@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/xingchen-vox-collage/SKILL.md

@@ -1,251 +0,0 @@
----
-name: xingchen-vox-collage
-description: "Create or revise Xingchen scenes in an original Vox-inspired editorial paper-collage language with separate image assets, code-owned typography and evidence graphics, deterministic AICUT motion, hero-frame lookdev, and anti-PPT visual QA. Use when the user explicitly asks for Vox-style, editorial collage, paper-cut explainer, kinetic editorial typography, collage B-roll, or wants to reconstruct this visual grammar inside a Xingchen video. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "Phantomlau3674/voxstylehub-steven"
----
-
-# Xingchen Vox Collage
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-Own the selected editorial-collage visual branch. Preserve the approved Xingchen thesis, proof, scene job, timing, and safe regions. Do not become a second story router or render truth.
-
-## Route The Picture
-
-Choose one route per scene:
-
-- `editorial-explainer` (default): generate or source independent assets, keep exact text and evidence code-native, and animate deterministically in Remotion.
-- `metaphor-broll`: use one visual proposition, a focused cast of object groups, and a concise assembly or transformation. Prefer deterministic layers; allow generated image-to-video only when exact layout, identity, text, counts, and proof do not matter.
-- `source-collage`: preserve real source geometry, crop it into editorial sheets or cutouts, and add code-owned annotation. Never present generated media as evidence.
-
-Do not call a paper texture or a Ken Burns move a Vox scene. The screen must reveal, compare, trace, build, transform, inspect, or resolve something the narration cannot show alone.
-
-## Read Inputs
-
-Read only the current scene truth:
-
-- accepted script or real-audio timing;
-- `scene_id`, `beat_id`, `knowledge_change`, `dominant_visual`, `motion_action`, `proof_ref`, `safe_region`, and `semantic_relation`;
-- selected whole-film art direction and any explicit reference-style record;
-- existing source media and project-local assets.
-
-If the project runs under `xingchen-next`, keep its `Content Lock` and `Preview Lock`. A style bake-off, asset check, or hero-frame choice is a working decision, not another approval lock.
-
-## Explore Before You Commit
-
-Before locking routes or shot cards, write one short creative exploration in free-form prose — a paragraph, not a form. In your own words, answer:
-
-- Which visual routes could carry this film, and which one did you choose and why?
-- What material opportunities does the subject offer — papers, inks, objects, archives, or textures the argument can physically live in?
-- Where might the film become sparse or repetitive, and what would rescue those passages?
-- How does motion carry meaning here — what does each entrance, transformation, or settle explain — instead of decorating?
-
-This exploration is an authoring default, not validator input: no schema, no required line count, no score. Keep it with the project's working notes and let it steer the shot cards. For a copyable visual-ambition brief and a worked example, read [creative-exploration.md](references/creative-exploration.md).
-
-## Lock The Shot Before Assets
-
-For every scene, write a shot card before prompting or cutting: the visual proposition, focal order, subject scale, facing, common ground or deliberate suspension, occlusion order, and subtitle-safe lane. Treat a change of shot as a new composition; do not reuse a character blindly when its scale, facing, or narrative role changes.
-
-Keep moving people and props out of the environment plate. A matching character sheet is allowed as a staging reference, but split accepted actors into individual alpha-checked files before animation. The settled layout must make the narrative hierarchy legible while completely still.
-
-## Build The Visual Contract
-
-Create `visual/vox/` with:
-
-- `DESIGN.md`: material world, palette, type roles, edge language, shadow logic, texture, motion character, anti-reference;
-- `scene-spec.json`: scene jobs, hero frames, independent layers, camera, transitions, checkpoints, and rendered review evidence;
-- `assets.json`: every source, generated, local, and code-native asset with status and provenance;
-- `prompts/`: one prompt per generated bitmap asset;
-- `lookdev/`: intended-resolution hero frames and playable critical clips;
-- `renders/`: branch previews and visual evidence.
-
-Initialize the structure:
-
-```powershell
-python .\scripts\init_vox_branch.py <project-root> --slug <slug> --width 1920 --height 1080 --fps 30
-```
-```
-Read [visual-language.md](references/visual-language.md) while writing `DESIGN.md`. Read [asset-layer-contract.md](references/asset-layer-contract.md) before generating or cutting assets.
-```
-
-## Design The Hero Frame First
-
-Hero-frame-first is the authoring default, not a universal hard gate. For the hook, hardest explanation or proof, and payoff:
-
-1. Write one visual proposition, not a list of nouns.
-2. Sketch two genuinely different spatial mechanisms using the same accepted scene job.
-3. Select the strongest composition and construct its settled hero frame at intended resolution.
-4. Verify focal hierarchy, negative space, proof scale, subtitle safety, and material consistency.
-5. Decompose the accepted frame into independent actors before animating.
-
-Do not approve motion from a prompt or storyboard. Render the hero frame and then a playable clip.
-
-When a scene's meaning exists only in motion — a traversal, a transformation, a settle — a settled still cannot carry the argument. In that case, review a representative in-motion frame extracted from the playable clip and record a one-sentence rationale next to it stating why the still cannot stand in. This exception never relaxes source-collage readability: real source proof must stay legible and undistorted in whatever frame is reviewed.
-
-Before strict validation, each scene must point to real review artifacts:
-
-```json
-{
-  "hero_frame": {
-    "description": "...",
-    "focal_order": ["..."],
-    "subtitle_safe_region": "...",
-    "path": "visual/vox/lookdev/s01-hero.png",
-    "input_fingerprint": "<evidence_input_fingerprint>"
-  },
-  "playable_clip": {
-    "path": "visual/vox/lookdev/s01-preview.mp4",
-    "audio_ref": "audio/narration.wav",
-    "input_fingerprint": "<evidence_input_fingerprint>",
-    "phone_reviewed": true,
-    "phone_review_ref": "visual/vox/lookdev/s01-phone-review.md",
-    "checkpoint_paths": {
-      "entry": "visual/vox/lookdev/s01-entry.png",
-      "settled": "visual/vox/lookdev/s01-settled.png",
-      "exit": "visual/vox/lookdev/s01-exit.png"
-    }
-  }
-}
-
-## Generate And Prepare Assets
-
-Use `$imagegen` for raster plates, isolated cutouts, props, archival-style illustrations, or textures. Follow its built-in-first policy.
-
-- Generate each independently moving or reusable actor as a separate asset.
-- Keep headlines, Chinese text, dates, numbers, charts, maps, labels, captions, and proof highlights out of generated bitmaps.
-- Request isolated opaque cutouts on a removable flat chroma background when practical; use the installed ImageGen removal helper and inspect alpha edges.
-- Use a generated actor sheet only to establish a coherent cast; split it into independently positioned cutouts and reject it as a final animated plate.
-- Use a reference sheet only for a recurring person, place, or object that must stay recognizable across scenes.
-- Save accepted assets inside the project. Do not reference `$CODEX_HOME/generated_images` from Remotion.
-- Give every `reviewed` or `approved` code-native asset a project-relative `path`. When a path uses `file.html#fragment`, that literal DOM/SVG `id` must exist in the file.
-- Preserve real source crops without non-uniform scaling or invented content.
-
-For a dense existing poster, prefer regenerating or extracting separate actors. Manual bounding-box extraction is a last-mile repair, not the default production method.
-
-## Compile Deterministic Motion
-
-Load `remotion-render-adapter` for formal Remotion work. Copy the neutral primitives when the project lacks an equivalent library:
-
-```powershell
-python .\scripts\install_remotion_primitives.py <existing-remotion-project-dir>
-```
-```
-Use the copied primitives as ingredients, not a scene template. Compose a film-specific picture.
-
-Apply paper-native verbs:
-
-- `slide`, `reveal`, `peel`, `pivot`, `stamp`, `drop`, `slap`, `settle`, `trace`, `wipe`;
-- shallow depth-separated parallax;
-- mask and crop reveals;
-- finite grain, dust, scanner-light, or paper flutter;
-- short impact shake only on a meaningful landing.
-
-Build the settled layout first. Animate backward from final coordinates to hidden or off-frame states. Preserve the final hero frame until the transition begins.
-
-Assign every independent actor a `primary`, `secondary`, `tertiary`, or `ambient` motion tier. Let primary actors travel and land most strongly, secondary actors arrive after the focal read is established, tertiary actors move least, and the plate remain almost still. Do not give three or more independent actors the same entrance frame unless the scene contract explains why their simultaneous arrival carries meaning.
-
-Keep motion deterministic: derive every value from the Remotion frame and stable seeds. Do not use wall-clock time, unseeded randomness, asynchronous timeline construction, or infinite repeats.
-
-Read [remotion-scene-contract.md](references/remotion-scene-contract.md) for layer props, motion timing, camera rules, and handoff fields.
-```
-
-## Prevent The PPT Failure
-
-Block the branch when any of these is true:
-
-- the dominant visual is a centered title or generic card;
-- every scene repeats the same torn rectangle, grid, or left-text/right-image layout;
-- narration is pasted onto the canvas;
-- a flattened poster is described as independent object animation;
-- charts, maps, or evidence are decorative placeholders;
-- labels, connector lines, glow, or particles are being added to rescue a weak picture mechanism;
-- a generated image contains essential Chinese text or audited numbers;
-- a template selects scene type by index instead of scene meaning;
-- a scale-only move is described as camera choreography.
-- subtitles or labels are the only changing layer across an explanatory interval;
-- the essential semantic actor becomes a postage stamp in the phone downsample while decorative space dominates;
-- an adjacent-scene cut abandons material, object, rail, aperture, color-field, or motion continuity without recording an intentional reset.
-
-Allow quiet frames and hard cuts. More layers and more movement are not automatic improvements.
-
-## Validate And Review
-
-While the branch is still being authored, validate its contract with pending evidence allowed:
-
-```powershell
-python .\scripts\validate_vox_branch.py <project-root> --allow-pending
-```
-```
-`--allow-pending` converts missing hero-frame and playable-clip evidence into warnings. A `PASS` with warnings means the structure is usable; it is not visual approval.
-
-The validator can only confirm that a review artifact exists, decodes, and matches its contract. It cannot confirm that anyone watched the clip at phone size, and it cannot judge readability or aesthetics — that judgment stays human. A `phone_review_ref` path or `phone_reviewed: true` flag proves the artifact is present, never that a human looked at it or approved the picture.
-
-After the intended-resolution hero frame, accepted-audio playable clip, entry/settled/exit checkpoints, phone review, and full FFmpeg decode exist, run strict validation:
-```
-```powershell
-python .\scripts\validate_vox_branch.py <project-root>
-```
-```
-Strict validation checks project-local code-native files and fragments, hero-frame resolution and decodability, playable video and audio streams, full clip decode, checkpoint files, phone-review artifacts, adjacent-scene transition contracts, and synchronization between `project-state.json` and `scene-spec.json` when Lean state exists.
-
-After rendering a clip, create visual evidence:
-```
-```powershell
-python .\scripts\make_visual_evidence.py <clip.mp4> <evidence-dir>
-```
-```
-After assembling a full master, extract consistent per-scene clips and entry/settled/exit frames from the Lean scene timings:
-```
-```powershell
-python .\scripts\lock_vox_inputs.py <project-root> --write
-python .\scripts\build_scene_evidence.py <project-root> <evidence-dir>
-```
-```
-The lock command hashes the accepted audio inputs, canonical scene timing, and assembled source master. Run it again whenever any of those inputs changes. The extractor refuses to overwrite a non-empty evidence directory, verifies the lock, and writes the same `evidence_input_fingerprint` into its machine-readable index. Record that value in each reviewed `hero_frame.input_fingerprint` and `playable_clip.input_fingerprint`; a later input change then makes the old evidence fail strict validation instead of silently passing.
-
-Inspect the contact sheet and the entry, settled, and exit frames. Then watch the playable clip at phone size. Use [visual-qa.md](references/visual-qa.md). Do not accept from process exit, code size, component count, or stills alone.
-
-Check the settled frame for focal order, common grounding, facing, occlusion, and subtitle clearance. Check the clip for staggered entrances and whether each impact, whoosh, or tick lands on the visual event it is meant to explain.
-
-Handoff accepted scene implementations to `xingchen-lookdev`, then use `remotion-render-adapter` for the formal render job. Keep rejected visual hypotheses visible until the full preview confirms the winner.
-```
-
-## Reference Study Boundary
-
-Read [learned-branches.md](references/learned-branches.md) when reconstructing a public Vox-style result. Reuse causal grammar, not logos, title packaging, exact compositions, proprietary assets, or copied prompts.
-
-## Source
-
-Rewritten for AICUT from [`Phantomlau3674/voxstylehub-steven`](https://github.com/Phantomlau3674/voxstylehub-steven)
-(MIT), xingchen-vox-collage/SKILL.md. The method is that skill's own; the way it runs is AICUT's.

package/skills/xingchen-vox-collage/references/asset-layer-contract.md

@@ -1,98 +0,0 @@
-# Asset And Layer Contract
-
-Use this reference before generating, extracting, or accepting assets.
-
-## Shot Lock
-
-Before creating assets, record for each shot:
-
-- focal order and relative scale;
-- facing and sight-line relationships;
-- common ground line, or an explicit reason the actors float;
-- front-to-back occlusion stack;
-- subtitle-safe lane.
-
-Do not treat a new shot as a crop of the previous shot when it changes any of those decisions. The environment plate contains no actor that must enter, leave, resize, or change z-order independently.
-
-## Layer Roles
-
-Every `scene-spec.json` layer must use one role:
-
-- `plate`: substrate, environment, or large static field;
-- `source`: real evidence with preserved geometry;
-- `cutout`: isolated person, object, place, or archival illustration;
-- `prop`: supporting object with independent timing;
-- `type`: exact live headline or label;
-- `data`: chart, map, number, timeline, or evidence graphic;
-- `annotation`: arrow, circle, underline, stamp, marker, or connector;
-- `occluder`: foreground paper edge or object used for depth/reveal;
-- `texture`: non-semantic grain, dust, light, ink, or print imperfection.
-
-## Separation Test
-
-Keep an actor separate when any answer is yes:
-
-- Does it enter, exit, pivot, scale, mask, or parallax independently?
-- Must it remain exact while another element changes?
-- Will it be reused in another scene?
-- Does its z-order change?
-- Does it carry proof, identity, or an exact label?
-
-Merge repeated low-value decoration into a plate when separation adds no visual control.
-
-## Generated Raster Rules
-
-- Use one prompt per distinct asset or coherent variant.
-- Record the final prompt path and accepted file path.
-- Leave exact Chinese text, numbers, brand copy, and evidence labels out of the bitmap.
-- Ask for crop room based on intended placement, not a generic centered portrait.
-- For opaque cutouts, generate on a flat removable chroma background and validate the final alpha edge.
-- A cast or actor sheet may establish visual consistency, but it is not an animation-ready layer: split every accepted actor into an individual alpha-checked cutout.
-- For hair, glass, smoke, liquids, reflections, or translucent material, use a plate, mask, or confirmed native-transparency route; do not pretend a broken matte is acceptable.
-- Reject duplicated limbs, accidental words, watermarks, identity drift, fake proof cues, and lighting inconsistent with `DESIGN.md`.
-
-## Existing Poster Decomposition
-
-If only a flattened poster exists:
-
-1. identify bounding boxes and intended final centers;
-2. extract crops;
-3. remove backgrounds for opaque subjects;
-4. erase duplicated neighboring objects;
-5. inpaint or locally blur vacated landing regions;
-6. animate actors back to original centers so the settled frame reconstructs the approved poster.
-
-Do not leave sharp duplicate actors in the backdrop. Prefer a clean plate or separately generated actors over repeated manual cleanup.
-
-## Asset Manifest Minimum
-
-Record:
-
-```json
-{
-  "id": "hero-cutout",
-  "role": "cutout",
-  "source_kind": "generated",
-  "path": "visual/vox/media/images/hero-cutout.png",
-  "prompt_ref": "visual/vox/prompts/hero-cutout.txt",
-  "status": "approved",
-  "alpha_checked": true,
-  "notes": "No embedded text; designed for right-third placement"
-}
-```
-
-Use `source_kind`: `source`, `local`, `generated`, or `code-native`. Generated assets never become proof.
-
-For a reviewed or approved code-native actor, `path` is still required. Point to a project-relative source file, and use a fragment when several independently addressable actors share one SVG or HTML file:
-
-```json
-{
-  "id": "viewing-rays",
-  "role": "annotation",
-  "source_kind": "code-native",
-  "path": "visual/vox/media/code/hero-settled.html#viewing-rays",
-  "status": "reviewed"
-}
-```
-
-The fragment must be a literal DOM or SVG `id` in that file. A manifest name without an existing file and addressable fragment is not an independent layer.

package/skills/xingchen-vox-collage/references/creative-exploration.md

@@ -1,19 +0,0 @@
-# Creative Exploration And Visual Ambition
-
-Working notes for the `Explore Before You Commit` step in [SKILL.md](../SKILL.md). The exploration is free-form prose, an authoring default, and never validator input.
-
-## Visual Ambition Paragraph For Bounded Briefs
-
-When a root agent delegates a bounded implementation task, paste this paragraph into the brief and adapt it to the scene. It is provider-neutral and deliberately carries no numeric quotas:
-
-> Aim high visually. Source or generate independent materials for this scene instead of flattening everything into one plate. Give the frame one clear composition with an unmistakable focal carrier and intentional negative space. Let motion explain something — an entrance that establishes hierarchy, a transformation that shows cause, a settle that marks resolution — rather than adding movement for energy. If any passage turns sparse or repetitive, rescue it by changing the spatial mechanism, scale, material, or crop, not by stacking more layers or effects. Exact text, numbers, and evidence stay code-native and readable at phone size.
-
-## Example Exploration (Fictional, Publication-Safe)
-
-Topic: why the writing on a thermal receipt fades away. All characters and objects below are original paper-collage inventions.
-
-> Two routes could carry this film: a chemistry route that turns the receipt into a lab bench of paper test strips, or a time-lapse route that treats the receipt as a tiny landscape slowly bleached by sun and time. I choose the lab bench, because the fading is a chemical story and the bench lets heat, light, and air each appear as a physical actor. The material opportunities are unusually rich: thermal paper itself is the substrate, so the film can live on one long roll of receipt paper; a paper-cut lab technician moth, "Doctor Dot," can guide the eye without ever carrying text; heat can be a vermilion iron-shaped cutout that stamps letters into existence and later steals them back. The sparse risk is the middle — once the mechanism is explained, two more examples of fading would repeat the same picture — so the middle passage changes mechanism instead: the receipt is folded into a timeline accordion, and each fold is a month the ink survives. Motion carries the meaning: the iron stamps, the moth peels a letter off the page to show dye molecules letting go, and the accordion settles flat when the receipt is finally blank. Nothing moves just to sparkle.
-
-## Case-Study Note: Observed Density Facts, Never Quotas
-
-One successful 86-second production in this visual language shipped with 8 scenes, 103 layers, and 78 approved assets. These are observed facts from that single production, recorded here only to show what one healthy film happened to contain. They are not targets, recommended ranges, or quotas of any kind. A stronger film may use fewer or more; the ambition paragraph above deliberately prescribes no counts, and no review should ever cite these numbers as a bar to clear.

package/skills/xingchen-vox-collage/references/learned-branches.md

@@ -1,34 +0,0 @@
-# Learned Public Branches
-
-Use this reference only when reconstructing the public July 2026 wave of Vox-inspired agent skills. Treat repositories as observations, not authorities.
-
-## What To Keep
-
-| Branch | Keep |
-| --- | --- |
-| `pyang5166/gbro-collage-broll` | one-line visual proposition, final-frame approval, small-object metaphor mode, first/end-frame QA |
-| `Alisa0808/vox-director` | topic-matched style bake-off, whole-poster versus element-level routing, local layer motion, original-position reconstruction |
-| `kakaxi12/vox-director-codex` | separate generated assets, live HTML/CSS text, deterministic code motion, project-bound asset manifest |
-| `CK42BB/vox-explainer-skill` | approved style anchor and reference-image chain for consistent stills |
-| `sadekxD/vox-style-collage-video-skill` | reference-sheet economy, one-shot/one-point discipline, hero-frame and empty-plate distinction |
-| `ZeroZ-lab/vox-video-maker` | distinguish evidence sheets, generated stickers, and code-native graphics |
-
-## What To Reject
-
-- Provider or model IDs as core workflow logic.
-- Treating two flattened endpoints as guaranteed layer control.
-- Essential text baked into generated posters.
-- Fixed scene types selected by array index.
-- A generic torn-paper card around every object.
-- Audio, research, or packaging complexity used to hide weak visuals.
-- Repository popularity as proof of render quality.
-
-## Reconstructed Causal Grammar
-
-Use this sequence:
-
-`scene job -> visual proposition -> style DNA -> settled hero frame -> independent actors -> deterministic motion -> rendered evidence -> revision`
-
-For short metaphors, an optional generated-video route may replace deterministic motion only after the scene proves it has no exact text, proof, identity, count, or endpoint requirement.
-
-Do not copy Vox logos, proprietary title packaging, exact layouts, private prompts, or source assets.

package/skills/xingchen-vox-collage/references/remotion-scene-contract.md

@@ -1,77 +0,0 @@
-# Remotion Scene Contract
-
-Use this reference while implementing selected scenes.
-
-## Layer Fields
-
-Each actor needs:
-
-- stable `id`, `role`, and `z`;
-- `motion_tier` of `primary`, `secondary`, `tertiary`, or `ambient` for independently timed cutouts, props, and occluders;
-- final `x`, `y`, `scale`, `rotation`, and optional `anchor`;
-- facing and grounding notes when the actor represents a person, creature, or placeable object;
-- `source_ref` or live `content`;
-- entrance verb, start frame, and duration;
-- optional finite ambient motion;
-- proof and safe-region constraints.
-
-Final coordinates are authoritative. Derive entrance states from them.
-
-## Hierarchy And Timing
-
-Set the still composition before adding motion. Establish the primary read first, then place secondary and tertiary actors around it without obscuring a face, hand, prop, or proof target. Use the `z` stack to make foreground occlusion intentional rather than accidental.
-
-Let the primary tier travel furthest and have the clearest landing; give secondary actors a smaller delayed entrance; keep tertiary actors restrained; keep ambient plate movement near-static. Do not start three or more independent actors on the same frame unless `intentional_group_entrance` records the narrative reason.
-
-Treat paired sound as an event contract: `impact`, `whoosh`, `tick`, or silence must align with the action it supports, not merely with a convenient timestamp.
-
-## Motion Vocabulary
-
-Use material-native actions:
-
-- `slide`: translate from one edge with ease-out and slight settle;
-- `slap`: start slightly enlarged, become opaque quickly, then settle;
-- `drop`: fall with one controlled bounce;
-- `pivot`: rotate around a believable tape, staple, hinge, or corner;
-- `peel`: clip or mask reveal from an anchored edge;
-- `stamp`: fast scale/opacity landing with a short impact hold;
-- `trace`: SVG stroke draw for a real relationship;
-- `wipe`: full-frame paper field used sparingly for chapter change;
-- `resolve`: scattered actors settle into the final explanatory composition.
-
-Start from scale around 0.92-0.97 or 1.15-1.35 when using scale. Do not use `scale(0)`.
-
-For 30 fps, use these as starting points, then tune by viewing:
-
-- small paper/label entrance: 8-14 frames;
-- hero cutout entrance: 12-22 frames;
-- grouped stagger: 3-6 frames;
-- annotation draw: 14-28 frames;
-- impact shake: 3-7 frames with rapid decay;
-- final reading hold: at least 12-24 frames when the argument needs it.
-
-## Camera
-
-Keep the camera parallel to the editorial plane unless the scene has a real spatial reason.
-
-- Default to hard cuts, locked camera, or a restrained push from 1.00 to about 1.04-1.08.
-- Use parallax by moving separated depth groups, not by pretending one raster has layers.
-- Use overscan for camera shake or pan so no black edges appear.
-- Preserve proof geometry and type legibility.
-
-## Determinism
-
-Derive animation from `useCurrentFrame()`, `useVideoConfig()`, `spring()`, and `interpolate()`. Use stable seeded functions for torn edges, dust, and grain. Do not use `Math.random()`, current time, asynchronous mutations, or infinite CSS animation.
-
-## Critical Evidence
-
-For a motion-graphic scene, render:
-
-- `entry`: proposition is visible but unresolved;
-- `settled`: composition and hierarchy are complete;
-- `exit`: outgoing frame is preserved until transition ownership begins;
-- playable clip with accepted audio.
-
-Stills diagnose composition. The clip decides rhythm, easing, continuity, and physicality.
-
-Write these artifacts back into the scene contract: `hero_frame.path` plus `playable_clip.path`, `playable_clip.audio_ref`, `playable_clip.phone_reviewed`, and `playable_clip.checkpoint_paths.entry|settled|exit`. Pending validation may warn about missing evidence; strict validation must probe and fully decode the referenced media.

package/skills/xingchen-vox-collage/references/visual-language.md

@@ -1,74 +0,0 @@
-# Editorial Collage Visual Language
-
-Use this reference to write `visual/vox/DESIGN.md` and judge hero frames.
-
-## The Visual Thesis
-
-Define one sentence describing how the subject becomes a material world. Avoid naming a style alone.
-
-Good:
-
-> The argument appears as evidence clipped onto a warm newsroom worktable; one vermilion mark interrupts monochrome source fragments as the causal chain is assembled.
-
-Weak:
-
-> Vox style, paper collage, cool animation.
-
-## Locked Style DNA
-
-Lock these dimensions before expanding scenes:
-
-1. `substrate`: off-white paper, kraft worktable, flat color field, archive sheet, or another film-specific ground;
-2. `edge_language`: torn, scissor-cut, or geometric; choose a dominant family;
-3. `image_treatment`: full color, monochrome, duotone, halftone, engraving, or screenprint;
-4. `palette`: substrate, paper, ink, one primary accent, and at most one supporting accent;
-5. `shadow_logic`: one direction, one depth range, one softness family;
-6. `grain`: paper fiber, risograph, newsprint, photocopy, or clean editorial;
-7. `type_roles`: display, information, data, and subtitle roles using project-approved fonts;
-8. `annotation`: marker, pencil, grease pencil, stamp, tape, or editorial rule;
-9. `motion_character`: precise, punchy, restrained, elastic, urgent, or lyrical;
-10. `anti_reference`: what must not appear.
-
-## Composition
-
-- Give the frame one unmistakable argument carrier.
-- Use scale contrast before decoration.
-- Place generated cutouts directly on the substrate or a localized color field. Do not wrap every asset in another paper card.
-- Use foreground occluders to create shallow depth and reveal direction.
-- Keep evidence larger and quieter than decoration.
-- Reserve negative space intentionally for live type, subtitles, or a later reveal.
-- Let adjacent scenes vary scale, density, temperature, crop, or spatial mechanism for a semantic reason.
-
-## Typography
-
-Keep exact text live in Remotion. Use generated text only as deliberately unreadable texture.
-
-- Display copy: short assertion, normally one or two lines.
-- Information copy: labels, captions, source notes, and explanatory text.
-- Data copy: years, percentages, prices, counters, and chart values.
-- Subtitles: separate layer; never let a title composition determine caption placement.
-
-For a 1920x1080 Douyin master, begin around 58-64 px subtitles, 44-52 px thesis labels, and 38-46 px important nodes. Test on a phone downsample.
-
-## Vox Grammar Versus Collage Decoration
-
-The branch is successful when the frame demonstrates an argument through one of these mechanisms:
-
-- evidence is exposed and annotated;
-- two states are compared in the same spatial system;
-- a causal chain is assembled;
-- a number changes scale against a stable reference;
-- a map or timeline is traversed;
-- an object transforms into the consequence it represents;
-- an established pattern is interrupted and reframed.
-
-Paper texture without one of these jobs is decoration.
-
-## Hero Frame Checklist
-
-- Can a cold viewer name the focal object in one second?
-- Can they infer the relationship without production notes?
-- Does the visual remain specific with subtitles hidden?
-- Are source proof and generated illustration visually distinguishable?
-- Does every independently moving actor exist as a layer or code-native object?
-- Is the final composition worth holding for at least half a second?

package/skills/xingchen-vox-collage/references/visual-qa.md

@@ -1,59 +0,0 @@
-# Visual QA
-
-Review the rendered result, not the prompt or source code.
-
-## Cold Frame Review
-
-Hide subtitles and production notes. Inspect the contact sheet at thumbnail size.
-
-- Is the focal object immediate?
-- At phone downsample size, is the essential semantic actor still identifiable, or has it become a postage stamp inside decorative space?
-- Does the sequence show a real state change?
-- Are adjacent frames meaningfully different without becoming visually unrelated?
-- Does the material world remain coherent?
-- Does any frame collapse into a centered slide or generic card grid?
-- Are proof and exact type readable without zooming?
-
-## Motion Review
-
-Watch once at normal speed, once muted, and once frame-by-frame around entrances and transitions.
-
-Also watch once with subtitles and exact labels hidden. Caption changes do not count as picture-state changes.
-
-- Does each movement have a semantic or physical reason?
-- Do paper actors settle rather than float forever?
-- Are entrances staggered by hierarchy instead of everything arriving together?
-- Does a camera move direct attention rather than manufacture energy?
-- Are text, faces, logos, and evidence stable?
-- Does the final hero frame reconstruct cleanly without duplicate ghosts?
-- Do cuts preserve continuity anchors where needed?
-- Can each adjacent-scene handoff name a carried material, object, rail, aperture, color field, or motion vector, or an explicit intentional reset?
-
-## Anti-PPT Blockers
-
-Block when:
-
-- six or more consecutive scenes share the same container geometry;
-- the video is readable as a slide deck with transitions removed;
-- large text is the dominant visual because the picture is missing;
-- motion consists mainly of opacity, scale, labels, and connector lines;
-- subtitles or labels are the only changing layer through an explanatory interval;
-- essential semantic actors are too small to identify in the phone downsample even though the subtitles are readable;
-- a cut breaks the material world without a carried anchor or an intentional reset;
-- generated plates are used as full-frame backgrounds behind summary bullets;
-- code paths or component counts are offered instead of viewing evidence.
-
-## Technical Evidence
-
-Require:
-
-- positive duration and expected dimensions/fps;
-- full FFmpeg decode without error;
-- nonblack/nontransparent sampled frames;
-- contact sheet and named checkpoint frames;
-- actual asset paths, with no placeholders or remote-expired URLs;
-- final preview duration aligned to accepted audio when audio exists.
-
-These requirements only prove that review artifacts exist and decode. Artifact existence never proves that a human watched the result or approved it — the `Approve` / `Revise` / `Block` call below is a human judgment.
-
-Return `Approve`, `Revise`, or `Block`, with scene ids or timestamps and the narrowest repair.