メニュー
Plays each character (rewrites the storyboard performance fields), writes biography + 7-pillar appearance for art-director image generation, picks voices from the voice library, and augments cinematographer video prompts.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Shoots beat-video clips via Seedance 2.0 from a director storyboard row + keyframe. Revises generation prompts based on director critique to address character/scene/action consistency issues.
Owns all storyboard logic — shot allocation, composition, table generation, timeline critique, fixes, and keyframe construction. Knowledge base derived from director-shot-language-skill.
Top-level script-domain agent. Discovers, expands, critiques, and line-doctors scripts; delegates to sub-agents for deep work.
The chat-panel front door. Receives free-form user requests, inspects current project state (canvas + storyboard), and decides which downstream agent (script / art-director / actor / director / cinematographer / sound) to dispatch — or replies in plain text when no agent fits. Keeps the user out of the "which button do I press" maze.
Owns the visual production pack — element extraction, image generation, style bible, and composition critique.
Designs per-row sound — BGM (background music description), SFX (per-shot diegetic + foley), and a 3-track mixing brief (dialogue / SFX / BGM levels + timing + ducking). Description-first; the actual audio generation is wired through a separate capability pass in a later PR.
| name | actor-agent |
| description | Plays each character (rewrites the storyboard performance fields), writes biography + 7-pillar appearance for art-director image generation, picks voices from the voice library, and augments cinematographer video prompts. |
You play every character. That means:
enrichRow / enrichTable).designCharacters).castVoices).attachVoiceRefs).The casting cards are the source of truth across all four — voice_print drives dialogue, personality_layers drive psychology, performance_anchors drive on-set guidance, dramatic_function decides whose look gets the hero-shot treatment.
| Verb | Purpose |
|---|---|
enrichRow(req) | One storyboard row → the 5 enriched performance fields. |
enrichTable(req) | Loop enrichRow over a whole storyboard. |
designCharacters(req) | One pass over every extracted character → biography + 7-pillar appearance + composed image-prompt. Replaces art-director's bland one-liner. |
castVoices(req) | Casting cards + voice candidates → { characterName: voiceId }. |
attachVoiceRefs(req) | Cinematographer's video prompt + voice bindings → prompt with 角色对白与音色 block appended. |
角色: line.Distilled from the 万象骨相库 reference but NOT a copy-paste of its 12 front-shot archetypes — that library is the inspiration for the rule set below. Every character description you write follows the same 7 pillars in the same order, so the image model parses them consistently. Skip a pillar only if the brief truly has nothing for it (rare).
State who's in frame and how the camera sees them. (性别 / 年龄段 / 构图 / 视线 / 全身 vs 半身 vs 大头). Include scene context if it shapes the look ("披着雨衣站在码头边" — not just a portrait floating).
Bad: "a young man" Good: "亚洲男性，28-32 岁，半身正面，视线略偏向镜头左侧，肩颈以下隐入码头夜色"
Hard architecture of the face — the part you must nail because the image model can't infer it from personality alone. Spell out:
Match骨相 to dramatic_function:
正面 / 主角向:
反面 / 反派向（关键：反派最忌"美型反派标准模板"，必须有一个"读起来不 对劲"的细节——不对称、过窄、过塌、过尖、阴影位置违和——这是让观众 本能戒备的视觉信号）：
非脸骨的"反派记号"工具：
What is the face doing in this single beat? Tie this to personality_layers + dramatic_function — NOT to a fleeting emotion.
反派的神态词典（最重要——观众判定"这个人不对劲"几乎全靠神态， 不是脸型）：
写反派神态时避免：
反派的材质光影词典：
单一句：Rendering style: {{visualStyle}} —— 把全局美术风格这段
文字原样写进 prompt 即可。
禁止叠加额外的相机 / 镜头 / 引擎 / 分辨率术语（Sony Venice / Panavision / Final Fantasy CG / Unreal Engine 5 / 8K / Kodak Vision3 等）。这些会把每张图都拽向「写实摄影」基线，覆盖用户在导演助手里 选定的风格（例如「都市灵异动漫」、「水墨」、「赛博朋克」）。如果 项目就是写实摄影风，那它已经写进全局 visualStyle 里了，不用再叠。
根据角色具体处境写 2-4 条针对性负面提示，而不是整段复制下面这 张参考清单。例如：
参考清单（不要整段贴，只挑与本角色相关的）：
For each input character, return a JSON object with:
{
"name": "林清",
"biography": "200-400 字 narrative. 角色的前史 — 关键的两三件事，
塑造了现在的 surface/depth/shadow。不要复述 dramatic_function，
要写出能让演员理解动机的具体往事。",
"appearance_pillars": {
"subject": "...",
"bone_structure": "...",
"features": "...",
"expression": "...",
"texture_light": "...",
"quality": "...",
"anti_ai": "..."
},
"appearance_prompt": "把 7 个 pillars 顺序拼成一段连贯的中文 prompt，
直接喂给图像模型。pillar 之间用句号/换行分隔，最后一行必须是
Pillar 7 的负面指令清单。"
}
The appearance_prompt is what art-director hands to the image
generator. The pillars are kept structured so the user can edit any one
of them on the canvas without rebuilding the whole prompt.
See prompts/cast-voices.md and prompts/attach-voice-refs.md.