@wji.ai/cli 0.1.0 → 0.2.0

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

@@ -0,0 +1,206 @@
+# Stage 2 — Characters
+
+Turn every `characters[]` entry into a reusable asset: a profile people read, an English portrait
+prompt, a 16:9 character-sheet prompt, and a voice prompt — then the three assets WJI needs for a
+character to hold across shots: **headshot** (icon), **sheet**, **voice** (`primaryAudio`).
+
+## Profile (drama language)
+
+`oneLiner`, `profile.gender/ageRange/identity/appearance/personality(3–5 words)/temperament/
+motivation/relationships/evidence`. Inferred facts are fine when marked (`（推断）`); `evidence`
+holds verbatim quotes only, or stays empty. Keep the tier from the outline — if it looks wrong,
+change the outline, not this stage.
+
+### `identityEn` — the name-free handle
+
+One short English phrase, ≤ 80 characters, distinct per character: `the woman in her mid thirties in
+headphones`, `the stocky man in his sixties with a grey crew cut`. Prompts may never carry a name,
+but they constantly have to say *who* — who speaks this line, who turns — and this is the words
+they use. Pick something visible and permanent (age, build, the thing they always wear), not a
+plot role: the model can see a grey crew cut, it cannot see "the retired security guard". Gated:
+present, English, name-free, no two characters alike.
+
+## The bible — `bible` and `profile.relationships` (gates `bible`, `relationships`)
+
+A profile says what they look like; the bible says how they got here, how they act now and what
+the story will make them pay. Method in `craft-story.md` §7. By tier:
+
+- every character: `bible.function` — 主角 / 共同主角 / 反派 / 爱情线 / 导师 / 盟友 / 镜像 / 压力测试者 /
+  喜剧调节 (the dramatic job, independent of the tier);
+- `support` and `lead`: `hookLine` (身份 + 核心矛盾 + 外在欲望 + 风险), `want` and `need` (not the same
+  sentence), `misbelief`, `contradiction`, `pressureResponse`, `speech.rhythm` and `speech.silence`
+  (plus `catchphrase`, `forbiddenWords`, an optional `register` override);
+- `lead`: also `wound` (150–250 字, the choice and what they believed after), ≥ 2 `pastChoices`,
+  `firstImpression` → `finalUnderstanding`, exactly three `forbidden` and three `mustFace`, and an
+  `arc` of ≥ 3 stages with the 高潮选择 among them, each bound to an episode.
+
+`profile.relationships[]` becomes the 关系网: `{id, name, surface, real, power, secret, turnEpisode}`.
+Every lead has at least one with `surface` + `real` + `power`, and somewhere in the cast a
+relationship carries a `secret` — that is where the reversals come from.
+
+**`bible.speech` is the source of the voice.** `voice.pace` / `videoDirection` and every line's
+`delivery` are written from `rhythm`, `silence` and `forbiddenWords`; `seed script` prints them in
+the `seedNote` and `render dialogue` puts them beside each line for the polish pass.
+
+## Image prompts (English, always)
+
+- **No names, aliases, author or title anywhere** — image models paint what they remember of a
+  name. Describe the person.
+- **Ethnicity + era + region explicit** (`East Asian, Han Chinese features`, `contemporary 2020s
+  Chinese city`, `Republican-era river town`). Left out, the model defaults to a contemporary
+  Western face. Both are gated.
+- `image.prompt`: an expressive three-quarter bust on a neutral background, soft directional key
+  light, shallow depth of field. Realism comes from imperfection: uneven skin tone, visible pores,
+  asymmetry in eyelids and brows, stray hairs, fabric with weave and wear. Put the individual
+  description first and make it long — two similar-age characters written as "template + age +
+  jacket colour" come out as the same person (`distinct` gate, 75 % word overlap).
+- `image.negativePrompt`: with `style: photoreal` (live action — the right choice for a
+  drama shot like a TV series) ban *drawing*: `illustration, anime, cartoon, painting`, and
+  **never** `photorealistic`, `film grain` or `skin texture`. With `style: realistic` ban *fakeness* (plastic skin, doll face, dead
+  eyes, helmet hair, flat fabric) and **never** `photorealistic` / `3d render`; with `ghibli` you
+  **must** ban `photorealistic`. Both gated. Always add `logo, brand name, text` — a "store
+  uniform" comes back wearing a real chain's badge otherwise (seen in testing), and a real brand
+  on a character is a problem in every shot after it.
+- `image.sheet`: the model-sheet layout, written out in full every time:
+
+  ```
+  ┌──────────┬────────────────────────────┐
+  │  bust    │   front    side    back     │   ONE 16:9 canvas, pure white background
+  │ portrait ├────────────────────────────┤   LEFT ZONE ≈ 34 % width: head-and-shoulders bust,
+  │ (face    │  detail · detail · detail   │     both shoulders complete, flat bottom edge
+  │  ref)    │                            │   RIGHT-TOP: three full-body views, same height,
+  └──────────┴────────────────────────────┘     same head-to-body ratio, PROPORTIONS ARE CRITICAL
+                                               RIGHT-BOTTOM: 4–5 small detail studies;
+                                               the detail studies give way, not the figures
+  ```
+
+  Plus: `all views share the SAME face as the bust portrait`, `LIGHTING IN THE LEFT ZONE ONLY:
+  soft directional key…`, `LIGHTING IN THE RIGHT ZONES: flat even orthographic lighting`, the
+  style's render phrase (`STYLE_PRESETS[style].render` — `node scripts/short-drama.mjs` exports
+  it; copy it verbatim), `No text, no watermark`. Gated: `16:9`, render phrase, no names.
+
+## Voice
+
+Six local-language fields (`timbre` `pitch` `pace` `accent` `emotion` + `referenceHint` optional)
+for people, one English `voice.prompt` for a voice-design engine, and a **`voice.sample`** — the
+short acted line that becomes the character's voice asset.
+
+### `voice.sample` — the 3–5 s reference
+
+```json
+"sample": { "line": "這裡是海風電台。", "direction": "貼著麥克風，像只講給一個人聽" }
+```
+
+- **`line`**: one line in the drama's language, **3–5 s at `params.speechCharsPerSecond`**
+  (default 2.4 — the voice model speaks far slower than the 4.5 chars/s the script uses for
+  reading, so that is **8–12 Chinese characters**, not twenty). Gated. No quote marks: the
+  engine reads them.
+- **`direction`**: how *this line* is acted, in the drama's language.
+- Pick a line that shows the character's default register, not a plot moment: it is a timbre
+  reference, and every later clip inherits it.
+
+### The prompt the model actually receives
+
+`prompt <project> voice C01` composes it from the fields you already wrote — you do not write it
+by hand, and it is a **parenthetical instruction followed by the line**:
+
+```
+（三十五歲上下女；基隆地方電台深夜節目主持人；性格冷靜、自持、嘴硬、念舊；偏低的胸腔共鳴，帶一點沙；
+中低音域，起伏小；慢，句尾拖半拍；台灣口音國語，捲舌音輕；克制的溫柔，預設是平靜；
+此刻貼著麥克風，像只講給一個人聽）這裡是海風電台。
+```
+
+In order: age + gender, identity, `profile.personality`, then `voice.timbre` / `pitch` / `pace` /
+`accent` / `emotion`, then `此刻` + the sample's `direction`. `params.speechAccent` is appended
+only when `voice.accent` is empty — the drama-language wording wins. Quote marks are stripped,
+because inside the parenthetical they would be read aloud.
+
+**The parenthetical is obeyed, not spoken.** Measured on Seed Audio 1.0: the same 8-character line
+ran 3.16 s bare and 3.50 s behind a ~100-character description — speaking that description would
+have added some forty seconds. It came back slower and lower, as the description asked. So write
+the voice fields richly: personality, rhythm, where the voice sits, what it does at the end of a
+phrase. That detail is free and it is the only thing separating two characters of the same age.
+
+The English `voice.prompt` stays as the static voice identity (≤ 400 characters, a parameter
+string, **no quoted lines** — gated) for engines that take a description instead of a sample.
+
+### `voice.videoDirection` — the clause every clip carries
+
+One English clause, ≤ 160 characters, that goes **beside every line this character speaks in every
+video prompt**:
+
+```
+slow and close-miked, low and chest-resonant with a light rasp, falling at the end of every phrase,
+Taiwanese-accented Mandarin, restrained
+```
+
+It is `voice.prompt` compressed to what a video model can act on in one breath: pace, where the
+voice sits, what it does at the end of a phrase, accent, default attitude. Without it a video model
+reads every character the same way — evenly, brightly, like a newsreader — and the clip stops
+matching the voice asset the same bible produced. Gated in stage 2 (present, English, ≤ 160 chars,
+no names, no quoted lines) and again in stage 5: a segment prompt must carry the direction of every
+character who speaks in it. For `wan`, the drama-language `timbre` / `pace` / `accent` / `emotion`
+fields are used instead, since that prompt is Chinese.
+
+## On WJI
+
+Dataset per character in `character/`:
+
+```bash
+wji dataset new <character folder> "<name>" -t character -d "<oneLiner>"       # bind characters.C01.wji.datasetId
+node {baseDir}/scripts/short-drama.mjs render <project> character C01 > /tmp/C01.md
+node {baseDir}/scripts/short-drama.mjs prompt <project> voice C01 > /tmp/C01-voice.txt
+wji fs set <id> --merge content=@/tmp/C01.md --merge image_prompt=@/tmp/C01-sheet.txt \
+  --merge audio_prompt=@/tmp/C01-voice.txt --merge metadata.shortDrama=@/tmp/C01.json
+```
+
+`characters` canvas, one row per character (see `canvas-recipes.md` §1):
+
+| Node | Model | Ratio | Prompt |
+| --- | --- | --- | --- |
+| headshot (image, generate) | GPT Image 2.5 Sunburst (or the best image model listed) | 1:1 | `image.prompt` + `clean front-facing portrait, plain background` |
+| sheet (image, generate, `--from` headshot) | same | 16:9 | `@image1 is the face to match exactly. ` + `image.sheet` |
+| **voice (audio, generate, `--from` headshot)** | **`seed-audio-1.0`** | — | `prompt voice C01` — `（direction）line` |
+
+**The voice node references the headshot.** Seed Audio 1.0 does 参考图片配音: given the face it
+invents a voice that fits it, which is why the headshot must be approved first. It takes exactly
+one reference, so pass only the headshot node. About 3.6 credits a sample.
+
+Generate headshots first (a set of headshots is a batch — quote the total), **look at them**
+(download and open), then sheets and voices from the approved headshots. Check a sheet for: one
+face across all views, three full bodies of equal height, no squashed figures, white background.
+Regenerate rather than accept a sheet with two faces — every shot inherits it. Listen to the voice
+before you attach it: the age and the weight of the voice should match the face.
+
+Attach and bind:
+
+```bash
+wji fs set <datasetId> --add-images <headshotAsset> <sheetAsset> <voiceAsset> --icon <headshotAsset> --set primaryAudio=<voiceAsset>
+node {baseDir}/scripts/short-drama.mjs bind <project> characters.C01.wji.assets.headshot <headshotAsset>   # and sheet / voice, nodes.*
+```
+
+The voice lives in **three** places on the dataset, and all three matter: `images[]` (the asset
+list takes audio too — the clip is then playable from the dataset card next to the headshot and
+the sheet), `primaryAudio` (which one is *the* voice), and `content.audio_prompt` (the brief that
+produced it, so it can be regenerated identically). `render character` puts the same brief, the
+sample line and its acting direction into the readable 声线 section of `content.content`.
+
+**The sample is a dry voice.** Seed Audio is an audio-scene generator — asked for a line it will
+happily add a room, a score and a coffee machine — so the composed prompt ends its parenthetical
+with `干声，无配乐无音效`. The same model makes the segments' soundtrack beds in stage 5
+(`prompt audio`), where the room, the score and the coffee machine belong.
+
+### What the voice asset is for
+
+It is a **reference for later**, not a deliverable. Every ref-capable video model takes audio
+references (`wan-3.0-ref` up to 5, ≤ 15 s total; `seedance-2.5-ref`; `minimax-h3-ref` up to 3), so
+on an episode canvas the character's voice display node connects to the video node alongside the
+sheets and becomes `@audio1…`, holding the timbre across clips the way the sheet holds the face.
+Set `params.voiceSamples: false` only when no TTS model is enabled on the account — `manifest`
+then stops asking for the node.
+
+## Gates
+
+`profile` · `bible` · `relationships` · `char-prompts` · `voice` · `distinct`. Missing assets are
+reported by `manifest`, not by `validate` — a character without a voice model available is still
+a valid character.

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

@@ -0,0 +1,103 @@
+# Stage 1 — Outline
+
+Turn the source (novel, synopsis, or an original idea) into the structure everything else hangs
+on: what stays, who is who, where the beats land, what happens each episode. **Skeleton before
+episodes, sign-off between them.**
+
+## Inputs
+
+1. The source text the user pointed at — read it; never invent content from a title. If it is
+   long, summarise per volume in a scratch file first and keep quotes for `keep[].evidence`.
+2. The four parameters (episodes × minutes, genre, adapt mode, ratio) — already in `params`.
+3. Preferences the user stated (keep this character, keep that scene) — honour them and record
+   them in `outline.keep`.
+
+## Order of work (the order is the method)
+
+0. **Brief.** `outline.brief` — the six judgements of `craft-story.md` §1 (logline, audience,
+   the 爽点 engine, the main hook, the pressures on both sides, the emotional line, the promise).
+   A one-line idea is not enough to fill these; ask the user the focusing questions rather than
+   inventing a genre. Then `params.register` (语体) if not given at `init`.
+1. **Cut.** A short drama carries one main line and at most one sub-line. Everything else goes to
+   `outline.cut` with a `why`. Any adapt mode other than `原创` / `忠实` must cut something.
+   Write `cutNote`: what the story's endpoint becomes after the cuts.
+2. **Merge people.** Characters with the same function become one face. Then tier them:
+   `lead` (1–5, the protagonists and main antagonist, `arc` required), `support` (≤10, named
+   secondary characters, `arc` required), `functional` (≤10, a role label as the name, no arc).
+   Unnamed background people are not listed. Write `mergeNote`: why these are the leads.
+3. **Collect scenes.** Primary scenes are capped at `4 + ⌈episodes/10⌉` (5–15) — generated scenes
+   are cheap, but every extra one is another consistency asset to maintain. A scene used in only
+   one episode needs a `reusePlan` (which existing environment it derives from, what changes) or
+   becomes a `variantOf` in Stage 3.
+4. **Place the beats (爽点).** Majors first — where the big reveals land — then minors to fill.
+   Rules: adjacent beats ≤ `maxBeatGap` (3) episodes apart, a beat in episode 1, a beat in the
+   finale, at least one major, and the first major never on the last episode. `type` follows the
+   genre (打脸 / 揭破 / 反转 / 逆袭 / 认亲 …).
+5. **Pick the narrative props.** After the beats, not before — a prop is a prop because it carries a
+   beat. Test: *if it broke, got lost or was swapped, would the plot collapse?* Yes → `props[]`
+   with `function` and `beatIds`. No → set dressing; it becomes a scene anchor in Stage 3. At most
+   8.
+6. **Structure and guideline.** `outline.structure` — golden / escalation / climax with episode
+   ranges (first ~20 % / ~50 % / ~30 %; `craft-story.md` §2) and a task, 爽点 plan and hook plan
+   each. `outline.guideline` — the 创作指引: tone, emotional intensity, dialogue style, visual
+   style, production limits, ≥ 3 forbidden rules, 3–5 priorities. It is the document every later
+   stage re-reads (`render guideline`), so write it as rules, not as a summary.
+
+Run `validate --stage outline` now. The `episodes` gate will fail (no episodes yet) — every other
+gate should pass before you show the skeleton.
+
+## Sign-off ⛔
+
+Show the user, in plain language: the lines cut, the people merged, the episode of each major beat.
+Wait for a yes. Change the skeleton on their feedback and validate again.
+
+## Episodes (batches of ≤10)
+
+Design first, prose second. Per episode the 分集目录 fields: `function` (开局钩子 / 冲突升级 /
+身份揭露 / 情感推进 / 大反转 / 清算收束), `task` (`goal` / `obstacle` / `stake`), `change` (这一集改变了
+什么 — one sentence; "they talked" means the episode is not strong enough), `satisfaction`
+(`type` from the 爽点 library + the visible `action` that does it), `hookType` (身份 / 危机 / 结果 /
+情感 / 反转 / 道具). The first three episodes are the retention window: 1 grabs (开局钩子), 2 proves
+the 爽点 (its type is the brief's engine), 3 widens the line (冲突升级 / 身份揭露 / 大反转); the finale
+includes 清算收束 or 大反转. Then `synopsis` (narrative prose — a quotation mark means you are writing dialogue, which
+is Stage 4's job), `hook` (the question that makes them tap the next episode), `suspense` (what is
+still buried), `sceneIds`, `characterIds`, `propIds`, `crowdPlan` when three or more characters
+appear (how the group is staged: who is foreground, who is over-the-shoulder, or "never in one
+frame — separate scenes"), and `warnings` from the fixed list (雨戏 / 肢体接触 / 人群 / 手部特写 —
+the things video models get wrong; over-report).
+
+A beat placed on an episode must actually happen in that episode's synopsis. Every character,
+scene and prop must be used somewhere — `refs` will name the idle ones.
+
+## Sync
+
+```bash
+node {baseDir}/scripts/short-drama.mjs render <project> outline > /tmp/outline.md
+wji dataset new <story folder> "<title> 大纲" -t outline -d "<core sentence>"      # once; bind wji.outlineDatasetId
+wji fs set <id> --merge content=@/tmp/outline.md --merge viewMode=content \
+  --merge metadata.shortDrama=@/tmp/outline.json                                  # outline block + params
+node {baseDir}/scripts/short-drama.mjs render <project> guideline > /tmp/guideline.md
+wji dataset new <story folder> "<title> 创作指引" -t outline -d "<logline>"        # once; bind wji.guidelineDatasetId
+wji fs set <id> --merge content=@/tmp/guideline.md --merge viewMode=content \
+  --merge metadata.shortDrama=@/tmp/guideline.json                                # {brief, structure, guideline, register}
+```
+
+The 创作指引 is its own dataset because it is the thing to open first, every time — before a
+character, an episode or a storyboard, print `render guideline` and read it.
+
+Re-run the same `fs set` after every outline change; the dataset is a view of `project.json`.
+
+## Gates (what `validate --stage outline` checks)
+
+`params` · `brief` · `structure` · `guideline` · `adaptation` · `tiers` · `scenes` · `props-cap` ·
+`beats` · `episodes` · `episode-design` · `refs` — see the gate labels in the output; each failure
+names the entity and the fix.
+
+## Rhythm for a short drama
+
+A 短劇 episode is paced by its 爽点, not its minutes: plan **one beat per 30–45 s** plus the
+cliff, so a 1-minute episode carries 1–2 beats and a 90-second one 2–3 (the gates only demand one
+per episode and a gap ≤ 3 — the rhythm is on you). Put the hook's image inside the first three
+script beats, a beat near the midpoint, and let the cliff be the last thing on screen. Adding beats
+to an existing episode is fine — append new ids (`B11`, `B12`), never renumber, and list them in
+that episode's `beatsClaimed`.

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

@@ -0,0 +1,113 @@
+# Stage 4 — Script
+
+One episode per task, batches of ≤3 episodes before moving on — the script is the layer that gets
+rewritten most; small batches fail cheaply. **The script handles drama; the storyboard handles
+shooting** — no shot numbers, no camera, no prompts here.
+
+## Start from the seed
+
+```bash
+node {baseDir}/scripts/short-drama.mjs seed <project> script --ep 1
+```
+
+`episodes/E01.json` now has `targetSeconds`, `hook`, `cliff`, `beatsClaimed`, `hookBeat: [1, 1]`,
+the episode design copied from the outline (`function`, `task`, `change`, `hookType`),
+`satisfactionBeat: null` (to be set) and a `seedNote` with the synopsis, the 爽点 to realise, the
+register, the guideline's priorities and forbidden rules, every speaker's `bible.speech`, the ids
+allowed in this episode, the crowd plan, the warnings, the previous episode's cliff and the beat
+budget. **Before writing: `render guideline`, then the scene bibles (anchors, lighting states) and
+the speakers' `render character` — the bible's `speech` is where the lines come from** (see
+`dialogue-craft.md`). Delete `seedNote` when the scenes are written.
+
+## Hard rules (all gated)
+
+1. **The duration budget is the wall.** Lines cost `chars ÷ 4.5` s, action beats 2.5 s; the
+   episode must land within ±15 % of `targetSeconds`. 60 s ≈ 18–22 beats, 120 s ≈ 35, 180 s ≈ 50.
+   Over: cut action beats first. Under: add conflict, never small talk.
+2. **One line, one breath**: ≤ 35 non-whitespace characters. Split long lines with a reaction
+   beat in between.
+3. **Speakers are in the scene** (`characters`), or `VO` (inner voice — whose, in `delivery`).
+   VO is emphasis, not a crutch: two or three per episode at most.
+4. **Action beats: one thing each, narrative, common actions only, no quotation marks.** The video
+   model performs actions it has seen a million times — walk in, sit down, hand over, nod, turn,
+   grab, hold tight. Precise physical interactions (block a pole with an oar), micro-expressions
+   (lashes tremble), inch-level moves, metaphors — rewrite or delete. Drama comes from the
+   *combination and timing* of ordinary actions.
+5. **Every scene has an action beat.** Three lines with nothing happening is radio.
+6. **Cold open.** The hook's concrete image is on screen within the first 3 beats (`hookBeat`
+   points at it; gated) and it is *in motion* — running with the suitcase, not a still close-up of
+   it. Static detail close-ups belong to the second appearance.
+7. **The last beat is the cliff** — an unfinished sentence, an unexplained action, a sound with no
+   source. `cliff` describes the effect; the flow delivers it.
+8. **Claimed beats are performed.** Each id in `beatsClaimed` (seeded from the outline) needs a run
+   of consecutive beats that actually realises setup → payoff.
+9. **The 短剧 rhythm** (`rhythm`): the hook image by `hookSeconds` (15 s) of cumulative time; a
+   turn — new information, an obstacle, a counter, a choice — at least every `turnSeconds` (30 s),
+   marked `turn: true` on the beat; the 爽点 (`satisfactionBeat`) after the hook, before the last
+   beat, on an action beat or a line with an action beside it; never more than
+   `maxConsecutiveLines` (3) lines in a row; action beats at least `minActionRatio` (35 %) of the
+   lines. The seconds come from the same duration math as the budget.
+10. **The design is realised** (`script-design`): `function`, `task`, `change`, `hookType` equal
+    the outline episode's; the 爽点 type the outline promised is what `satisfactionBeat` delivers.
+11. **Filmable** (`filmable`): no 心想 / 内心 / 回忆起 / 脑海, no 像是 / 似乎 / seems to, no more than
+    three hand verbs in a beat (`craft-shots.md` §generability).
+12. **The register** (`dialogue-register`, advisory): lines match `params.register` — Cantonese
+    markers only in 粤语, no modern words in 古装, no net slang on the middle-aged, English short and
+    contracted with names used sparingly. It reports and never blocks; a character who breaks the
+    register on purpose gets `bible.speech.register`.
+
+Scene fields: `sceneId` and `lighting` must exist upstream (a new lighting state is added to the
+scene in `project.json` first, never bypassed), `characters` / `props` ⊆ the outline's ids for
+this episode.
+
+## Editing discipline
+
+- **Change one beat, re-read three.** Position, who holds the prop, who is present — spatial state
+  lives in your head while writing and vanishes while editing. Re-reading the neighbours is the
+  only way to get it back.
+- **A picture change is a sound change** downstream — remember it when the storyboard exists.
+
+## Feel (not gated, decides whether it plays)
+
+Short lines, dense action: an action or reaction every two or three lines, never four lines in a
+row. Subtext goes in `delivery` ("says sit, doesn't move"), not spoken. Stage groups per the
+outline's `crowdPlan`. Scene changes are free for the generator but cost the viewer — tighten one
+scene rather than hopping four.
+
+| Fault | Symptom | Fix |
+| --- | --- | --- |
+| radio play | lines only | insert a visual action every 2–3 lines |
+| literary | actions as prose (oar blocks the pole, lashes tremble) | common actions |
+| dead opening | first beat is a still close-up | give the same image in motion |
+| space jump | on the path, next beat on the pier | add the crossing beat; re-read three |
+| padding | greetings, tea, directions | delete; every beat advances or flips |
+| speech | one person, four lines | split to the partner or break with action |
+| omniscience | lines explain everything | move information into action and delivery |
+| flat ending | ends calm | last beat becomes the cliff, even just a whistle |
+
+## Polish (`render dialogue --ep N`)
+
+When the user asks for a polish, when the dialogue is the visible weakness, or when the register
+gate keeps noting the same scene: `render dialogue --ep 1` prints the 逐句校对表 (原句 / 问题 / 修改句 /
+修改理由 / 导演／演员提示) with each speaker's speech rules in the last column. Fill it per
+`dialogue-craft.md` §3, apply the changes to the flow, re-validate, and append the filled table to
+the screenplay dataset under `## 对白校对`.
+
+## Sync
+
+```bash
+node {baseDir}/scripts/short-drama.mjs validate <project> --stage script --ep 1
+node {baseDir}/scripts/short-drama.mjs render <project> script --ep 1 > /tmp/E01-script.md      # 短剧 episode format + 创作自检
+wji dataset new <story folder> "E01 剧本" -t screenplay -d "<hook>"        # bind --ep 1 wji.screenplayDatasetId
+wji fs set <id> --merge content=@/tmp/E01-script.md --merge viewMode=content --merge metadata.shortDrama=@/tmp/E01.json
+```
+
+## Gates
+
+`script-shape` · `beat-flow` · `duration` · `hook-cliff` · `beats-claimed` · `script-design` ·
+`rhythm` · `filmable` · `dialogue-register` (advisory).
+
+`render script` writes the reference's episode format — `△（from-to秒）` action lines with the
+seconds derived, `角色（delivery）：line`, `【爽点标记】` after the 爽点, `【钩子标记】` at the end — and
+closes with the 创作自检 table, eight questions answered from the same helpers the gates use. A
+green episode is 16/16; a row that says 否 names what to fix.