, room or framing, then re-fire the chosen direction on Sunburst. ⚠️ Reasoned from OpenAI's positioning, not measured — if you A/B one plate, write the answer into `slates-prompting-gpt-image-2-5`.\r\n\r\n**GPT Image is the rail for photoreal people.** Head-to-head against the alternatives it is the one that returns photographic rather than plastic skin. Routing detail: `slates-model-selection`.\r\n\r\n### 🚨 GPT Image 2.5 routes references through its EDIT endpoint\r\n\r\nTwo consequences, both of which waste a generation:\r\n\r\n1. **A prompt that opens by describing the reference gets the reference edited.** Open with *\"one reference image is a three-panel character identity sheet\"* and it returns a brand-new character identity sheet. The head block below is written the way it is precisely to prevent this.\r\n2. **One reference alone drags the scene into that reference's background.** A sheet-only roll puts her on the sheet's plain grey backdrop with no location at all. **Always pass two references:** the identity sheet, plus a second image supplying either the environment or the skin quality to match.\r\n\r\n### 🚨 THE PLATE IS FRAME ZERO\r\n\r\nThe plate is the video roll's first frame, so **whatever is going to happen has not happened yet.** If the beat is something breaking, the thing is intact in the plate. A plate showing the aftermath gives the model nothing to do and opens the shot on a result instead of an event.\r\n\r\n**Sole exception:** a background element that must be visible in frame one to stop the scroll.\r\n\r\n### Build every plate from three blocks\r\n\r\n**HEAD — identical every time. The anti-sheet clause is load-bearing.**\r\n\r\n> Reference image 1 is a character identity sheet showing one woman across several panels — the face in the large portrait panel is the authority for her identity. Use that exact woman. Reference image 2 is [the second reference and its job]. Produce a single photograph, not a sheet and not multiple panels.\r\n\r\n**MIDDLE — the only part that changes:** location, camera position, wardrobe, **expression, and hair state.**\r\n\r\n**TAIL — identical every time.**\r\n\r\n> Shot on a phone, slightly soft, mild wide-lens distortion, no colour grade. Camera-original skin with open pores across the nose and cheeks, an oily sheen on the forehead, faint creases under the eyes. These are photographs off someone's camera roll that were never opened in an editor. No retouching and no skin smoothing. Nobody is holding a camera and no phone, camera or device is visible anywhere in the picture. Clean unmarked frame with no text and no logo.\r\n\r\n### 🚨 Expression and hair state go in the MIDDLE, never the tail\r\n\r\nPutting *\"caught mid-sentence, mouth slightly open, eyes on the lens\"* in the shared tail produces the identical face in every location.\r\n\r\n**Hair STYLE stays fixed** — that plus wardrobe is what carries identity. **Hair STATE varies:** tucked behind an ear, blown across her face, tangled from lying down, scraped back.\r\n\r\nExpressions that worked: mid-sentence and locked on the lens · laughing with eyes half shut · concentrating on something, not the camera · squinting into daylight · flat and unimpressed, waiting · bored and blinking, about to speak and not.\r\n\r\n### 🚨 The plate decides whether a physical feat is possible at all\r\n\r\n**The model will mime an action it cannot resolve.** Three rolls failed the same way on 2026-08-24: a dumbbell whose bar stayed straight while her hands moved, a pan that floated free of her grip and folded against her chest, a steel pole that leaned instead of folding.\r\n\r\n**All three were plate problems.** The pan plate had her holding it one-handed at arm's length with the phone in her other hand — no second hand existed to fold it with, so the model invented one and the object drifted.\r\n\r\n- **If the act needs both hands, the plate must show both hands free.** That means the camera is propped, set down or wedged — never held at arm's length.\r\n- **Direct the resistance, not the result.** Knuckles whitening, tendons standing, a beat where it holds, then the give.\r\n- **State the finished shape.** *\"Folded in half so the two ends touch\"* is checkable. *\"Bends it\"* permits a lean.\r\n\r\n### 🚨 NEVER composite a character onto a finished scene. Generate the whole frame at once.\r\n\r\n**Receipt, 2026-08-24.** A spectacle plate was generated first (a street, correctly lit), then the character was added by passing that plate as a second reference. Every attempt came back looking green-screened: her face carried a soft frontal key that existed nowhere in the scene, she was sharper and cooler than the background, and no haze or flare crossed in front of her. **A finished scene handed in as a reference tells the model to paste, not to light.**\r\n\r\n**The fix is to describe one scene from scratch with only the identity reference attached.** Same subject, same location, same everything — the difference is that the model renders one frame in one pass instead of merging two. The first from-scratch attempt landed.\r\n\r\nUse a second reference for *design* (a creature's build, a style) — never for a whole composed frame the character has to be inserted into.\r\n\r\n### 🚨 Write the light as ONE physical system, or the character reads as pasted in\r\n\r\nThe single most common tell is a key light on the subject's face that has no source in the picture. Name all four of these explicitly:\r\n\r\n1. **One source, and where it is.** *\"The sun is low and directly down the avenue behind everything.\"*\r\n2. **What that does to HER specifically.** If she is backlit, say her face is in ambient bounce with no highlight and no catchlight, and put the hard rim on her hair, one shoulder and the collar.\r\n3. **Forbid the key that does not exist.** *\"There is no light in front of her: no frontal key, no fill, no soft source on her face.\"* Without this clause the model adds one every time.\r\n4. **Atmosphere in FRONT of her, not only behind.** Haze, dust and veiling flare must cross the foreground, so her edges are no sharper than the parked cars beside her. **Every shadow in frame runs parallel**, hers included.\r\n\r\n**Let the camera's exposure be part of the physics.** Backlit means the phone lifts for her face and the sky blows to white. Night means the fire blows out and the blacks lift. That single decision ties subject and background into one image better than any amount of description.\r\n\r\n⚠️ **Do not stack grain adjectives.** *\"Heavy luminance noise\"* plus *\"aggressive noise reduction\"* plus *\"milky lifted blacks\"* in one prompt produced a clip so noisy it read as broken. Modern phones are clean at night — name the light, not the sensor damage.\r\n\r\n### Other plate rules\r\n\r\n- **Never describe the phone as an object.** *\"She lifts the phone\"* renders a phone in her hand. Describe what the camera does.\r\n- **Write negatives as positives.** *Matte uneven skin*, not \"not smooth\". *Flat daylight*, not \"no studio lighting\". *Caught mid-step*, not \"not posed\". A \"no X\" list is either ignored or quietly summons X.\r\n- **Text on props renders cleanly and is worth using.** A label reading `YOUR PRODUCT` comes back legible.\r\n- **Content refusals are stochastic. Re-roll the identical prompt before rewriting it.** A plate can refuse once and pass on an unchanged second call. A refused job returns no image and costs nothing. Only redesign after two or three refusals.\r\n- **Generate a pool and pick.** A mediocre plate wastes a metered video roll, and video is the only expensive stage.\r\n\r\n---\r\n\r\n## PHASE 5 — THE VIDEO. One generation per shot.\r\n\r\n`slates_generate_video`, vertical `9:16`, the plate as the first frame. Model routing lives in `slates-model-selection` — read it rather than guessing, and check `slates_estimate_generation_cost` before firing.\r\n\r\n### 🚨 ONE SHOT PER GENERATION. Never a multi-scene take.\r\n\r\nMeasured directly: a six-scene roll returned **nothing usable**, a three-scene roll returned about one usable second, and **the same beats shot as ten separate single-scene generations all landed on the first attempt.**\r\n\r\nCost is linear per second, so splitting is free. What it buys:\r\n\r\n- **A bad shot costs one shot, not the whole ad.** In a long take one broken scene kills every good scene beside it and the retry re-buys all of them.\r\n- **Every shot re-rolls independently**, so a note lands on one shot.\r\n- **8 concurrent jobs means the whole ad shoots in one wave anyway.** A long take buys no wall-clock.\r\n- **No transition can fail**, because there are no transitions inside a generation. Cuts happen in the edit, where they are free and reversible.\r\n\r\nA short line padded into the minimum duration is trimmed in the edit for nothing. **That padding is a far cheaper problem than a dead long take.**\r\n\r\n### The plate is the first frame, always\r\n\r\nPassing the plate as the first frame locks the opening on a composition you already approved, so the model animates a frame you chose instead of inventing one. Pass the identity sheet alongside it as a reference.\r\n\r\n**The one exception** is a shot whose first frame must not exist yet — a reveal, an assembly, anything resolving *into* the plate. There, pass the plate as a reference only.\r\n\r\n⚠️ **Re-read the plate before writing its prompt.** A prompt describing a camera resting on a counter, against a plate that is plainly a held selfie, is a contradiction the model resolves badly.\r\n\r\n### 🚨 Eyeline decides whether a beat lip-syncs or becomes voiceover\r\n\r\n**A character not addressing the lens reads as B-roll, and the model lays the voice over silent footage.** This is not a frame-size problem — the identical shot lip-syncs correctly once the eyeline is locked.\r\n\r\n**Write where her eyes ARE for the whole duration:** *\"her eyes are locked on the lens for the entire shot and she talks straight into it.\"*\r\n\r\n⚠️ **A brief exception becomes the whole shot.** *\"Glances at her feet once and back up\"* produces five seconds of staring at her feet. *\"Never looks down at her hand\"* produces her looking down at her hand. The model takes the vivid clause and runs the beat on it. State the eyeline as the dominant instruction and add exceptions only if you must.\r\n\r\nMixed voiceover and on-camera is a legitimate choice in this format. It has to be chosen, not discovered.\r\n\r\n### 🚨 Name the rig, THEN the operator's body\r\n\r\n**\"Single continuous handheld shot\" tells the model what the camera IS. The operator's body tells it what the camera DOES.** Ship one without the other and it renders its default: a smooth slow push with almost no shake.\r\n\r\n**Write a person's body, never a camera path.** Not *\"the camera drifts along the tailgate\"* but *\"the phone is lying flat on the tailgate metal, held loosely by a hand at the corner, so the frame sags to one side and gets nudged back level once.\"* Arrives, sags, overshoots, corrects, gets bumped, catches up. **And the world must register it** — a move nothing reacts to is a move nobody made.\r\n\r\n**An unusual camera position is still handheld.** Inside a fridge, on a car roof, in the gutter — someone is holding the phone there. A locked-off version reads as a tripod and loses the whole register.\r\n\r\n### Micro-movement is directed, never assumed\r\n\r\nEvery beat carries an explicit list of small involuntary actions. A beat that only says what she says produces a mannequin talking.\r\n\r\n> Small things: she blinks twice, her weight shifts onto one hip, a strand of hair slips forward over her shoulder, and the framing drifts and corrects the way a held phone does.\r\n\r\nOthers that work: a container tipping as she takes its weight · her shoulder rising as she reaches · her hand sliding along a handrail · picking dry skin off a lip · blowing hair out of the corner of her mouth.\r\n\r\n🚨 **Small involuntary business only — never a big physical action at the lens.** Directing her to climb onto something threw her whole body across a close frame and wrecked it. Re-rolled with nothing but a hair flip, a weight shift and two fingers tapping, it landed. **Small business reads as captured; a big action reads as staged, and at close framing it also destroys the composition.**\r\n\r\n⚠️ **When a re-roll must hold the plate's pose, say so as a prohibition:** *\"She does not climb onto the truck, she does not sit down, and she does not lean any further toward the camera than she already is.\"* Pair it with *\"the camera never travels and nothing in the shot jolts it.\"*\r\n\r\n**For a physical feat, direct the struggle, not the event.** Something that gives way on contact reads as a trick; something that **resists for a beat and holds** before it gives reads as real — knuckles whitening, tendons standing out, the forearm tensing, the grip shaking.\r\n\r\n### Audio\r\n\r\n**Always on.** \"No music\" means no score; it never meant no audio. Say what you want present — her lip-synced dialogue, the sound the action makes, footsteps, room tone — then close with **\"No music. No score. No background track of any kind.\"**\r\n\r\nGenerating with audio disabled produces a mouth moving in silence and the roll is wasted.\r\n\r\n---\r\n\r\n## PHASE 6 — THE EDIT\r\n\r\nFree, local, and where the ad is actually made. Trim dead air off a beat that ran slow. Drop a re-rolled beat over the original. Burn captions — all-caps, centred low, one or two words at a time, cut on the stressed syllable.\r\n\r\n**A one-second overrun is an edit problem, not a re-roll.** Re-rolling a shot to fix a second of air is the expensive way round.\r\n\r\n### Verify by WATCHING, never by exit code or by sampling stills\r\n\r\nA job that exits cleanly can still be a refusal with no file, and a landed take can still have a beat that never cut. **Check the file has an audio stream and the right dimensions, then look at a frame from each beat.**\r\n\r\n---\r\n\r\n## Anti-patterns\r\n\r\n- **Don't** write cinematic. Polish is the tell.\r\n- **Don't** put quality words in a prompt. Spend it on imperfections.\r\n- **Don't** combine scenes into one generation to save money. It costs more and returns less.\r\n- **Don't** generate video without a plate behind it.\r\n- **Don't** generate text overlays in an image. Captions happen in the edit.\r\n- **Don't** rewrite a prompt after one content refusal. Re-roll it unchanged first.\r\n- **Don't** describe the reference at the top of a prompt. Describe the photograph you want.\r\n- **Don't** fire the next roll because the last one failed. Diagnose against this file first.\r\n",
+    "slates-vision-feedback-loop": "---\nname: slates-vision-feedback-loop\ndescription: Lower-level utility skill for any Slates workflow that needs to \"generate, look at the result, refine, regenerate.\" Defines the standard inline-vision pattern. Other Slates skills compose this. Use when generating images and you need to confirm they match the brief before moving on, or when the user asks to \"iterate\" on an image.\n---\n\n# Vision feedback loop — Slates utility skill\n\nSlates returns generated images inline as base64. You see the actual pixels. Use that — don't trust prompt-following blindly.\n\n## Asset codes are your shared vocabulary with the user\n\nEvery asset in Slates has a short stable code (e.g. `IMG-A12`, `VID-V3`, `AUD-S1`) and a label derived from its prompt (e.g. `Beach Sunset`). These are visible in the gallery as a corner badge on each thumbnail. **Always refer to assets by their code in chat** so the user can match what you're saying to a specific card in their gallery.\n\n- ✅ \"I'm using **IMG-A12 — Beach Sunset** as the first frame. The second-frame candidate **IMG-A15** has the right composition but warmer light — want me to use that one instead?\"\n- ❌ \"I'm using the beach sunset image...\" (user has four beach sunset variants — which one?)\n- ❌ \"I'm using asset `7a3f9e4b-...`\" (UUIDs aren't readable; user can't match to a badge)\n\nThe code is the FORMAL reference. The label is human texture. Use both: `IMG-A12 — Beach Sunset`.\n\n## Vision tools at your disposal\n\n- `slates_get_asset_image` — pull one image into context. Returns its code+label.\n- `slates_get_assets_batch` — pull up to 8 images in one call. Use when picking from a candidate set; cheaper than N individual fetches.\n- `slates_get_asset_video_frames` — extract N keyframes (default 3) from a video and inline them as JPEGs. You can't see video natively; this is how you \"look at\" a clip before refining its motion prompt.\n\n## Pre-flight is automatic on the gen tools\n\n`slates_generate_video`, `slates_generate_motion_transfer`, and `slates_generate_lip_sync` now show you their reference assets **inline** on the confirm response. You don't need to fetch them yourself — but you DO need to look at what comes back, revise the prompt if the references suggest a different motion/framing, and only then re-call with `confirm=true`.\n\n## 🔴 The still-gate — never animate a bad frame\n\n<!-- @inject:still-gate -->\n**A visible defect in the still is already a STOP.** Do not animate it. Fix the frame first, then move to motion — and go to motion only when the crop passes the still scan and you genuinely need movement to confirm an uncertain edge, reflection, or object.\n\nThis is a **cost** rule as much as a craft rule: a 1080p/10s premium video generation costs many multiples of an image re-roll, and video is where a defect stops being fixable. Anything wrong in the still gets worse in motion — soft geometry mushes, broken-but-plausible objects fall apart, oily textures start crawling. **Animating a known-bad frame is the single most expensive mistake in the pipeline.** Re-rolling the image is the cheap move; re-rolling the video is not.\n<!-- @end:still-gate -->\n\n## The pattern\n\n1. **Generate.** Call `slates_generate_image` with a prompt. The result is in your context as an image content block.\n2. **Evaluate on TWO axes — they are different questions:**\n   - **Brief-conformance** — what did the user actually want? Are the elements right? Composition? Lighting? Subject identity?\n   - **Defects** — run the slop rubric below. *A frame can match the brief perfectly and still be slop that mushes the moment it moves.* Checking only the first axis is how a bad frame reaches an expensive video call.\n3. **One of three outcomes:**\n   - **Right** → save it (bind to a frame, character slot, etc.) and move on.\n   - **Close, but adjustable** → refine with a specific delta, regenerate **once**.\n   - **Wrong direction** → ask the user before regenerating. Don't burn credits on prompt-thrashing.\n\n## The defect rubric — four slop tells\n\n| Tell | What it looks like | Why it matters downstream |\n|---|---|---|\n| **Light with no transitions** | Flat-black pits instead of a shadow ramp; light that stops rather than falls off | Transfers onto every character or object added into that plate later |\n| **Broken-but-plausible objects** | Crates, railings, hardware, mechanisms you can *almost* read but that don't resolve | Turn to mush in motion, and the model multiplies them |\n| **Local logic breaks** | An effect present in only part of the frame — rain scratching one corner, wet ground under one figure | The video model's physical logic breaks along with it |\n| **Oily textures** | Soapy, licked-smooth surfaces that have lost their material identity | Reflections crawl in motion; the plate can't hold continuity |\n\n### Per-model accents — check the one you actually used\n\n- **Nano Banana Pro** (`nano-banana-pro`) — ruler-straight symmetry, everything parallel and square, flat even light, pretty but staged/stock, textures reading as 3D render rather than photograph. **It hyperbolizes every edit**: ask for graffiti on one wall and the whole location gets tagged.\n- **GPT Image** (`gpt-image-2-5-flare`, `gpt-image-2-5-sunburst`) — microcontrast to the ceiling, hard halos on every edge, no depth or bokeh, white balance pulled warm until the frame yellows, plastic licked-smooth materials. Worst tell: **one sickly texture pattern laid over the entire frame**. ⚠️ Catalogued on `gpt-image-2`, which 2.5 replaced on 2026-09-09 — an accent is a per-model observation, so treat this as a prior to check rather than a finding, and correct it here the first time a 2.5 frame disagrees.\n\n> ⚠️ These are accents for **`nano-banana-pro`** and the **GPT Image** line specifically. `nano-banana-2` is a **different model** (Gemini 3.1 Flash Image vs NB Pro's Gemini 3 Pro Image) and we have **no evidence** about its accent. Do not inherit one — say nothing rather than warn about a failure mode you can't substantiate. That caution applies to the GPT Image entry above too: it was measured on `gpt-image-2`, not on either 2.5 seat.\n\n## Where the fault lives — triage before you change anything\n\nWe say \"one specific delta per regeneration\" but that only helps once you know *which* variable to move. Diagnose first:\n\n| Visible pattern | Diagnosis | Fix |\n|---|---|---|\n| The defect exists in the source asset, or stays tied to the same feature when the direction changes | **Source asset** | Fix the sheet / plate, not the prompt |\n| Source is clean, and the defect changes when only the suspect motion clause changes | **Motion direction** | Fix the prompt |\n| Controls conflict, or the failure follows neither variable | **Inconclusive** | Narrow the test — change less, not more |\n\n**Review routes; it is not pass/fail.** Geography melts → fix the location. Identity drifts → fix the character sheet. Assets are sound but the action is wrong → fix the video direction. Wrong idea entirely → reopen the brief with the user.\n\n**Correct the earliest broken handoff.** Polishing a downstream symptom hides the source and guarantees it resurfaces in the next shot built from the same asset.\n\n## Baseline hygiene — isolate the variable you're testing\n\nWhen the **character** is the question, keep the location out of it: test on a plate that already holds its own geometry, depth, materials, and light. **A broken plate gives every character failure a second plausible cause**, and you will spend re-rolls deciding which one you're looking at. The same applies in reverse — test a plate empty before you populate it.\n\n## Refinement rules\n\n- **One specific delta per regeneration.** Don't change five things at once — you won't know what helped.\n- **Rewrite the FULL prompt on every iteration — never a diff, never a fragment.** Change one decision, then re-emit the whole prompt so every slot still agrees with every other slot. This composes with the rule above rather than replacing it: *one delta* governs **what changes**, *full rewrite* governs **how you re-emit it**. A patched fragment leaves the old slots stale and silently contradicting the new one.\n  - On **Seedance**, a re-emit must keep the `Shot N` structure intact — see `slates-prompting-seedance`.\n  - **Exception — Omni Flash Edit.** Long prompts documentedly destroy its fidelity. There the rule inverts: one short instruction plus *\"Keep everything else the same.\"*\n- **Anchor with references.** If the result drifted from the user's intent, attach the *previous best* generation as a reference image alongside the original brief.\n- **Use `slates_get_asset_image`** to pull a previously-generated image back into context if you need to compare against a fresh generation.\n- **Use `slates_edit_image`** for surgical tweaks instead of full regeneration when ~90% of the image is right — `sourceAssetId` = the asset, `prompt` = the change only. Edits preserve composition and identity; full regen rolls the dice. Recipe: `slates-edit-and-iterate`.\n\n## Cost discipline\n\n- Track total credits spent across the loop. Surface to the user every 3 iterations.\n- Stop after 3 failed iterations on the same prompt — escalate to the user with what you tried and what's not working. The slot machine never converges.\n- For a high-cost generation — anything past the confirm gate in `slates-cost-discipline` — confirm before *every* attempt, not just the first.\n\n## When to break the loop\n\n- The user said \"good enough\" or \"ship it.\" Stop iterating.\n- You've burned >5 generations on one frame. Hand back and ask.\n- The user changes brief mid-loop. Treat it as a new brief, not a continuation.\n\n## Voice when narrating to the user\n\nTight, observational, no editorializing.\n- ✅ \"Frame 2 has the wrong lighting direction — back-lit instead of side. Regenerating with side light.\"\n- ❌ \"I notice that the lighting in frame 2 isn't quite what we were going for. I'll go ahead and try again with a different approach.\"\n",
 };
 //# sourceMappingURL=content.js.map