@wji.ai/cli 0.5.0 → 0.6.0

package/skill/ads/reference/product-photography.md

@@ -0,0 +1,119 @@
+# Product photography — lighting, shots, backgrounds, camera, grade
+
+The studio vocabulary (棚拍語彙) an image or video model understands, organised as the parts of
+a prompt: lighting setup, shot type, background, camera and lens, colour grade. Each row gives
+the look, what it is for, and the phrase to write. The style presets in `reference/styles.md`
+pick from these tables; `reference/ai-consistency.md` says how to keep the product identical
+between shots.
+
+## Lighting setups (布光)
+
+| Setup | Look | Use for | Prompt phrase |
+| --- | --- | --- | --- |
+| Three-point (三點光) | key from 45°, fill opposite at half power, back light on the edge; even, dimensional, "catalogue" | any product, safe default | `three-point studio lighting, soft key from upper left, gentle fill, thin rim light separating the product from the background` |
+| Large softbox (柔光箱) | wrap-around soft light, broad soft highlight, faint shadow; closer = softer | packaging, skincare, food, anything matte | `single large softbox close above and to the left, soft wrap-around light, one soft highlight, faint shadow` |
+| Rim / backlight (輪廓光 / 逆光) | bright edges, dark centre, glow through liquid | glass, bottles, dark products, anything translucent | `strong backlight through the bottle, glowing rim light on both edges, dark background, no front fill` |
+| Clamshell (蚌殼光) | key above, reflector below, nearly shadow-free, luminous | cosmetics, jewellery, beauty, faces | `clamshell beauty lighting, soft light from above and below, shadowless, dewy highlights` |
+| High-key (高調) | white background lit to pure white, few shadows, airy | pharmacy, tech accessories, e-commerce packshots | `high-key lighting, pure white seamless background, minimal shadow, bright and clean` |
+| Low-key (低調) | one key at 45°, deep shadow, product carved out of black; premium | spirits, watches, luxury, automotive | `low-key lighting, single hard key from the right, deep black shadows, product emerging from darkness` |
+| Hard light (硬光) | small source, crisp shadows, graphic, punchy | fashion, streetwear, retail-loud, editorial | `hard direct light, sharp-edged shadow cast on the wall, high contrast, graphic` |
+| Strip lights (條形光) | long thin highlights that follow the product's curve; gradient reflections | gloss and metal: cans, bottles, cars, phones | `two vertical strip lights either side, long gradient highlights along the glossy surface, controlled reflections` |
+
+Combine at most two (rim + softbox is the standard bottle setup). Name the *direction* and the
+*result on the product* ("soft highlight on the shoulder of the bottle"), not the equipment
+brand; the model renders results, not gear.
+
+## Shot types (景別 / 拍法)
+
+| Shot | What it shows | Use for | Prompt phrase |
+| --- | --- | --- | --- |
+| Packshot (產品照) | the product alone, label to camera, on white / neutral, eye-level, f/8–f/11 sharpness edge to edge | identity reference, e-commerce, end card | `clean product packshot, front label facing camera, centred, white seamless background, evenly lit, sharp throughout` |
+| Hero (主視覺照) | low angle, dramatic light, product as monument | KV, poster, TVC end frame | `hero shot from a low angle, product towering, dramatic rim light, dark gradient background` |
+| Lifestyle (情境照) | product in use in a real setting, natural light, shallow depth | social, feed, testimonial stills | `lifestyle photo, product held at a sunlit kitchen counter, window light, shallow depth of field, candid` |
+| Flat lay (平鋪) | top-down, product with props arranged on a surface | cosmetics, food, stationery, social | `flat lay, top-down view, product centred among a few props on linen, soft even light` |
+| Macro (微距) | texture, condensation, fabric weave, a logo emboss | hooks, cut-aways, quality proof | `extreme macro, 100 mm macro lens, condensation droplets on the label, shallow focus` |
+| In-hand (手持) | a hand holding or opening the product; scale and use | UGC, demo, app / device | `product held in one hand at chest height, natural skin, hand steady, label readable` |
+| Floating (懸浮) | product suspended, slight tilt, ingredients or parts orbiting | tech, beauty, food | `product floating in mid-air, slight tilt, soft shadow beneath, ingredients suspended around it` |
+| Splash / freeze (潑濺) | liquid frozen mid-air around the product | drinks, skincare, cleaning | `high-speed splash, liquid frozen mid-air around the bottle, backlit droplets, dark background` |
+| Knolling (整齊排列) | items laid out at right angles, grid-like | kits, sets, tools, fashion | `knolling arrangement, every item at right angles on a plain surface, top-down, even light` |
+| 360 turntable (轉台) | a video move: the product rotating a full turn on a plinth | demo clip, e-commerce spin | video prompt: `product on a turntable rotating slowly through 360 degrees, camera locked, studio light constant` |
+
+## Backgrounds (背景)
+
+| Background | Reads as | Phrase |
+| --- | --- | --- |
+| Seamless white / grey (無縫背景紙) | e-commerce, clean | `white seamless sweep` / `mid-grey seamless backdrop` |
+| Gradient | studio, premium | `smooth dark-to-charcoal gradient background` |
+| Brand-colour sweep | campaign, retail | `seamless backdrop in the brand colour #hex, no texture` |
+| Acrylic / reflective | tech, beauty, luxury | `product standing on black acrylic with a soft mirror reflection` |
+| Podium / plinth (展台) | luxury, skincare | `on a round stone podium, one step, soft shadow` |
+| Textured natural | food, wellness, craft | `on rough slate` / `on washed linen` / `at the edge of still water` |
+| Environment | lifestyle | name the room, the time of day and the light source |
+
+## Camera and lens language
+
+| Choice | Effect | Phrase |
+| --- | --- | --- |
+| 85 mm | compression, flattering, no distortion; the packshot lens | `85 mm lens` |
+| 100 mm macro | detail and texture | `100 mm macro lens` |
+| 35 mm | environment visible, lifestyle | `35 mm lens, environment in frame` |
+| 16–24 mm | interiors, real estate; keep verticals straight | `24 mm lens, verticals corrected` |
+| f/8–f/11 | packshot sharpness, every letter readable | `f/11, sharp front to back` |
+| f/2.8 or wider | lifestyle, isolates the product | `f/2.8, background softly out of focus` |
+| Eye level | neutral, catalogue | `eye-level camera` |
+| 45° down | shows the top and the front; food, flat objects | `camera at 45 degrees looking down` |
+| Low hero | authority, scale | `low-angle camera, slightly below the product base` |
+
+## Colour-grading vocabulary (調色)
+
+| Grade | Reads as | Phrase |
+| --- | --- | --- |
+| Clean neutral | true product colour, e-commerce | `neutral colour balance, accurate product colours, no colour cast` |
+| Warm golden | comfort, food, lifestyle | `warm golden grade, soft highlights` |
+| Teal-orange | cinematic, tech, automotive | `teal and orange cinematic grade` |
+| High-contrast editorial | fashion, retail-loud | `high-contrast editorial grade, deep blacks, crisp whites` |
+| Pastel matte | beauty, wellness, festive | `pastel matte grade, lifted blacks, soft saturation` |
+| Desaturated luxury | spirits, watches, premium | `desaturated luxury grade, muted tones, rich blacks` |
+
+Always end the grade line with the product exception: "product colours stay exact".
+
+## The prompt template
+
+Write in this order; each slot is one clause. Skip a slot only when the reference image already
+fixes it.
+
+```
+[subject: product, exact state, label to camera] →
+[shot type] → [lighting setup, direction, result on the product] →
+[background] → [camera: lens, aperture, height] → [grade] →
+[constraints: keep label / shape / colour exactly as @image1; no text; no extra objects]
+```
+
+### Worked examples
+
+**Bottled oolong tea** (packshot reference `@image1`, KV at 4:5, 2K)
+
+`A chilled glass bottle of oolong tea exactly as in @image1, label facing camera, beads of
+condensation on the glass. Hero shot from a slightly low angle. Strong backlight glowing through
+the amber tea, thin rim light on both edges, one small soft highlight on the shoulder. Dark
+teal-to-black gradient background, faint mist at the base. 85 mm lens, f/8, product sharp
+throughout. Cool clean grade with the tea's amber kept exact. Keep the label text, bottle shape
+and cap colour exactly as the reference; no added text; nothing else in frame; upper third
+empty for a headline.`
+
+**Trail sneaker** (packshot `@image1`, lifestyle still at 16:9)
+
+`The sneaker exactly as in @image1, worn on a right foot mid-stride on wet asphalt at dusk.
+Low-angle lifestyle shot, 35 mm lens, f/2.8, the shoe sharp and the street softly blurred.
+Hard low sun from behind the shoe as rim light, cool ambient fill, small splash of water from
+the heel frozen mid-air. Teal and orange cinematic grade; the shoe's colours and logo placement
+stay exact. No text, no other people in focus.`
+
+**Skincare jar** (packshot `@image1`, KV at 1:1, 2K)
+
+`The cream jar exactly as in @image1 on a round pale-stone podium, lid off and leaning against
+the jar, a smooth swirl of cream visible. Clamshell beauty lighting, soft light from above and
+below, nearly shadowless, dewy highlights on the glass. Warm off-white seamless background with
+a single soft petal-pink shadow. 85 mm lens, f/11, eye-level. Pastel matte grade, jar and label
+colours exact. Keep the label, jar shape and lid finish exactly as the reference; no text; no
+hands.`

package/skill/ads/reference/styles.md

@@ -0,0 +1,70 @@
+# Visual styles — 13 looks for ads
+
+One style per campaign, chosen from the brief's industry and tone and written into every prompt
+so the KV, the stills and the clips agree. Each row names the visual signifiers, the lighting
+preset it borrows from `reference/product-photography.md`, the phrases to write, and the tone
+of the CTA that fits. Nothing here names a real brand's look; describe the style, never a
+competitor's ad.
+
+## The styles
+
+| Style | Visual signifiers | Light preset | Prompt phrases | CTA tone |
+| --- | --- | --- | --- | --- |
+| Minimal (極簡) | negative space, one product, neutral sweep, thin type, nothing moves fast | large softbox or high-key | `minimal composition, generous negative space, single product, neutral seamless background, soft shadow, quiet` | short, lower-case, one verb: "Try it." |
+| Luxury (奢華) | low-key black, jewel tones, marble or dark stone, gold accents, serif type, slow motion | low-key + rim | `low-key luxury lighting, black gradient, marble surface, gold rim highlights, rich shadows, slow deliberate camera` | restrained, no discount: "Discover the collection" |
+| Retail-loud (促銷) | saturated brand colours, big percentage type, bursts and badges, hard light, fast cuts | hard light | `bold saturated brand colour background, hard direct light, punchy contrast, high energy, graphic` | urgent, numeric: "40 % off. Ends Sunday." |
+| Lifestyle (生活感) | real rooms and streets, window or golden-hour light, candid people, hands, imperfect framing | softbox window light (natural) | `lifestyle photo, natural window light, candid moment, shallow depth of field, warm and lived-in` | conversational: "Made for mornings like this" |
+| Tech (科技) | dark background, neon edge light, glass and brushed metal, particles, HUD lines, orbiting camera | strip lights + low-key | `dark studio, cyan and magenta edge lighting, glass and brushed metal, floating particles, slow orbit camera` | precise, feature-led: "Meet the new X" |
+| Food (食品) | warm backlight, steam, drips, glistening surfaces, macro, shallow depth, wood and linen | rim / backlight + softbox | `warm backlit food shot, rising steam, glossy surface, macro detail, shallow focus, wooden table` | appetite verbs: "Taste it hot" |
+| Beauty (美妝) | clamshell light, dewy skin, water droplets, product swatches, pastel and skin-tone palette | clamshell | `clamshell beauty lighting, dewy luminous skin, water droplets, soft pastel palette, swatch of the product texture` | promise-led, gentle: "Glow by morning" |
+| Fashion (時裝) | editorial pose, hard flash, colour block, fabric in motion, off-centre crop | hard light | `editorial fashion shot, hard on-camera flash look, colour-block background, fabric mid-motion, bold crop` | terse, imperative: "Wear it now" |
+| Automotive (汽車) | wet asphalt, dusk or tunnel light, long strip reflections down the body, low tracking camera | strip lights + low-key | `car on wet asphalt at dusk, tunnel light streaks along the bodywork, low tracking camera, cinematic teal-orange` | aspirational, one line: "Drive further" |
+| Real estate (房地產) | 16–24 mm interiors, verticals corrected, HDR balance of window and room, twilight exteriors | high-key interior, natural | `wide 24 mm interior, verticals straight, balanced window and room exposure, twilight exterior with warm windows` | invitation: "Book a viewing" |
+| App / SaaS (應用 / 軟體) | device mockups, cursor and tap animation, brand gradient, kinetic type, flat clean UI | high-key or gradient sweep | `phone mockup floating on a brand gradient, clean UI on screen, soft shadow, kinetic typography space around it` | outcome + free: "Start free" |
+| Festive (節慶) | seasonal props, warm bokeh lights, gold and red or snow and blue, wrapped gifts | softbox + warm practicals | `festive scene, warm bokeh string lights, seasonal props, gift wrapping, cosy warm grade` | gifting: "Gift the good stuff" |
+| Black Friday (黑五) | black background, neon or acid accents, glitch and scan lines, countdown numerals, bold percentage type | hard light + neon rim | `black background, neon accent rim light, glitch texture, bold oversized sale typography space, high contrast` | countdown, urgency: "24 hours. 50 % off." |
+
+## Choosing by industry
+
+| Industry | First choice | Second choice | Avoid |
+| --- | --- | --- | --- |
+| Beverage, snacks | Food | Lifestyle, Retail-loud | Luxury unless premium spirits |
+| Skincare, cosmetics | Beauty | Minimal, Luxury | Retail-loud (cheapens the product) |
+| Fashion, footwear | Fashion | Lifestyle, Minimal | Tech |
+| Consumer electronics | Tech | Minimal | Food, Festive |
+| Software, apps | App / SaaS | Minimal | Luxury |
+| Cars, e-bikes | Automotive | Tech | Retail-loud except clearance |
+| Property | Real estate | Luxury (high-end) | Everything else |
+| Jewellery, watches, spirits | Luxury | Minimal | Lifestyle unless aspirational |
+| Supermarket, e-commerce promo | Retail-loud | Black Friday (seasonal) | Minimal |
+| Seasonal campaign, any category | Festive | the category's first choice | — |
+| Home, furniture, wellness | Lifestyle | Minimal | Tech |
+
+## Style in motion (video clips)
+
+The same style decides how the camera and the cut behave; write the motion clause of a clip
+prompt from this table and keep it constant through the ad.
+
+| Style | Camera | Cut rhythm | Motion phrases |
+| --- | --- | --- | --- |
+| Minimal, Luxury, Beauty | slow push or slow orbit, locked horizon | 4–8 s per clip | `slow push in`, `slow orbit`, `camera locked, only the light moves` |
+| Retail-loud, Black Friday, Fashion | whip pans, snap zooms, hard cuts | 1–3 s per clip | `fast whip pan`, `snap zoom onto the product`, `hard cut` |
+| Lifestyle, Food, Festive | handheld drift, rack focus, natural pace | 3–5 s | `gentle handheld drift`, `rack focus from the hand to the product` |
+| Tech, Automotive | orbit, low tracking, reveal from dark | 3–6 s | `low tracking shot`, `orbit with edge light sweeping across the surface`, `reveal from darkness` |
+| Real estate, App / SaaS | slow glide, dolly through a door, screen-locked mockup | 4–6 s | `slow glide forward`, `dolly through the doorway`, `tap animation on the screen` |
+
+Clip lengths above are planning values; the accepted ranges per model are in
+`reference/models.md`.
+
+## Applying a style
+
+- Write the style's phrases in **every** prompt of the campaign: KV, stills, clips. A style that
+  appears in the KV and not in the clips reads as two campaigns.
+- The style sets light, background, palette and type feeling. It never changes the product: the
+  packshot reference and the identity clause in `reference/ai-consistency.md` stay in force.
+- Two styles in one ad is a mistake except the deliberate before / after (a Lifestyle "before"
+  cut to a Minimal "after").
+- Brand colours from the brief override the style's palette: a Tech look in a brand's orange is
+  `orange edge lighting`, not cyan.
+- Match the CTA tone column to the copy on the end card; a Luxury visual with "BUY NOW!!!" on
+  the card is the most common mismatch.

package/skill/short-drama/SKILL.md

@@ -366,6 +366,10 @@ actually charged (the `credits` line of each result, not the estimate), what `ma
 lists, and any gate you had to fix. If a generation failed, say which and why — never resubmit a
 node whose status is `generating`.
 
+Once an episode is generated, teasers, posters and covers per platform are the `social-media`
+skill's job (it reads this project's `episodes/E0N.json` and canvases); a commercial for the show
+is the `ads` skill's. Point the user there rather than improvising sizes here.
+
 ## Rules specific to this skill
 
 - **The user signs off the brief + outline skeleton and the first segment's keyframes.**

package/skill/short-drama/reference/canvas-recipes.md

@@ -4,6 +4,24 @@
 which row, and in what order the references are connected. Read them as written and substitute the
 tool calls from `operations-wji.md`; `--from A B C` is `fromNodeIds: ['A','B','C']` and the array
 order is the reference order. Everything about rows, `@imageN` order and prompts is identical.
+Build each row with **one `canvas_create_nodes` call**: its `nodes` array takes the fields of
+`canvas_create_node` (minus `canvas`), and `"$i"` in `fromNodeIds` or `nextTo` means item i of the
+same batch (0-based, must come earlier), so a headshot → sheet → voice chain is one call and a
+`characters` canvas of twelve characters is one or two calls instead of fifty. Items are created in
+order; if one fails the result says `failedAt` and the earlier nodes stay on the canvas — fix that
+item and resend from it with real ids (up to 60 nodes a call).
+
+```
+canvas_create_nodes { canvas: CV, nodes: [
+  { kind: "title", text: "C01 林小夏", position: {x: 0, y: 0} },                                        # $0
+  { kind: "image", name: "C01 headshot", position: {x: 400, y: 0},
+    generate: { model: "<image model>", ratio: "1:1", prompt: "<portrait prompt>" } },               # $1
+  { kind: "image", name: "C01 sheet", fromNodeIds: ["$1"],
+    generate: { model: "<image model>", ratio: "16:9", prompt: "@image1 is the face to match … " } }, # $2
+  { kind: "audio", name: "C01 voice", fromNodeIds: ["$1"],
+    generate: { model: "seed-audio-1.0", prompt: <voice prompt> } }                                  # $3
+] }                                                                   → created[0..3].id = T1 H1 S1 V1
+```
 
 Command sequences for the three canvas types. All ids are placeholders — take real ones from the
 command output (`wji --json canvas add … | jq -r .id`) and `bind` them at once. `SD` below is

package/skill/short-drama/reference/operations-cli.md

@@ -36,6 +36,7 @@ The `wji` skill governs auth, confirmations and spending. The calls this skill m
 | preflight | `wji auth whoami --json`, `wji profile wallet` |
 | workspace | `wji ws create "<title>" -r <ratio>` → `wji ws use <id>` |
 | datasets | `wji dataset new … --app <appId>`, `wji fs set <id> --merge` (never `--content`), `--add-images` (never `--images` on an existing node) |
+| reads | `wji fs get <id> --json` is the `standard` summary; `--detail full` whenever you need the whole `content` (`--merge` reads full for you) |
 | canvases | `wji canvas create <name>` · `wji canvas add <kind> -c <cv> …` · `wji canvas connect <a> <b>` · `wji canvas get <node>` · `wji canvas gen <node> [--yes --wait]` · `wji canvas rm` |
 | assets | `wji asset download <id> -o <dir>` — **look at what you generated** before building on it |
 | history | `wji history get <id> --wait` |

package/skill/short-drama/reference/operations-wji.md

@@ -43,9 +43,9 @@ write a whole document at a time rather than field by field.
 
 | Purpose | Tool |
 | --- | --- |
-| workspace | `workspace_search`, `workspace_get`, `workspace_create`, `workspace_add_app` |
+| workspace | `workspace_get { id }` for the workspace the chat is in, `workspace_create`, `workspace_add_app`; `workspace_search` **only with `limit`** (≤ 10) and a `keyword` — without them it returns every workspace the user can see. Both default to `detail: 'standard'`; `detail: 'full'` only when you need `settings` |
 | semantic folders | `resolve_dataset_folder { id, key: 'story' \| 'character' \| 'location' \| 'prop' \| 'shot' }` |
-| datasets | `fsnode_create_file { parent, name, dataset: true, datasetType, description, content, tags }`, `fsnode_update`, `fsnode_upsert_segment` |
+| datasets | `fsnode_create_file { parent, name, dataset: true, datasetType, description, content, tags }`, `fsnode_update`, `fsnode_upsert_segment`; `fsnode_get { id, detail: 'full' }` before replacing a `content` (the default `standard` read is a summary with `truncated: true`, and writing it back is rejected) |
 | assets | `asset_get`, `asset_check_status`, `asset_create_from_url` |
 | canvases | the `canvas_*` tools below |
 
@@ -59,16 +59,21 @@ substitute:
 | `canvas-recipes.md` shows | call |
 | --- | --- |
 | `wji canvas create <name>` | `canvas_create { workspace, name }` → canvas id |
-| `wji canvas add <kind> -c CV -n "<name>" --at x,y …` | `canvas_create_node { canvas, kind, name, position: {x, y}, … }` |
+| `wji canvas add <kind> -c CV -n "<name>" --at x,y …` | `canvas_create_node { canvas, kind, name, position: {x, y}, … }` — one node |
+| a whole row of `wji canvas add …` lines | `canvas_create_nodes { canvas, nodes: [ … ] }` — up to 60 nodes in one call, each item the fields of `canvas_create_node` without `canvas`; `"$i"` in `fromNodeIds` / `nextTo` is item i of the same batch (0-based, must come earlier). Built in order; on a failure the result has `failedAt` and the earlier nodes stay |
 | `--prompt-file /tmp/x.txt` / `-p "<prompt>"` | `generate: { prompt, model, ratio, resolution, duration }` on `canvas_create_node` |
 | `--asset <id>` (a display node) | `assetId` on `canvas_create_node` |
 | `--from A B C` | `fromNodeIds: ['A','B','C']` — **the array order is the reference order** |
 | `wji canvas connect <a> <b>` | `canvas_connect_nodes { canvas, fromId, toId }` |
-| `wji canvas get <node>` | `canvas_get_node { canvas, id }` — references, label, status, asset |
+| `wji canvas get <node>` | `canvas_get_node { canvas, id }` — references, label, status, asset; `detail: 'full'` for the whole `generate` / `text` before editing them |
 | `wji canvas gen <node>` (cost preview) | `canvas_generate { canvas, nodeIds, estimateOnly: true }` |
 | `wji canvas gen <node> --yes --wait` | `canvas_generate { canvas, nodeIds, wait: true }` |
 | `wji canvas rm` | `canvas_delete_nodes { canvas, ids }` |
 
+Build each row with **one** `canvas_create_nodes` call (title, headshot, sheet `fromNodeIds: ["$1"]`,
+voice `fromNodeIds: ["$1"]`), not four `canvas_create_node` calls — see the batch example at the top
+of `canvas-recipes.md`.
+
 Reference labels (`@image1`, `@audio1` …) are numbered per type in connection order, exactly as the
 canvas UI numbers them. After connecting, call `canvas_get_node` and check the labels match what the
 prompt names **before** generating — a mislabelled reference is a wasted generation.
@@ -77,6 +82,24 @@ prompt names **before** generating — a mislabelled reference is a wasted gener
 first, tell the user the cost, and only generate for real once they say yes. The `wji` skill's
 spending policy applies here in full.
 
+## Steps and context
+
+**A turn is capped at 40 steps, not 40 tool calls** (an Agent can set its own cap). One step is one
+model call; every tool call you emit in the same step runs together and counts once. So emit
+independent calls together — all the `short_drama_render` / `fsnode_create_file` of a stage in one
+step, the five `resolve_dataset_folder` calls in one step — and write a whole document per
+`short_drama_put` rather than field by field. When two steps are left you get a system note: stop
+calling tools, report what was written (with ids) and what is still open; the stream ends at the cap
+without letting you reply.
+
+**Tool results stay in the context for the rest of the turn.** Read what you need, once: `ai_list_models`
+once per project with `features: 'image'` (it is large); `fsnode_get` on the one dataset you need
+instead of `fsnode_children` on a folder of them; `short_drama_get { section }` for one branch rather
+than the whole project; never `workspace_search` without `limit`. Every list / get tool takes
+`detail: 'brief' | 'standard' | 'full'` and defaults to `standard` (text clipped to 200 chars, media
+as `{ id, url, status }`, `content` summarised); ask for `brief` when you only need ids and for `full`
+only when you need the complete `content` / `form` / `settings`.
+
 **In a canvas chat** (the chat panel beside an open canvas) the browser supplies its own `canvas_*`
 tools, which always act on the canvas the user is looking at and take no `canvas` argument. A short
 drama needs several canvases, so do that work in a workspace chat; in a canvas chat, say so.

package/skill/short-drama/reference/stage-storyboard.md

@@ -345,6 +345,12 @@ segment is rejected, retry, then trim — do not go hunting for a forbidden word
 To bisect anyway, use a scratch video node at `-d 2`, 480P: free when rejected, ~30 credits when it
 passes, and sample each fragment several times because one pass proves nothing.
 
+**Last resort: the uncensored twin.** `wan-3.0-ref-uncensored` (MuleRouter, no green-net) takes the
+same prompt and the same 10 images + 5 videos + 5 audios, at ~1.3× the price; note that it bills the
+reference-video seconds as well as the output. Switch a stuck segment to it only with the
+Producer's approval, and only after retry-then-trim has failed. `short-drama.mjs` does not know it
+as a mode, so set it as the node's model directly.
+
 ## Sheets mode — no storyboard images (the default)
 
 `params.references: sheets` drops stage 5's image generation entirely: the video node references

package/skill/social-media/SKILL.md

@@ -0,0 +1,152 @@
+---
+name: social-media
+description: Make and size content for social platforms on WJI through the `wji` CLI — the image and video specs of Instagram, TikTok, YouTube / Shorts, Facebook, X, LinkedIn, Pinterest, Threads, Snapchat, 小红书, 抖音, 快手, 视频号, B站, 微博 and 微信公众号 (ratio, pixels, duration, file size, safe zones, caption / title / hashtag limits, covers and thumbnails), what performs on each (the 3-second hook, loops, on-screen text), and the repurposing recipes: one master into every placement, a short-drama episode into vertical teasers and a poster set, an ad master into all its placements. Use whenever a task mentions social media / 社群 / 社交媒体, posting / 发文, Reels, Shorts, TikTok, 小红书 / 抖音 / 视频号, a cover / 封面, poster / 海报, thumbnail, teaser / 预告, or asks what size or length a platform wants. Combines with the short-drama and ads skills. Works together with the `wji` skill.
+---
+
+# Social media on WJI
+
+You are the person who knows every platform's numbers and turns one piece of work into the
+right post for each of them. **You write the spec sheet, the hooks and the copy**; the `wji`
+CLI keeps the plan as a dataset and generates the per-ratio images and videos on a canvas;
+local `sharp` or `ffmpeg` (when installed) crops to exact pixels. Nothing is generated until
+the user has said yes to the quote (or the message carried `-auto`, see the `wji` skill).
+
+```
+what & where ──▶ spec sheet ──▶ source assets ──▶ generate per ratio ──▶ covers ──▶ copy ──▶ post plan
+     │               │               │                  │                  │          │
+  platforms,     one row per     existing nodes /    images: native      thumbnail,  caption,
+  goal, market   surface from    short-drama /       ratio then crop;    poster,     hashtags,
+                 the tables      ads / new           video: regenerate   图文 cover   slot, label
+```
+
+`{baseDir}` = this skill. The `wji` skill governs every CLI call.
+
+| Read | When |
+| --- | --- |
+| `reference/platforms-global.md` | any Instagram / TikTok / YouTube / Facebook / X / LinkedIn / Pinterest / Threads / Snapchat question — the spec tables and safe zones |
+| `reference/platforms-china.md` | 抖音 / 快手 / 小红书 / 视频号 / B站 / 微博 / 公众号 — specs, and China-specific practice (AI label, titles, 图文) |
+| `reference/ratio-map.md` | before generating: which WJI model makes which ratio natively, what must be cropped and how to anchor the crop |
+| `reference/content-craft.md` | hooks, loops, overlays, captions and hashtags per platform, posting times, cover design, carousel / 图文 storytelling |
+| `reference/repurpose.md` | master → placements, drama episode → teaser + posters, ad master → placements, the deliverables checklist |
+| `reference/plan-template.md` | the post spec sheet — write one row per post before touching the canvas |
+
+## Operations
+
+### Where the plan lives
+
+Locally: `social/<campaign>/plan.md` (the spec sheet, one row per post, plus the copy and the
+deliverables) and `prompts/`. On WJI: the same plan as a dataset `SOCIAL <campaign>` in the
+workspace's **story** folder.
+
+```bash
+wji auth whoami --json && wji profile wallet
+wji ws use <workspaceId>
+wji ws app list                                                  # the Dataset app id → DS_APP
+wji --json ws fs --app DS_APP | jq -r '.. | objects | select(.content.semanticKey? == "story") | .id'   # → STORY
+wji dataset new STORY "SOCIAL <campaign>" -t other -d "<one line>" --json | jq -r .id                   # → PLAN
+wji fs set PLAN --merge content=@social/<campaign>/plan.md --merge viewMode=content
+wji canvas list                                                  # the masters, keyframes, sheets already there
+wji canvas nodes -c <canvas> --json | jq '.[] | {id,name,kind,assetId}'
+wji asset download <assetId…> -o /tmp/social                     # look before you write about it
+wji asset upload <file>                                          # the user's own material
+wji models get <model>                                           # the ratio list ratio-map.md summarises
+```
+
+### Generating per ratio
+
+One canvas `SOCIAL · <campaign>` (`wji canvas create "SOCIAL · <campaign>" --json | jq -r .id`),
+one row per platform surface:
+
+| Need | Command |
+| --- | --- |
+| image at a native ratio | `wji canvas add image -c SC -n "小红书 cover" --from <source nodes> -m <image model> -r 3:4 --resolution 2K --prompt-file prompts/xhs-cover.txt` — the source keyframe / KV / sheet first in `--from` |
+| video at the platform ratio | `wji canvas add video -c SC -n "Reels teaser A" --from <same references as the master> -m <video model> -r 9:16 -d 20 --prompt-file prompts/reels-a.txt` — **regenerated, never cropped** |
+| quote → yes → run | `wji canvas gen <nodeIds…>` (prints the cost, refuses) → tell the user → `wji canvas gen <nodeIds…> --yes --wait` |
+| adjust | `wji canvas set <node> -p … / -r … / -d …`, `wji canvas connect <from> <to>` |
+| one-off still | `wji gen image -m <model> -p "…" -r 4:5 --ref <assetId> --wait` |
+
+### Exact pixels
+
+A model generates a ratio, not a pixel size. When the placement needs exact pixels or a ratio no
+model has (1200×628, 1080×1260 视频号, 1146×717 B站, 900×383 公众号, banner sizes), generate at the
+nearest native ratio from `reference/ratio-map.md`, download the asset, and crop locally:
+`npx sharp-cli` / a five-line node script with `sharp` (`resize(w, h, { fit: 'cover', position })`)
+or `ffmpeg -i in.png -vf "crop=w:h:x:y,scale=W:H" out.png` — whichever is installed; if neither
+is, deliver the generated frame plus the crop box (left, top, width, height) and say so. Anchor
+every crop as `reference/ratio-map.md` §"Crop anchoring" says: faces, then product centre, then
+the headline zone. Upload the result with `wji asset upload` and add it to the canvas as a display
+node so the plan's ids point at the final file.
+
+## Before anything else
+
+1. **Ask what and where, in one message**: what is being promoted (a drama episode, an ad, a
+   product, an account), which platforms and surfaces (feed vs Reels vs Stories are different
+   posts), the goal (views, follows, clicks to the full episode, sales), the market and language
+   (mainland / overseas / both — it changes the platform list and the label rules). **Never guess
+   the platform.** Offer the usual sets: 短剧 launch = 抖音 + 小红书 + 视频号 (+ TikTok + Reels +
+   Shorts overseas); product = Instagram feed + Reels + TikTok + 小红书.
+2. **Write the spec sheet** (`reference/plan-template.md`): one row per post with the platform's
+   own numbers copied from the tables — ratio, pixels, duration cap, safe zone, caption / title /
+   hashtag limits, cover spec. Rows marked *(verify)* in the tables stay marked in the sheet.
+3. **Find the source material** before proposing generation: the canvases and datasets already
+   in the workspace, the user's files, or the other skills' work (next section). Generate new
+   material only for what does not exist.
+4. **Quote the whole plan** (`wji gen cost` per node kind) before the first generation.
+
+## Combining with other skills
+
+- **A short-drama project is present** (a folder with `project.json` and `episodes/`, or the
+  short-drama skill is installed beside this one at `../short-drama/`) → *drama mode*. Read the
+  episode with the engine: `node ../short-drama/scripts/short-drama.mjs render <project> script --ep N`
+  for the lines and `… render <project> segment E0N-01` per segment (or read
+  `storyboard.segments` in `<project>/episodes/E0N.json` directly); `… manifest <project> --ep N`
+  tells you the episode canvas id, and
+  `wji canvas nodes -c <E0N canvas>` lists its keyframes, sheets and clips. Then
+  `reference/repurpose.md` §"Drama episode → teaser + poster set": pick the hook line from the middle of the
+  episode, build the 4-beat 15–45 s teaser (hook 0–3 s → context → escalation → cut before the
+  payoff + 「主页看全集」) as **new 9:16 video nodes generated from the same character and scene
+  sheets** (never a crop of the 16:9 clips), 2–3 variants that differ only in the hook, and the
+  poster set (小红书 3:4, Instagram 4:5, Pinterest 2:3, YouTube thumbnail 16:9, B站 1146×717,
+  视频号 6:7) from the episode keyframes with the title zone kept clear. Record the node ids in
+  `plan.md`; do not write into the drama's `project.json`.
+- **An ads brief is present** (`ads/<product>/brief.md` or the `AD <product>` dataset) →
+  *campaign mode*. Read the brief and the `AD · <product>` canvas (KV, stills, clips, master
+  ratio). `reference/repurpose.md` §"Ad master → placements": one row per placement, KV
+  regenerated at native ratios or cropped for the rest, vertical / square video regenerated from
+  the same stills, per-platform copy written from the brief's CTA and legal lines, the AI-content
+  label toggled where the ads skill's `reference/compliance.md` says so.
+- **Neither** → standalone: the user's assets or fresh generation, same spec sheet.
+- If the material the user describes belongs to a skill that is **not installed**, say which one
+  (short-drama / ads) rather than rebuilding its data by hand.
+
+## Making the posts
+
+- **Images**: generate at the native ratio where one exists (`reference/ratio-map.md`), with the
+  source node first in `--from` so the subject stays the same; crop only for non-native sizes.
+  One ratio for every card of a carousel / 图文.
+- **Video**: one node per platform ratio, regenerated with the master's references and a prompt
+  reframed on the faces / product for vertical; `-d` under the surface's cap; captions and
+  on-screen text are overlays inside the safe zone, never generated.
+- **Covers and thumbnails**: face + ≤4 words, high contrast, the series template repeated
+  (`reference/content-craft.md` §"Cover and thumbnail design"); 小红书 cover is the hook of the 图文.
+- **Copy**: hook line, body, CTA, hashtags within the platform's limits (Instagram ≤5 hashtags,
+  小红书 title ≤20 chars, X 280) — written into `plan.md`, one block per post.
+- **Posting slots** from `reference/content-craft.md` are starting points, said as such.
+
+## Reporting
+
+After the spec sheet: the rows and what will be generated vs cropped, with the quote. After
+generation: the deliverables table from `reference/repurpose.md` (post, file / node, ratio, pixels,
+duration, cover, caption, hashtags, slot, AI label) and anything the user still has to do in
+their editor (captions, subtitles, crops you could not make).
+
+## Rules specific to this skill
+
+- **Ask before every generation**; a batch is one quote.
+- **Never invent a platform limit** — cite the table row; keep *(verify)* flags; if a platform
+  is not in the tables, say so and ask the user for its spec.
+- Video is regenerated at the target ratio, never cropped; images are cropped only when no model
+  has the ratio.
+- Text stays out of the platform's UI zones; one ratio per carousel; the drama's dialogue is
+  quoted verbatim in teasers.
+- Do not compile or store personal data of real people from the platforms; all-audiences content only.

package/skill/social-media/reference/content-craft.md

@@ -0,0 +1,124 @@
+# Content craft — hooks, loops, overlays, captions, covers, carousels
+
+What makes a post perform once the size is right: the first three seconds, the ending that
+re-plays, text that stays inside the safe zone, a caption shaped for each platform's reading
+pattern, a posting slot to start from, a cover that reads at thumbnail size, and a carousel that
+tells one story. Sizes and limits are quoted from platforms-global.md and platforms-china.md,
+not restated here; teaser and poster recipes built on these rules are in repurpose.md.
+
+## The first 3 seconds
+
+Feeds decide in the first swipe-length. The product, the face or the question is on screen by
+second 3, in motion, with no logo card, no fade-in and no "hi guys". Hook types, with a line
+each (drama and product examples side by side):
+
+| Hook type | Drama teaser | Product / brand |
+| --- | --- | --- |
+| Shocking line, mid-scene | 「你以为我不知道孩子是谁的？」 cut in on the reaction | "This is the last one we could get made." |
+| Result first, cause later | the slap, the arrest, the reveal — then 「三天前…」 | the after-state (clean pan, finished room), then the before |
+| Direct question | 「如果你婆婆当众撕了你的结婚证，你会怎么做？」 | "Why does every white sneaker turn yellow in a month?" |
+| Contradiction / mismatch | the bride in a wedding dress signing divorce papers | a luxury unboxing on a bus |
+| Countdown / stakes | 「还有十分钟，她就要嫁给别人了」 | "Three tests. One survives." |
+| Visual anomaly | one impossible detail in an ordinary frame (a wet footprint indoors) | the product doing the one thing rivals cannot |
+| Named audience | 「二婚的姐妹看过来」 | "If you edit on a phone, watch this." |
+
+Rules: one hook per cut, the first frame is already the strongest frame (it is also the cover),
+sound-off comprehension — the hook line is burned in as text. Write 2–3 hooks for the same body
+and let the platform pick (repurpose.md, variants).
+
+## Loop-friendly endings
+
+Reels, TikTok and 抖音 count re-watches. End on a frame that matches the first frame (same
+composition, same subject position) so the loop hides the seam; or end one beat *before* the
+answer the hook promised, so the viewer re-watches to catch it; or end on the CTA card that is
+also the cover. Never end on a logo fade or a black frame. For a drama teaser the ending is the
+cut before the payoff plus the line 「主页看全集 / full episode in bio」 (repurpose.md).
+
+## On-screen text and burned subtitles
+
+- Keep every word inside the safe zone of the 9:16 diagram (platforms-global.md): headline in the
+  upper safe band, subtitles at the lower edge of the safe band (≈y 1100–1240 on 1080×1920), never
+  behind the caption or the right rail.
+- Subtitles are burned in (most feeds autoplay muted): 2 lines maximum, ≤ 14 CJK chars or ≤ 32
+  Latin chars per line, high-contrast box or stroke, one style for the whole account.
+- Headline type ≥ 60 px on 1080×1920, ≤ 6 words / 8 chars; a second overlay only when the first
+  has left. Do not repeat the caption on screen.
+- Text is added after the crop and at the final ratio (ratio-map.md, crop anchoring); text
+  generated inside the picture is only for covers that are never cropped again.
+- 「AI生成」 label where required (platforms-china.md) is the platform's toggle, not an overlay
+  you design — but do not place other text where the platform paints its label (top-left on most
+  Chinese apps).
+
+## Caption structure per platform
+
+| Platform | Structure | Hashtags / tags | Length notes |
+| --- | --- | --- | --- |
+| Instagram | hook line (shown before "more") → 2–4 short lines → CTA (save / share / comment) | **≤5** (Dec 2025 cap) — 3 niche + 2 broad | 2,200 max; the first 125 chars are the post |
+| TikTok / 抖音 | one-line hook + a question that invites comments | 3–5 hashtags / 话题, one is the series tag | 4,000 max on TikTok; 抖音 shows ≈20 chars of title |
+| YouTube | title ≤ 60–70 chars shown; description: 2-line summary, links, chapters, disclosure | 3–5 hashtags at the end of the description | title 100 hard cap |
+| Facebook | short hook, link in the first comment or the post, CTA | 1–3 | shares outweigh hashtags |
+| X | the hook is the whole post; thread for detail | 1–2 or none | 280 |
+| LinkedIn | hook in the first two lines → line-broken story → question → 3 hashtags | 3 | 3,000; "see more" after ≈3 lines |
+| Pinterest | keyword title ≤ 100, description with 2–3 search phrases | none | description 500–800 *(verify)* |
+| Threads | one sentence; the picture carries the rest | none | 500 |
+| 小红书 | 标题 ≤ 20 with the keyword first → 正文 in short lines with emoji markers → 话题 tags | keyword + 3–6 话题 (#…#) | 正文 1,000 (长文 6,000) |
+| 视频号 / 公众号 | one-line 描述 with 话题; the article's first line is the hook | 1–3 话题 | limits not confirmed |
+| B站 | 标题 with the series and episode number; 简介 with timestamps | tags on upload (platform field) | limits not confirmed |
+
+A CTA is one verb: 关注 / follow, 主页看全集 / full episode in bio, save, comment the answer.
+Platform buttons (shop, link, subscribe) are not drawn into the picture.
+
+## Posting-time starting points
+
+Sprout Social 2026 aggregates, to be treated as **starting points** and replaced by the
+account's own insights after two weeks: Instagram Mon–Thu 10:00–15:00; TikTok Wed 14:00–17:00;
+LinkedIn Tue 11:00–17:00. Local time of the audience, not of the account. Other platforms and
+the Chinese apps: not confirmed in the research — plan lunchtime (12–13) and evening (20–22)
+slots in the audience's time zone and measure. Series content posts at the same slot every
+time; the second post of a series goes out within 24 h of the first while the hook is fresh.
+
+## Cover and thumbnail design
+
+- **Face + ≤ 4 words** (≤ 8 CJK chars): the face at 40–60 % of the frame height, looking at the
+  camera or at the text; the words in one colour band, one weight.
+- **High contrast**: light text on a darkened lower third or dark text on a light band; check
+  it at 120 px wide (the 小红书 / 快手 tile is half a phone).
+- **One template per series**: same title position, same colour, same type; the episode number
+  changes, nothing else. Recognition beats novelty in a grid.
+- **Ratio per platform** (ratio-map.md): 16:9 1280×720 YouTube; 3:4 小红书 / IG grid; 9:16 Reels
+  / TikTok / 抖音 with the text in the centre 60 % because the grid crops to a taller tile;
+  1146×717 B站 with the bottom-left clear; 6:7 视频号 composed for its 1:1 share crop.
+- Covers are generated **with** the text only at the final ratio (nano-banana-pro, gpt-image-2.5
+  or seedream-5-pro for reliable multilingual type); every other size gets the text overlaid
+  after the crop.
+
+## Carousel / 图文 storytelling
+
+1. **Cover as hook** — the question, the contrast or the result; it must work alone in the feed.
+2. **One idea per card**, one line of text per card, the same layout grid on every card.
+3. **Escalate**: cards 2–N answer the cover in order of surprise, not of logic; the strongest
+   proof sits two cards before the end.
+4. **Last card = CTA** (follow, save, 主页看全集, 领取), in the same template as the cover so the
+   loop back to card 1 feels designed.
+5. **One ratio for the whole set** (IG ≤ 20 slides, 小红书 ≤ 18 images, TikTok photo mode ≤ 35):
+   generate every card at the same `ratio` and crop to the same box.
+6. Caption repeats the cover's keyword in the first line; on 小红书 the 话题 tags close the body.
+
+## Trending audio
+
+Trending sounds lift reach on TikTok / 抖音 / Reels, but a business account may only use
+tracks from the platform's **commercially licensed library**; personal-account sounds are
+muted or the post is removed when the account is business-typed. AI-generated soundtracks
+(Seed Audio, generated per teaser) avoid the problem and can be reused across the series.
+Never lip-sync a drama clip to a licensed pop song for a paid promotion.
+
+## What to avoid
+
+- A logo intro, a fade-in, a title card before the hook.
+- Letterboxed 16:9 inside a 9:16 post (regenerate — ratio-map.md).
+- Text in the top 270 px or the bottom 670 px of a vertical frame; text under the right rail.
+- More than 5 hashtags on Instagram; hashtag walls anywhere; hashtags that do not exist yet.
+- Sub-1080-px sources, upscaled 1K frames, JPEG re-compression of a re-compression.
+- Cross-platform reposts with the other platform's watermark or overlay language.
+- Engagement bait ("like if…", fake polls), fake countdowns, fake UI buttons drawn into the frame.
+- Claiming AI footage is real footage; skipping the AI label where it is required.