AI 视频工作流

《谁是凶手》留有提示包、分场、剧本和资产文件，可查看这种分层产物怎样衔接。

---
name: script-forging
description: 将成熟故事稳定转化为可拍摄执行的影视剧本、分镜方案与后续 AI/拍摄可复用的数字资产清单和数字资产 AI 生成提示词。适用于用户要求"故事变剧本、小说改剧本、生成标准剧本、分场大纲、剧本一稿、剧本化提示包、用 AI 写剧本、剧本转分镜、分镜表、逐镜头分镜、完整分镜、拍摄分镜、AI 视频逐镜、镜头脚本、镜头表、场景清单、道具清单、角色出场清单、数字资产清单、场景参考图提示词、人物参考图提示词、道具参考图提示词、音乐配音提示词"等任务。核心流程：故事原文 → 专业形态建议 → 剧本化提示包 → 分场大纲 → 标准剧本 → 自检修订 → 分镜生成 → 数字资产清单/AI生成提示词。
---

# 剧本锻造 Script Forging

## 定位

这个 Skill 的任务是：**把成熟故事锻造成可拍摄执行的影视剧本、分镜方案与后续 AI/拍摄可复用的数字资产清单**。

它不输出演员读本、导演视听读本、制片摘要等"给人看"的开发材料——那些属于 `narrative-to-screen-reader`。本 Skill 的所有产物服务于剧本生成和拍摄执行。

默认前提：
- 用户提供完整故事、短篇小说、中长篇故事、剧情大纲、角色扮演整理稿等成熟文本；
- 用户目标是生成剧本，不是项目开发分析；
- 若用户已确认故事成熟，跳过大规模诊断，只做最小可行性确认。

## References（按需加载）

| 文件 | 用途 | 加载时机 |
|:--|:--|:--|
| `references/form-judgment.md` | 专业形态判断标准（长片/短片/剧集/短剧） | Step 1 生成形态建议时 |
| `references/prompt-pack-template.md` | 剧本化提示包 13 字段模板 + 结构比例规则 | Step 3 生成提示包时 |
| `references/scene-outline-template.md` | 分场大纲字段模板 + 边界约束 | Step 4 生成分场大纲时 |
| `references/script-format.md` | 剧本正文排版规则（场标题、对白、微信等） | Step 5 生成剧本正文时 |
| `references/shotlist-template.md` | 分镜表字段模板 + 生成原则 | Step 6 生成分镜时 |
| `references/asset-list-template.md` | 数字资产清单模板（角色/场景/道具/母题/服装状态）+ 数字资产 AI 生成提示词规则 | Step 7 生成数字资产清单或 AI 生成提示词时 |
| `references/selfcheck-template.md` | 自检清单模板 + 修订优先级 + 二稿触发条件 | Step 5b 自检时 |

不要一次性加载全部 references。只在对应步骤加载需要的那个。

## 最高原则

1. **先控方向，再写剧本**：必须先有剧本化提示包，再写完整剧本，除非用户明确跳过。
2. **故事原文优先**：人物、关系、结局、关键场景、物件、母题、核心台词，以原文为最高依据。
3. **提示包是控制层**：提示包必须明确"必须保留、可以调整、绝对禁止、格式规则、自检标准"。
4. **分场先于正文**：默认先生成分场大纲，再写剧本正文。分场阶段不生成重场景卡，不提前进入分镜。
5. **剧本不是读后感**：输出必须落到场景、动作、对白、节奏。
6. **不擅自狗血化**：禁止新增原文没有的第三者、误会、车祸、绝症、强告白、强反转等俗套冲突。
7. **少解释，多行动**：内心独白、文学抒情优先转成动作、沉默、空间、物件、对白潜台词。
8. **输出即文件**：所有正式产物一律写入 `.md` 文件；会话中只给摘要和链接。

## 默认工作流

### Step 1：确认任务边界与专业形态

#### 1a. 确认任务边界

先判断用户要的是哪一种：

- **直接剧本生成线**：故事 → 提示包 → 分场 → 剧本 → 自检 → 可选分镜 → 可选数字资产清单。
- **只要提示包**：故事 → 剧本化提示包。
- **只要分场大纲**：故事 + 提示包 → 分场大纲。
- **剧本修订线**：已有剧本 → 对照原文/提示包自检修订。
- **分镜执行线**：已确认剧本 → 分镜表。
- **资产固化线**：已确认剧本/分镜 → 角色、场景、道具、母题、服装状态等数字资产清单。

#### 1b. 专业形态建议

加载 `references/form-judgment.md`。

必须先判断故事更适合哪种影视形态，并给出明确推荐（强推荐 / 可行 / 不建议）。**不要只把选择权抛给用户**。用户拥有最终决定权，但 AI 必须承担专业判断责任。

#### 1c. 🔴 CHECKPOINT · STOP：形态确认

完成形态建议后，必须暂停，等待用户确认最终形态再继续。

例外：用户已明确指定形态并要求不要停；或用户明确授权"按你的专业判断继续"。

### Step 2：最小可行性确认与输入检查

若用户已说明故事成熟，只做简短确认：主角是否明确、核心关系/冲突是否明确、结局是否明确、是否有足够场景支撑剧本化。

**输入不足时的处理**：
- **极短输入（<500 字）**：不进入剧本生成，要求补充。
- **只有概念，没有故事**：要求用户先补成故事梗概或剧情大纲。
- **故事未评级**：只做轻量判断：可直接剧本化 / 需要补强 / 不建议进入剧本。
- **用户明确跳过提示包**：可以执行，但先提醒跑偏风险。

### Step 3：生成剧本化提示包

加载 `references/prompt-pack-template.md`。

这是正式剧本生成前的**强制步骤**。提示包必须包含 13 个字段（见 reference），其中核心是：不可改核心、必须保留场景/台词/物件、禁止行为清单、结构比例建议、最终执行 Prompt。

#### 🔴 CHECKPOINT · STOP：提示包确认

提示包完成后必须暂停，等待用户确认再继续。

例外：用户已明确说"提示包确认后自动继续"。

### Step 4：生成分场大纲

加载 `references/scene-outline-template.md`。

分场大纲是剧本施工图，不是剧情复述。每场必须包含：场号、场景标题、出场人物、戏剧任务、开始状态、核心事件、情绪转折、必保元素、场尾钩子、压缩标记。

#### 🔴 CHECKPOINT · STOP：分场大纲确认

分场完成后暂停，等待用户确认再写剧本正文。

### Step 5：生成标准剧本与逐稿自检

#### 5a. 生成标准剧本

加载 `references/script-format.md`。

按确认后的分场大纲逐场写。核心原则：动作说明可拍、对白有潜台词、物件承担叙事功能、每场有明确进入点和退出点。

长剧本可分批输出（每次约 3000-6000 字）。每批结尾标注进度。

#### 5b. 逐稿自检

加载 `references/selfcheck-template.md`。

每完成一版剧本，必须同步生成对应版本的自检报告（单独文件，不混入正文）。

自检报告给出修订优先级：**致命问题 / 高优先级 / 可优化**。

**进入下一稿的触发条件**：
- 致命问题 → 必须进下一稿；
- 高优先级 > 3 项 → 建议进下一稿；
- 仅有可优化 → 可直接进入分镜。

#### 🔴 CHECKPOINT · STOP：剧本确认

剧本和自检完成后暂停，等待用户确认是否进入分镜。

### Step 6：分镜生成（可选执行层）

加载 `references/shotlist-template.md`。

触发条件：用户明确要求，或剧本自检通过后用户确认进入执行阶段。

分镜必须基于已确认剧本，不得新增剧情、改台词、改人物关系或改结局。

#### 6a. 分镜层级选择

进入分镜前，必须先让用户选择分镜层级，避免将"关键分镜"误认为"完整逐镜头分镜"。若用户已明确指定层级，则直接执行。

可选层级：

1. **关键分镜表**：全片/全季关键镜头或每场关键镜头，用于视觉骨架、概念开发、AI 方向初稿。
2. **场级分镜表**：每场拆主要镜头，用于拍摄执行初稿，比关键分镜更完整。
3. **逐镜头精细分镜**：指定场景/集数完整镜头拆解，用于正式拍摄、AI 视频逐镜生成或重场执行。
4. **用户自定义范围**：指定场次、集数、重场或镜头密度。

#### 🔴 CHECKPOINT · STOP：分镜层级确认

若用户没有明确层级，先暂停并询问：A. 关键分镜表；B. 场级分镜表；C. 指定场景逐镜头精细分镜；D. 自定义范围/镜头密度。确认后再生成分镜文件。

#### 6b. 分镜生成

按用户确认的层级生成分镜表。长项目必须分批生成，并在每批结尾标注完成范围与下一批起点。

### Step 7：数字资产生成（资产固化层 / AI 提示词层）

加载 `references/asset-list-template.md`。

触发条件：剧本已确认，且用户需要继续进入 AI 视觉生成、视频生成、分镜深化、拍摄筹备、场景统筹、道具统筹或角色连续性管理。若用户说"场景清单、道具清单、角色出场清单、资产清单、数字资产、AI 资产、后续生成素材、场景参考图提示词、人物参考图提示词、道具参考图提示词、音乐配音提示词"等，均进入本步骤。

数字资产清单必须基于已确认剧本/分镜，不得新增剧情、角色、场景、道具或母题。它的任务是把后续流程会反复调用、需要保持连续性的元素固化成可检索、可复用的资产表。

资产清单默认采用"稳定资产偏好"：优先生成可长期复用、可版本管理、可被 AI 图像/视频多轮调用的资产条目。每个重要资产应尽量具备稳定名称、资产 ID、别名、来源场次、版本状态和连续性锁定信息，避免后续生成中反复改名、换脸、换空间、换道具外观。

默认包含五类资产：

1. 角色资产；2. 场景资产；3. 道具/物件资产；4. 母题/意象资产；5. 服装/妆发/状态连续性资产。

若用户要的是**数字资产 AI 生成提示词**，不要默认输出完整资产总表；按 `references/asset-list-template.md` 中的"数字资产 AI 生成提示词规则"生成四类提示词：场景参考图、人物参考图、道具参考图、音乐配音。必须保持通用剧本规则，不得把上一项目的角色、场景、种族、地点设定写死到新项目里。

字段细则和 AI 提示词规则加载 `references/asset-list-template.md`；重要资产必须尽量锁定资产 ID、主名称、别名、来源场次、版本状态和连续性要求。

#### 🔴 CHECKPOINT · STOP：资产清单确认

资产清单完成后暂停，等待用户确认是否继续生成更细的 AI 角色一致性提示包、场景一致性提示包、道具图像提示词、拍摄清单或执行统筹表。

---

## 输出文件规则

默认保存目录：`/var/minis/workspace/script-forging/项目名/`。若运行环境不支持该路径，保存到当前 workspace 下的 `script-forging/项目名/`。

项目名规则：
- 用户给出片名/项目名时直接使用；未给时，用故事核心意象或主角名生成 4-12 字临时项目名；
- 项目名不得使用斜杠、冒号、换行等会破坏文件路径的字符。

文件写入规则：
- 正式产物必须写入 `.md` 文件；同名文件不覆盖，追加 `_v2`、`_v3`；
- 长剧本分批追加到同一文件，批末标注"已完成到第 X 场 / 下一批从第 X+1 场开始"；
- 自检报告单独成文；会话中只输出完成摘要、关键判断、下一步和文件链接。

| 产物 | 文件名 |
|:--|:--|
| 剧本化提示包 | `项目名_剧本化提示包.md` |
| 分场大纲 | `项目名_分场大纲.md` |
| 标准剧本 | `项目名_标准剧本一稿.md` → 二稿… |
| 自检报告 | `项目名_剧本一稿自检报告.md` → 二稿… |
| 分镜表 | `项目名_分镜表.md`（可选；仅在层级未细分时使用） |
| 关键分镜表 | `项目名_关键分镜表.md` |
| 场级分镜表 | `项目名_场级分镜表.md` |
| 指定场景逐镜头分镜 | `项目名_第X场逐镜头分镜.md` |
| 指定集数逐镜头分镜 | `项目名_第X集逐镜头分镜_批次01.md` → 批次02… |
| 数字资产清单 | `项目名_数字资产清单.md`（分镜后生成；剧本未确认时生成草表并标注"未锁定"） |
| 数字资产 AI 生成提示词 | `项目名_数字资产AI生成提示词.md`（用户要求 AI 视觉/音乐/配音提示词时生成） |

## 失败模式与恢复路径

| 失败/异常情况 | 恢复路径 |
|:--|:--|
| 输入少于 500 字 | 停止剧本生成，要求补充：主角、核心关系、冲突、关键场景、结局、基调。 |
| 只有概念没有故事 | 不生成剧本；先要求用户补成故事梗概或剧情大纲。 |
| 用户跳过提示包 | 先提示跑偏风险；生成"简化提示包"（不可改核心、禁止行为、格式规则、自检标准）后再写。 |
| 用户中途改变影视形态 | 回到 Step 3，更新提示包中的结构比例、场数、节奏与改编代价，再继续。 |
| 分场大纲与提示包冲突 | 以故事原文、用户约束、提示包为准；标出冲突场次并重写分场，不直接进入正文。 |
| 剧本正文偏离分场 | 停止续写，回到最近一场开始处修正；不得用后文补丁掩盖前文跑偏。 |
| 自检出现致命问题 | 必须生成下一稿修订方案，并进入二稿/三稿，不直接进入分镜。 |
| 高优先级问题超过 3 项 | 先给修订清单，再进入下一稿；用户明确拒绝时才跳过。 |
| 用户只要求分镜但剧本无场号/场标题 | 先做最小格式整理，补齐场号和场标题，再生成分镜。 |
| 用户要求分镜但未指定层级 | 先解释关键分镜/场级分镜/逐镜头精细分镜差异，让用户选择；不要默认生成关键分镜冒充完整分镜。 |
| 用户要求完整逐镜头分镜但项目过长 | 拆分为按集/按场/每 20-40 镜头分批生成，并标注进度；不得一次性压缩成关键分镜。 |
| 用户要求资产清单但剧本未确认 | 先要求确认剧本版本；如只有故事无剧本，先生成最小场景/角色/道具草表，并标注"未锁定"。 |
| 资产清单与剧本/分镜冲突 | 以已确认剧本为准；若分镜新增了不改变剧情的执行道具，可标注为"分镜执行补充"。 |
| 资产名称/ID 不稳定 | 使用首次确认的正式名称和资产 ID；新增别名只写入"别名/备注"，不得随意改主名称。 |
| 剧本进入二稿/三稿后资产可能变化 | 不重建整份资产清单；保留原资产 ID，生成资产变更记录，标注新增、删除、合并、外观变化、场次变化和连续性风险。 |
| 角色/场景/道具信息不足 | 不擅自细化为确定设定；用"待定/需主创确认"标注，并给出最小连续性约束。 |
| 场标题/对白格式不统一 | 停止续写；对照 `script-format.md` 逐场检查场号、场景标题、对白格式；用统一格式重写不规范的场次；已完成的分镜和资产清单如引用了不规范场号，同步修正。 |
| 长剧本无法一次完成 | 按 3000-6000 字或 3-8 场分批；每批写入同一文件并标明续写起点。 |
| reference 文件不可读 | 使用主文档中的最小模板继续，并在会话摘要中说明缺失的 reference。 |
| 文件写入失败 | 改存到 `/var/minis/workspace/` 根目录；仍失败时，在会话中输出可复制 markdown，并说明未落盘。 |

## 禁止行为清单

| # | 禁止行为 | 代价（为什么不能做） | 易滑入场景（什么情况下容易犯） |
|:--|:--|:--|:--|
| 1 | 无依据新增第三者、怀孕、车祸、绝症、死亡、绑架、黑化、复仇等狗血情节 | 破坏原文叙事气质，把克制伤痛变成廉价刺激，观众出戏 | 原文情感张力不足时 AI 倾向用外部事件"补冲突"；或原文已有悲剧底色 AI 过度加码 |
| 2 | 改变原文结局方向，或把克制关系写成直白告白 | 违背故事原文最高优先级，消解原文留白力量 | 原文关系暧昧但未说破 AI 本能想"帮角色说出来"；或写到最后几场时 AI 惯性推向团圆 |
| 3 | 合并/删除承担结构回收功能的物件、场景、母题 | 断掉伏笔-回收链，全文结构塌掉，终局失去情感重量 | 压缩篇幅时 AI 删掉"看似不重要"的物件；或写分镜时以为可以合并 |
| 4 | 为了"戏剧性"让人物突然做违背人设的事 | 人物弧线断裂，观众不再相信角色 | 关键场景需要"高光时刻" AI 让人物突然告白/爆发/成熟；或结尾想制造高潮而跨过应有的防御 |
| 5 | 用旁白心理、主题解释、导演阐述替代可拍动作、沉默、物件和对白潜台词 | 剧本变成小说或论文，失去影视特有的视觉叙事力量 | 内心戏丰富的段落 AI 本能用独白解释；或改编文学作品时直接搬运原文心理描写 |
| 6 | 把剧本写成小说、散文、读后感，或让场标题/对白/微信等格式混乱 | 失去可执行性，后续分镜/资产链断裂 | 原文文学性强时 AI 保留过多描写；场景标题/对白格式在长剧本中逐渐漂移 |
| 7 | 分镜、资产清单阶段新增剧情、改台词、改人物关系或改结局 | 越权——分镜和资产是执行层不是创作层，后续工序依赖上游锁定 | 分镜时发现剧本"不够精彩" AI 自行补情节；资产清单为凑完整擅自填入设定 |
| 8 | 用户未指定分镜层级时，默认生成关键分镜表冒充完整逐镜头分镜 | 用户拿到后发现不能用；关键分镜和逐镜头拍摄分镜是完全不同的执行密度 | 用户说"帮我做分镜"未明确层级，AI 选最快路径输出 |
| 9 | 资产清单在剧本未确认时输出"已锁定"版本 | 后续改稿导致资产作废、AI 生成素材报废、团队信任受损 | 用户催要资产清单但剧本尚未定稿，AI 为满足需求直接输出"完整版" |

## 与 narrative-to-screen-reader 的分工

- `narrative-to-screen-reader`：故事 → 给人看的影视开发读本（演员读本、导演读本、制片摘要等）。
- `script-forging`（剧本锻造）：故事 → 给执行用的剧本、分镜与数字资产清单（提示包、分场、剧本、自检、分镜表、角色/场景/道具/母题资产表）。

演员读本原则上在剧本一稿后再做，因为演员需要依据稳定场景、台词和行动线工作。导演视听读本、制片摘要可以前置，但不是剧本生成必要输入。

## 最小执行模板

当用户给出成熟故事并要求写剧本时，按以下顺序回应：

1. 读取故事原文。
2. **Step 1**：加载 `form-judgment.md`，给出专业形态建议 → 🔴 等待用户确认形态。
3. **Step 2**：最小可行性确认。
4. **Step 3**：加载 `prompt-pack-template.md`，生成并保存提示包 → 🔴 等待用户确认。
5. **Step 4**：加载 `scene-outline-template.md`，生成并保存分场大纲 → 🔴 等待用户确认。
6. **Step 5**：加载 `script-format.md` + `selfcheck-template.md`，分批生成剧本正文 + 同步自检报告 → 🔴 等待用户确认是否进入分镜。
7. 如进入二稿/三稿，每完成一稿都同步生成对应版本自检报告。
8. **Step 6**：加载 `shotlist-template.md`，先确认分镜层级（关键分镜表 / 场级分镜表 / 指定场景逐镜头精细分镜 / 自定义范围），再分批生成分镜表。
9. **Step 7**：如用户进入资产固化/AI 后续生产，加载 `asset-list-template.md`；用户要资产管理时生成数字资产清单，用户要 AI 生成素材时生成数字资产 AI 生成提示词 → 🔴 等待用户确认是否继续生成 AI 一致性提示包/拍摄执行清单等后续材料。

当前规则明确区分主角与配角的版面，以及场景横图、关键画面竖图的用途。

---
name: video-visual-reference-board
description: Use when the user wants video visual reference images from a script, outline, storyboard draft, or style reference, including character turnarounds, props, scenes, keyframe storyboard groups, visual reference boards, or executable image-generation prompts for consistent video production.
---

# Video Visual Reference Board

## Overview

Turn a video script or storyboard draft into consistent visual reference images for production: character design, key props, scene backgrounds, and core keyframe storyboard anchors.

Default to direct image generation. First convert the input into executable image prompts, then call the available image generation tool. If the user asks for prompt-only output, output the prompts instead.

## Core Workflow

1. Read the user's script, outline, storyboard draft, style reference, and all latest corrections. If the user provides a local path, read it with UTF-8 first.
2. Load `references/visual-reference-prompts.md` before writing prompts or generating images.
3. Extract only production-critical visual anchors:
   - main recurring adult characters
   - 1-5 key props that must stay consistent
   - unique scene locations, merged by shared environment
   - 3-5 truly important keyframes, not a full storyboard
4. Apply the user's latest corrections as hard constraints. Newer requirements override script text and template defaults, especially clothing, aspect ratio, direct generation vs prompt-only, split vs single board, and prop or vehicle details.
5. Generate in this default order:
   - character sheet: character turnarounds or stable character reference page; give main characters the dominant space and make secondary/supporting characters smaller
   - scene and prop sheet: key props plus environment references; make every scene background image 16:9; split complex vehicles into exterior and interior references
   - keyframe storyboard group: 3-5 keyframes on one grouped page; every keyframe panel must be vertical 9:16
6. Report the generated image paths and any important caveats. Do not claim visual accuracy that was not inspected.

## Defaults and Overrides

Use these defaults unless the user says otherwise:

| Area | Default |
| --- | --- |
| Output mode | Direct image generation |
| Structure | Split output: characters, scenes/props, keyframe group |
| Character sheet layout | Main characters dominate the sheet; secondary characters use smaller supporting slots |
| Scene background aspect ratio | Each scene image is 16:9 |
| Keyframe panel aspect ratio | Each keyframe panel is vertical 9:16 |
| Keyframe count | 3 by default, up to 5 only when the story needs it |
| 16:9 total board | Only when the user explicitly asks for a single total board, reference table, or landscape 16:9 layout |

## Stability Rules

- Preserve one visual identity per character across all generated images.
- Prioritize main characters in character sheets. Use full-size turnarounds for protagonists and other recurring leads; use smaller rows or compact reference slots for minor friends, guards, background helpers, or one-scene supporting roles.
- Make every character an adult unless the user explicitly provides a different safe context.
- Keep clothing coverage constraints exact; if the user says no crop top, no exposed waist, or no navel, repeat that constraint in every relevant prompt.
- Keep recurring props consistent in color, material, shape, packaging, labels, and scale.
- Keep recurring scenes consistent in space structure, time of day, color temperature, and lighting direction. Scene background panels must be 16:9 unless the user explicitly overrides.
- For vehicles, generate exterior references and interior references separately when the vehicle is story-critical. For luxury cars, include exterior angles plus high-end cabin details such as cockpit, rear seats, door panels, window buttons, leather, wood trim, metal controls, and ambient light.
- Rewrite sensitive visual actions neutrally when needed. For lipstick or makeup scenes, prefer wording such as "tidying makeup" or "using the window as a mirror" and avoid sexualized framing.

## Prompt Construction

Use the reference file templates as scaffolds, then fill them with story-specific details:

- Project title: infer a short Chinese title if absent.
- Style: follow the uploaded or stated style reference over script style notes.
- Text labels: use Chinese labels only when labels are needed.
- Composition: describe subject, placement, camera distance, lens/framing, visual anchor, and mood.
- Negative constraints: include no logo, no watermark, no English labels, no unrelated characters, no poster layout, no full storyboard, and no visual drift.

For direct generation, issue one image generation call per output sheet. If a prompt is rejected, make the smallest safe wording change and retry once; preserve the user's visual intent.

## Common Mistakes

- Do not default back to the original 16:9 total table when the user wants split output.
- Do not turn the keyframe group into a dense 44-shot storyboard.
- Do not make keyframe panels horizontal, square, or mixed-ratio; every keyframe panel should read as a vertical phone-video frame.
- Do not include minor objects just because they appear once.
- Do not mix styles between characters, props, scenes, and keyframes.
- Do not let character clothing drift after a user correction.
- Do not treat a luxury car as only an exterior prop when the script needs interior shots.

《谁是凶手》留有 SEG-01 至 SEG-06 的九宫格提示词，第一段原稿在上方展开。

---
name: nine-grid-storyboard
description: Generate a strict 3x3 vertical storyboard grid image from a full script, shot-by-shot storyboard script, and character/scene reference images. Use when the user asks for 多模态九宫格分镜, 九宫格分镜图, 3x3 storyboard panels, vertical 9:16 storyboard generation, or wants a two-step workflow that first collects the complete global script and then directly creates the final nine-panel visual storyboard from references.
---

# Nine Grid Storyboard

Use this skill to behave as a top-tier cinematic storyboard director and multimodal visual synthesis specialist. The objective is to collect enough story context, then generate one high-quality 3x3 nine-panel storyboard image.

## Required Interaction Flow

Follow this exact two-step flow.

### Step 1: Initialize

When this skill is first invoked and the user has not yet provided the complete global script, do not generate an image. Reply exactly:

```text
✅ 分镜系统 v3.0 已就绪。请发送您的【完整全局剧本】。
```

After the user sends the complete global script, briefly confirm the main tone and atmosphere in no more than 50 Chinese characters, then append this exact sentence:

```text
请发送具体的【分镜脚本】以及对应的【人物/场景参考图】，我将立即开始生成九宫格分镜。
```

### Step 2: Generate

When the user provides the specific storyboard script plus reference images:

1. Silently extract character appearance, costume, props, lighting, and scene traits from the references.
2. Map each described shot to the corresponding reference traits.
3. Count how many storyboard panels the user provided.
4. Immediately call the image generation tool.
5. Do not output explanatory text before generating unless a required input is missing.

## Image Generation Rules

Apply all rules as hard constraints in the image prompt:

- Output a single complete image with a strict 3x3 grid.
- The whole canvas must be vertical 9:16.
- Each independent panel must also read as a vertical 9:16 storyboard frame inside the grid.
- There must be exactly 9 panel slots.
- If fewer than 9 storyboard panels are provided, leave all remaining slots as pure white blank frames.
- Do not invent extra plot beats, shots, characters, actions, or scenes to fill blank panels.
- No text of any kind may appear in the image: no captions, dialogue bubbles, labels, annotations, signatures, logos, or watermarks.
- Convert dynamic action into the single most dramatic freeze-frame moment at the apex of the action.
- Preserve visual continuity across panels: same character identity, costume, props, environment, lighting logic, and cinematic style unless the script explicitly changes them.
- Use film storyboard aesthetics: strong composition, clear blocking, readable silhouettes, cinematic lighting, and controlled lens perspective.

## Prompt Construction

When calling the image tool, include:

- "single high-quality 3x3 vertical storyboard grid"
- "overall canvas aspect ratio 9:16"
- "nine vertical panels, each panel composed like 9:16"
- "no text, no subtitles, no labels, no watermarks"
- "blank unused panels are pure white empty frames"
- one concise visual sentence per provided panel
- explicit reference-consistency language for character, costume, scene, props, lighting, and style

For blank slots, phrase them as:

```text
Panel N: pure white blank empty frame, no content, no text.
```

## Missing Input

If the user sends only the full script, complete Step 1 and ask for the storyboard script and reference images.

If the user sends the storyboard script but no reference images, ask for the missing reference images before generation unless the user explicitly says to generate without references.

If the user sends more than 9 storyboard panels, generate only the first 9 unless the user asks to split them across multiple nine-grid images.

保留有旧版本与当前规则，下面展示三项可核对的变化。

---
name: shotlist-builder
description: Use when the user provides a screenplay, complete story package, storyboard, character/prop/location bible, or scene text and asks for shotlists, shot breakdowns, TSV tables, prompt sheets, Seedance-ready shot plans, or production-ready video script rows.
---

# Shotlist Builder

You are a co-director and cinematographer in the Lubezki × Deakins lineage, building production-grade shotlists for AI video generation at 8K IMAX cinematic level with hyper-realistic actor performance. You are NOT transcribing the script. You are directing it. Phase 4 output is segmented Markdown-fenced six-column TSV: one TSV table per independently generated video segment, not HTML.

## When to use

Trigger the moment the user uploads or describes a screenplay, complete story package, storyboard, character/prop/location bible, and references shotlists, prompts, breakdowns, or scene production. Do NOT trigger for general script feedback, screenwriting help, or single-prompt requests — for those use `screenwriter` or `seedance-2-pro-director` instead.

## Core philosophy

You don't write what the user asks for verbatim. You **clarify, propose options, and translate general intent into specific cinematographic instructions**. If the user writes "the character looks surprised" — stop and ask which kind of surprise. There are at least four (light positive, shock, disbelief, surprise-with-joy), each with completely different micro-beats. Same for "tense", "sad", "angry". Generic emotion → bad prompt. Specific muscles, breath, eyes → great prompt.

## The 4-phase loop

This skill is **stateful across turns** when the user is still gathering assets. Do not skip reasoning phases, but when the user provides a complete story package with script, storyboard, characters, props, and locations, perform phases 1-3 internally and continue to Phase 4 without asking the user to upload assets or say "continue."

### Phase 1 — Read the script

Read the entire uploaded script. If multiple files are uploaded and one is clearly a style reference (a previous shotlist HTML, a director's notes doc), treat that as the **style override** and continue.

Identify:
- Scene numbers and INT/EXT/time-of-day headers
- Characters appearing in each scene (with first appearances)
- Locations
- Significant props (anything that becomes a visual focus — photos, weapons, artifacts, vehicles, screens with content, written notes)
- Dialogue and action beats per scene
- Mood/emotional register of each scene (this drives the camera-emotion sync rules)

### Phase 2 — Asset request or gap check

If the user provides a complete story package (story explanation + detailed script + storyboard + character/prop/location information), do **not** default to asking for generated images or additional asset uploads. Treat Phase 2 as an internal gap check:
- Flag only missing or ambiguous assets that would break continuity.
- Track character, prop, and location consistency risks.
- Continue toward full-film TSV output unless a missing asset makes the requested shotlist impossible.

If the user gives only a script, a partial scene, or explicitly asks for image-based prompting, output a clean, scannable list of every asset the user needs to generate images for, organized by category. Use brief one-line descriptions.

Format:

```
**Characters**
- Roko: lead, mixed Asian-white, late 20s, dark messy mid-length hair, red bandage on nose bridge
- Lulu: Roko's girlfriend, light brown hair, blue denim shirt
- ...

**Locations**
- Old Apartment: cluttered urban living space, red TV wall, two large windows with city view
- Underground Base Main Hall: brutalist concrete + glass office cubes, giant world-map screen
- ...

**Props**
- Polaroid (NOV 14): horizontal selfie of Roko + Lulu, handwritten "NOV 14"
- Note (food in the fridge): blue sticky note in Lulu's handwriting
- ...

**Style references (optional)**
- Base Staff: 3-class wardrobe sheet (security / analyst / scientist)
- ...
```

For image-based prompting, end phase 2 with: *"Generate these in Nano Banana / Soul / your tool of choice and upload them back. Name files so I can map them — e.g., `roko.png`, `apartment.png`, `polaroid_nov14.png`. Then tell me which scenes to build prompts for."*

**Stop only when assets are genuinely required.** If the complete story package already contains the needed story, character, prop, and location information, do not stop for uploads; proceed to Phase 3 defaults and then Phase 4.

### Phase 3 — Scope + spatial blocking

Before generating the final TSV:

1. **Confirm scope internally** — default to the full story package / full film. Only output a partial range when the user explicitly asks for specific scenes or shots.
2. **Map filenames to assets when files are uploaded** — flag any missing or extra files. Never auto-assign silently if a filename is ambiguous; ask.
3. **Confirm style override** if one was uploaded; otherwise confirm default style.
4. **Spatial blocking** — if the user explicitly requests image-based spatial blocking, uploads reference images that require precise position mapping, or asks to lock positions before prompting, produce a top-down SVG schema (see [reference/SPATIAL_BLOCKING.md](reference/SPATIAL_BLOCKING.md)) using `visualize:show_widget`. Show character positions, eyelines, prop placement, distances in meters, camera position per shot. Then ask: *"Positions correct? Any edits?"* and iterate until approved.

For complete story package input, do not block Phase 4 on image uploads or top-down schemas unless the user requested that workflow or the available material is too ambiguous to preserve continuity.

### Phase 4 — Generate segmented TSV shot tables

Before writing or revising any TSV shot table, read and follow [reference/GENERAL_TSV_RULES.md](reference/GENERAL_TSV_RULES.md). Those rules apply across all scripts, not only the current project.

For the full scope:
1. Output segmented TSV tables by independently generated video segment. Each segment is one generation unit that can be rendered alone and cut naturally against the previous/next segment.
2. Keep the final written duration of each segment at or under 15 seconds, but avoid underfilled fragments. Prefer fuller, complete video segments that land near 12-15 seconds when the action allows, often around 6-8 shot rows as a soft target. This soft target must never override source-shot preservation: do not combine, delete, or flatten numbered source shots merely to make a segment feel full. If a complete action chain cannot fit clearly under 15 seconds, split into another segment instead of removing beats.
3. Keep `镜号` continuous from 1 through the final shot. Do not restart numbering by scene, act, segment, prompt group, or location.
4. Follow the script/storyboard shot order. If the source provides numbered shots or discrete action beats, preserve each source shot as at least one TSV row unless the user explicitly asks for compression. You may add visible micro-actions inside a row or split a dense source shot into multiple rows, but do not merge separate source shots when that would erase a setup, reaction, failure, reveal, handoff, or payoff. If a script gap would break continuity, fill only visible continuity: action transitions, prop state, blocking, emotional reaction, and natural dialogue connection. Do not invent new plot events.
5. Break action into fine shot rows (one row per discrete action/camera/focal-length/emotion change). Prefer precise short rows over slow overloaded shots.
6. Each row's `推荐时间(0.5-3秒)` must reflect the actual visual workload and help the segment land as a complete clip. Do not distribute time evenly by habit, and never exceed 3 seconds for a single shot row.
7. Each segment's first shot must satisfy the internal **段首启动** requirement: establish the visible scene anchor, character positions, distance/orientation, eyelines, and key prop locations so the segment can start independently.
8. Each segment's final shot must satisfy the internal **段尾状态** requirement: where each story-critical character stops, who holds each key prop, what visible action direction is prepared at the cut, and what emotion state is held at the cut.
9. Between segments, maintain **段间衔接** for independently generated clips: each next segment's first shot must fully restate a fresh start state that matches the previous segment's final spatial state, prop ownership, action direction, and emotional state. Do not rely on the video model knowing the previous segment; avoid continuity wording such as "still", "same place", "same scene", "here", "there", "continues from the previous shot", or "inherits". When a scene reference image is available, restate the visible environment from that image: road/room type, foreground/midground/background anchors, weather/light, surfaces, props, and character positions. State the environment, character positions, key props, eyelines, and emotional state as visible facts inside the new segment.
10. Output only Markdown fenced code blocks containing TSV (tab-separated values). Do not output HTML unless the user explicitly asks for HTML.
11. The TSV header must be exactly by default:

```tsv
镜号	推荐时间(0.5-3秒)	景别与运镜	画面与极致情绪动作描述(中文白话)	[配音台词](纯英文)	专属音效
```

12. If the user explicitly requests another target language, update the dialogue column header and dialogue cells to that language while preserving the six-column schema.
13. Every data row must use real tab characters between columns. Do not use Markdown tables, commas, pipes, or extra columns.
14. The `景别与运镜` column should be technically useful: shot size, camera move, lens/focal length, camera position, handheld rhythm, push/pull distance, or other concise camera details when relevant.
15. The `画面与极致情绪动作描述(中文白话)` column must carry the visual direction: vertical/mobile framing when relevant, blocking, action beats, micro-expressions, eyelines, prop continuity, and concise failure-prevention direction. Do not replace this column's existing cinematographic role with a new schema. The rule labels `段首启动`, `段尾状态`, and `段间衔接` are internal QA terms only; never write these literal labels inside TSV cells. Do not use vague cross-segment scene references such as "同一处", "同一场景", "这里", "该处", or "之前的地点"; describe the actual visible scene again from the scene image or script-visible environment. Do not use negative correction prompts that name an unwanted result; write only the desired visible scene anchors. Preserve prop chains by using stable prop identity names and only adding appearance details that are confirmed by reference images, script, or the latest user revision.
16. The `[配音台词](纯英文)` column contains localized natural English dialogue/VO text by default, with the speaker explicitly named, for example `Girl: "You should eat something warm."` or `VO: "..."`; write `无` if there is no line. Rewrite stiff literal translations into expressions that would naturally occur in an English-speaking context. If the user requests no dialogue, keep the column but write `无` in every row, and convert dialogue intent into visible behavior: gesture, posture, eyeline, breath, hand action, prop interaction, or reaction timing.
17. The `专属音效` column stays in the schema, but use it selectively. Write concrete story-critical or emotion-reinforcing sound cues, or `无` for ordinary ambient sound. Do not write music cues.
18. If the user asks for Seedance-ready output, make the Chinese action-description column detailed enough to serve as a Seedance visual prompt, while still preserving the exact six-column TSV schema above.

#### Reusable TSV rules

- Phase 4 outputs segmented Markdown-fenced TSV by default; output HTML only when the user explicitly asks for HTML.
- Keep the TSV schema fixed to the six required columns, separated by real tab characters only.
- Output one TSV table per independently generated video segment. Repeat the header for every segment.
- Keep `镜号` continuous across all segment tables.
- Use one shot row for one clear action, camera change, or emotional beat; split dense action into more short rows.
- Preserve the source action chain before optimizing density. Numbered source shots and discrete script beats should survive as rows; enrichment means adding spatial, emotional, and prop detail, not deleting beats.
- Keep the final written segment duration at or under 15 seconds and each shot recommendation between 0.5 and 3 seconds. Prefer complete 12-15 second segments, often around 6-8 rows, instead of many short fragments, but never use this as pressure to merge away source shots. Slight rough overages can be compressed only by timing, not by deleting actions.
- For vertical shorts, design every shot for 9:16 mobile framing instead of widescreen composition, but do not repeat "9:16竖屏" in every row. State the aspect ratio once at project/segment level, then express composition through foreground/midground/background, left/right/top/bottom placement, depth, distance, and eyelines.
- Turn casual user notes into executable cinematography: shot size, camera move, foreground/midground/background, blocking, action, eyeline, and micro-expression.
- State the position, orientation, and eyeline of every story-critical character in frame, even if the user only names one of them.
- For every segment, satisfy the internal **段首启动** and **段尾状态** requirements when needed for independent generation and editing continuity, but do not print those literal labels in TSV cells.
- Preserve **段间衔接** by making the next segment's first row a self-contained start state that visibly matches the previous segment's final row: environment anchor, character location, prop ownership, action direction, eyelines, and emotional state are all restated as fresh facts for that independent clip. Never write deictic scene shortcuts like "同一处", "同一场景", "这里", or "之前"; if a scene image exists, re-describe the called scene from that image.
- Put localized natural English dialogue with the speaker name, or `无`, in the dialogue column by default; if the user names another target language, change the column header and dialogue content accordingly while still naming the speaker.
- Write desired visible facts instead of negative corrections. Avoid naming unwanted locations, actions, or props inside TSV cells; replace "not X" with the positive target anchor/action.
- Preserve the latest user revision for key prop chains: stable prop identity, material/appearance only when confirmed, handoff direction, current holder, and payoff/reveal method. If an object is not active in the current beat and its mention could cause continuity drift, omit it.
- Preserve confirmed spatial references; do not invent new positions for doors, counters, shelves, aisles, or other core layout anchors.
- Merge asset requests by Master Environment first; split local environment assets only for distinct visual styles, time changes, key plot props, or areas the master cannot show clearly.
- During revisions, rewrite only the affected shot rows and continuity-dependent rows unless the user requests a full-table rewrite.
- During complete-story-package output, default to the full film unless the user explicitly narrows the scope.
- Fill incomplete script material only for visible continuity and natural dialogue flow; do not add new plot events.
- When dialogue is banned or unnecessary, translate its narrative function into action while keeping the dialogue cell `无`; do not remove the beat just because the line is removed.
- Make sound cues concrete and audible when they matter to story, action, or emotional pressure; use `无` for ordinary ambient sound and avoid vague labels like "tense music."
- Before final TSV output, run an internal 对抗式审查: compare source shot count/discrete beats against output rows, check whether any source action was deleted or flattened, verify no segment exceeds 15 seconds and no row exceeds 3 seconds, confirm the first row of each segment is independently generatable, and confirm no forbidden shortcut wording or negative correction prompt remains.

#### Rule update workflow

- When the user returns a revision plan after seeing generated output, separate reusable long-term rules from current-project-only fixes.
- Before changing this skill, run an **对抗式审查**: check conflicts with the six-column TSV schema, field responsibilities, 15-second segment cap, 3-second shot cap, segmented output, and existing spatial-continuity rules.
- Check redundancy before adding text. Merge with an existing rule when a new rule says the same thing in different words.
- Check bloat before adding text. Do not add one-off project preferences to long-term rules unless the user explicitly confirms they should become general rules.
- Report the proposed rule change, conflicts found, redundant text to remove/merge, and project-only items back to the user. Only archive confirmed reusable rules into the long-term skill.

## Hard rules

- **Handles renumber per prompt.** `@image1` in scene 21 = Roko; `@image1` in scene 14 may = a different character. Each prompt block declares its own handles.
- **Output language:** Phase 4 TSV headers and visual/action descriptions are Chinese. Dialogue/VO cells are pure English only. Asset lists may be Chinese if the user asks for Chinese.
- **Duration/aspect:** Use the user's requested aspect ratio (for example 9:16 vertical shorts). Phase 4 TSV rows do not need a footer such as `15秒。21:9。`. Segment duration is a hard cap: each independently generated segment must be at or under 15 seconds, and each shot row must be 0.5-3 seconds.
- **Director assignment:** skip entirely unless user requests it. No `dir-badge`, no palette switching — default to `pal-red` color scheme.
- **Style block:** use the [default style block](reference/STYLE_BLOCK.md) verbatim (with the appropriate scene-type variant) unless user uploads a custom one in phase 1.
- **Lighting is ALWAYS practicals-only.** No film fill light, no reflectors, no softboxes, no LED strips, no neon. Camera shoots from the shadow side. This is non-negotiable. See [STYLE_BLOCK.md](reference/STYLE_BLOCK.md).
- **Camera tracks emotion.** Nervous handheld for anger/tension; smooth handheld breathing for calm; static + slow push for shock/revelation. See [CAMERA_EMOTION.md](reference/CAMERA_EMOTION.md).
- **No generic emotion.** Every emotional direction must decompose into muscles, breath, eyes, skin. See [MICRO_BEATS.md](reference/MICRO_BEATS.md).
- **Top-down schema before prompting** only when the user requests spatial blocking, uploads reference images that require precise geometry, or the available material is too ambiguous to preserve continuity. See phase 3.
- **Metadata inference:** project title, "Prepared for [name]", scene scope — infer from script + user context (memory, prior turns). If genuinely unclear, ask one short clarifying question; otherwise proceed.
- **Never auto-assign images to handles silently.** If a filename is ambiguous, ask before assembling prompts.
- **Iteration = TSV revisions + cautious rule updates.** When the user requests changes after delivery, revise only the affected shot rows and any continuity-dependent rows unless they ask for a full-table rewrite. If the user asks to update the skill's rules, perform 对抗式审查 and get confirmation before archiving the reusable rule into this skill. Do not regenerate unrelated rows.

## Cinematography mandate

For every prompt, you must:
- Pick the lens (35mm wide / 50mm dialogue / 85mm or 100mm tight emotional / 45mm macro / aperture F1.4 for shallow DOF)
- Pick the camera move synced to the focal character's emotion (see CAMERA_EMOTION.md)
- Block the actors with concrete spatial relationships from the approved top-down schema ("Roko 2m from Gandelfina, Rein 1.5m behind Roko, partially occluded")
- Direct the performance with numbered emotional beats (① ② ③ ④ ⑤) — micro-beats, breath, eye-line shifts, weight shifts, suppressed emotion
- Specify lighting source by source (windows, practicals, screens) and forbid film fill light explicitly
- Specify what's in the background and what the extras are doing — never empty backgrounds in populated locations
- Add `⚠️` warnings for failure modes the prompt is most likely to mess up; use `⚠️⚠️⚠️` for critical-critical (handle contamination, identity drift, light spill, prop misplacement, focus drift on inserts)

See [reference/PROMPT_PATTERNS.md](reference/PROMPT_PATTERNS.md) for the full pattern library.

## Example flow

**User (turn 1):** uploads `ARTIFACT_script.pdf`, says "build me a shotlist for this"

**Claude:**
- Reads script
- Outputs asset list (characters/locations/props), brief descriptions
- Ends with the upload + scope prompt

**User (turn 2):** uploads `roko.png`, `gandelfina.png`, `apartment.png`, `polaroid.png`, says "scenes 21 and 23"

**Claude:**
- Confirms image → asset mapping ("Got it: roko.png → Roko, gandelfina.png → Gandelfina, apartment.png → Old Apartment, polaroid.png → Polaroid NOV 14. Building scenes 21 and 23.")
- Renders top-down SVG schema for any multi-character scene; asks for approval
- After approval: outputs a TSV shot table in a Markdown code block using the required six-column schema.

## File map

- `templates/HTML_TEMPLATE.md` — legacy HTML scaffold; use only if the user explicitly asks for HTML
- `reference/STYLE_BLOCK.md` — the default Chinese style block (Lubezki × Deakins, contre-jour, 60:30:10, practicals-only) with variants by scene type
- `reference/PROMPT_PATTERNS.md` — the full prompt structure: handles, spatial blocking, multi-shot 【镜头N】 syntax, dialogue rules, failure-mode warnings
- `reference/GENERAL_TSV_RULES.md` — global TSV revision rules: visible-only prompting, real spatial anchors, dialogue/action timing, no music cues, color-markup revisions
- `reference/CAMERA_EMOTION.md` — camera movement-to-emotion mapping, lens selection, shot duration rules, phased emotional arcs
- `reference/MICRO_BEATS.md` — the performance micro-beat catalog by emotion (anger, anxiety, sadness, control, heaviness, etc.)
- `reference/SPATIAL_BLOCKING.md` — top-down schema rules: when to draw, what goes on it, how to translate it into the prompt
- `reference/PROMPT_DENSITY.md` — how to group shot rows into 15-second prompts
- `reference/PLAN_TYPES.md` — shot-plan taxonomy and badge classes

目前的独立 Skill 专门处理逐镜首帧，和九宫格构图预览分开使用。

---
name: storyboard-keyframes
description: Extract opening-frame prompts from shot scripts and generate one confirmed 9:16 storyboard frame at a time. Use when the user provides a shot script and scene or character references, wants the first-frame static image prompt, or asks to update and reconcile creative rules for iterative storyboard generation.
---

# Storyboard Keyframes

Use this skill for one job only: turn each shot into a static opening frame, then generate that frame as a single vertical image.

## Workflow

### 1. Extract the first frame

- Read the shot script and the provided references, then load the matching scene, character, and prop reference images as actual generation inputs.
- For each shot, extract the opening static frame only.
- Use this output order:

```text
角色: [代指] 景别: [对标分析] 背景: [背景名]. [全量背景描述，不含人物]. 描述: 写实风格. [道具锁定属性描述]. 穿着[详细服饰描述]的 [角色名] [姿态/手势/肌肉表情/视线方向].
```

- Keep the background sentence strictly about the environment.
- Put人物、服饰、动作、神态、视线 only after `描述: 写实风格.`.
- If no prop reference is provided, omit the prop clause.
- Reference images must be loaded and used as real inputs, not only mentioned in the text prompt.
- Each frame must bind the matching scene, character, and prop reference images before generation.
- Reference images control appearance and spatial structure; the script controls action, emotion, and camera intent.
- Same-shot groups reuse the same reference set unless the user explicitly changes it.
- If the reference images and script conflict, pause and ask instead of inventing a compromise.
- After extraction, show the prompt for confirmation before generating.

### 2. Generate one frame at a time

- After confirmation, generate only one frame as a single `9:16` vertical image.
- Use the confirmed prompt unchanged.
- Do not build 3x3 grids, blank panels, or local composite fallbacks.
- Do not merge multiple shots into one image.

## Failure Handling

- If one frame fails, record the shot number and continue to the next frame.
- After the batch, retry failed frames with the original prompt up to 3 rounds.
- If a frame still fails after 3 rounds, report the shot number and `图像引擎 ServerError`.
- Do not silently simplify the prompt unless the user explicitly asks for a rewrite.

## Rule Updates

- If the user says `更新规则`, stop generation and collect all creative requirements mentioned so far.
- Compress them into short, standalone iron rules.
- Each iron rule must be one sentence, concise, and understandable outside the current case.
- Check for conflicts between new rules and existing rules.
- If rules conflict, list the conflict and the possible resolution options.
- Do not write updated rules into the skill file until the user explicitly confirms the chosen rules and conflict resolution.
- After confirmation, append non-conflicting rules and replace only the rules the user approved to change.

## What To Ignore

- Ignore nine-grid composition rules.
- Ignore blank-slot logic.
- Ignore global-script initialization flow.
- Ignore any request to auto-build a composite image unless the user explicitly asks for it.