@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/brewreel/docs/custom-shot.md

@@ -1,103 +0,0 @@
-# 强模型的自由镜头
-
-便宜模型不要用 custom。便宜模型继续只挑镜头、填文字。
-
-这一节给会写代码的模型（Codex、Claude Code 这类）。模板镜头保住下限，也封住了上限。你想自己画一镜时，用 `custom`：分镜里只留这一镜的文字和数字，画面写在片子目录的组件里。
-
-只在本地技能模式开放。DeepSeek 插件看到 `custom` 会直接拒绝，原话是：自由镜头只能在本地用，插件里不运行模型自己写的代码。
-
-## 什么时候用
-
-- 现有镜头做不到你要的那一下动作，而且你能自己写 React。
-- 只要换一句文案、改一个已有字段，不要用 custom，继续填表。
-
-## 分镜
-
-和 `storyboard.json` 同级建 `shots/名字.tsx`。不要把组件放进 `template/`。
-
-```json
-{
-  "type": "custom",
-  "dur": 4,
-  "caption": "先留这一句，\n{少一步}",
-  "mood": 0.4,
-  "component": "shots/keyword-toss.tsx",
-  "slots": {"keyword": "少一步"}
-}
-```
-
-- `component` 只能是 `shots/` 加文件名，后缀 `.tsx`。不要写 `..` 或绝对路径。
-- `slots` 放这一镜要上屏的文字和数字。可以嵌套，最多 4 层。
-- `params` 不要写。写了也不会传给组件。
-- 字幕、广告法、带单位的数字、slots 里的纯数字，和别的镜头走同一套校验。效果数字（`from` / `to` / `value` / `percent` / `delta`）必须在 `meta.facts` 里有真来源，标了示例的 fact 不算。
-
-## 组件拿到什么
-
-```tsx
-import {CutShape, FONT, fitLine, toss, type CustomShotProps} from '../../custom-api';
-
-export default function KeywordToss({slots, theme, geo, t, dur, frame, fps, frames, lang, beat, index}: CustomShotProps) {
-  const word = typeof slots.keyword === 'string' ? slots.keyword : '';
-  const pose = toss(520).at(t);
-  return (
-    <div style={{position: 'absolute', left: geo.safe.x0 + 40, top: geo.safe.y0 + 200}}>
-      <CutShape seed={1} w={520} h={220} fill={theme.card}>
-        <div style={{fontFamily: FONT, fontSize: fitLine(word, 400, 72, 40), color: theme.cardText}}>{word}</div>
-      </CutShape>
-    </div>
-  );
-}
-```
-
-`props`：
-
-| 字段 | 是什么 |
-|---|---|
-| `slots` | 分镜里写的文字和数字 |
-| `theme` | 当前主题令牌（`card`、`cardText`、`accent`、`accentInk`、`line` …） |
-| `geo` | 画幅几何：`safe`、`card`、`cap` |
-| `frame` / `fps` / `t` / `dur` / `frames` | 镜头内帧号、帧率、秒、时长、总帧数 |
-| `lang` | `zh` 或 `en` |
-| `beat` | 一拍多少秒 |
-| `index` | 镜头下标，从 0 起 |
-
-`../../custom-api` 在出片拷贝之后才指向 `template/src/shots/custom-api.ts`。作者就按这个相对路径写。
-
-可以从 `custom-api` 拿到：
-
-- 字体 `FONT` / `SERIF` / `MONO`，`fitLine` / `fitSize`
-- 安全区 `SAFE` / `CARD` / `CAP` / `MAIN`，`geometryOf`
-- `useTheme`、`alpha`、`mixHex`、`toneColor`、`moodColors`
-- 纸纹 `PaperGrain`、手剪形状 `CutShape`、纸屑 `Confetti`
-- `toss` / `entrance` / `leave`、`floatMotion`、`settle`、`shadowByHeight`
-
-另外只允许 `import` `react`、`remotion`，或同目录的 `./` 辅助文件。不要直接去引 `template/src/core` 或别的正式镜头。
-
-## 字为什么必须放 slots
-
-广告法、数字出处、字数校验只看分镜，看不到你写进 JSX 的句子。出片前会扫组件源码：JSX 和字符串里的中文、英文句子（两个以上单词）、直接渲染的数字（`{12}`、标签文本里的数字）都报错，带文件和行号。改法是挪进 slots。
-
-坐标、`style` 里的数字、`seed` / `opacity` / `width` 这类属性、函数体里的计算，不算写死的文案。
-
-## 怎么出片
-
-和普通片子同一条命令。Remotion 只打包 `template/src`，所以出片时会：
-
-1. 先扫 `shots/*.tsx`。有写死的字就停下，不渲染。
-2. 拿到渲染锁之后，把 `shots/*.tsx` 拷到 `template/src/shots/_custom/<片子哈希>/`（这个目录在 `.gitignore` 里）。
-3. 生成注册表，跑 `tsc --noEmit`。类型错误按原来的文件和行号报，不进渲染。
-4. 渲染。组件抛错、字超出安全区、整屏空帧，都算到具体那一镜。动效停死用口播那套帧差（只量主体区）再查一遍。
-5. 出片结束把拷贝删掉，注册表还原成空的。
-
-样例：`examples/custom/storyboard.json`（关键词抛起落定 + 一条自己画的数字条）。
-
-## 常见报错
-
-| 报错 | 原因 |
-|---|---|
-| `shots/某.tsx:12:3` 写死了文字，怎么改：挪进 slots | JSX 或字符串里有中文 / 英文句子 |
-| `第 N 镜（custom）shots/某.tsx(10,7): error TS…` | 类型不对。看这一行，不要改 template |
-| `第 N 镜（custom）shots/某.tsx:8：boom` | 渲染时组件抛错。行号在文件里 |
-| `第 N 镜（custom）…出了 x150–930（组件 shots/某.tsx:9）` | 字画出了安全区 |
-| `第 N 镜（custom）…动效停死` | 这一镜有 0.4 秒几乎不动。加持续的位移，不要只靠很小的浮动 |
-| `自由镜头只能在本地用，插件里不运行模型自己写的代码` | 在插件里用了 custom。改回本地跑，或换成模板镜头 |

package/skills/brewreel/docs/faq.en.md

@@ -1,90 +0,0 @@
-# FAQ
-
-[← Back to README](../README.en.md) · [中文](faq.md)
-
-
-<details>
-<summary><b>Where does the background music come from?</b></summary>
-
-`scripts/make_bgm.py` synthesizes it on the spot with numpy / scipy: instruments, harmony, melody, mixing and mastering are all in the script, with no external sample library. `make.mjs` passes it the start time of every shot; each shot is one section, and its writing is picked by position and mood (hook, tension, lift, groove, peak, ending). Notes land on whole or half beats, and every shot change gets a cymbal or a fill, so the music follows the cuts. Loudness is normalized to -16 LUFS with ffmpeg's ebur128 (approximated by RMS if ffmpeg can't be found).
-
-Because it is synthesized on the spot, there are no copyright issues. When you publish on a platform such as Douyin, you can also swap in music from the platform's own library: render a silent version with `--no-bgm`, then add music on the platform.
-
-</details>
-
-<details>
-<summary><b>Is there a voice-over?</b></summary>
-
-Yes. v0.5.0 added MiniMax text-to-speech, and from v0.5.1 you can also use Alibaba Cloud (Model Studio CosyVoice) or Volcengine (Doubao speech). Write `meta.voice` and a `vo` per shot, set the matching environment variable (`MINIMAX_API_KEY` / `DASHSCOPE_API_KEY` / `VOLCENGINE_TTS_API_KEY`), and render. Shot lengths follow the narration, subtitles light up word by word, and the music ducks under speech. Without a key, add `--voice-provider mock` to preview the rhythm with a placeholder voice. Without `meta.voice` there is no voice-over, as before. When you publish with an AI voice, tick the platform's AI-generated content declaration as it requires.
-
-</details>
-
-<details>
-<summary><b>Why illustrations instead of real footage?</b></summary>
-
-The project does not generate images that look like "real photography". Without merchant material, the video falls back to components and simple line-drawn illustrations and labels them clearly, rather than passing them off as real photos. With the merchant's real photos (store, finished product, price list), use the real-photo shots; the result is far more convincing. Whether a photo is authorized goes on the human-review list.
-
-</details>
-
-<details>
-<summary><b>Can I publish the video as rendered?</b></summary>
-
-Not yet recommended. Treat the video as a first draft: watch the whole thing, fix awkward copy and repeated selling points, then publish. For industry videos, check every price, condition, date, opening hour and distance on screen against the brief, one by one. See ["Known limitations"](limitations.en.md) for test results.
-
-</details>
-
-<details>
-<summary><b>Does it cost anything?</b></summary>
-
-The project is free and open source. Model calls use your own API key and are billed by the provider for what you use; the project handles no money. Rendering runs on your own computer. The rendering engine Remotion charges for-profit organizations with 4 or more people; see the next question.
-
-</details>
-
-<details>
-<summary><b>Does Remotion need a license? What about upgrading to 5.0?</b></summary>
-
-Remotion is source-available, not open source: free for individuals, for-profit companies with 3 or fewer people, and non-profits (including commercial use); **for-profit organizations with 4 or more people must purchase Remotion's Company License**, see <https://www.remotion.dev/license>. This repo's Apache-2.0 license does not change Remotion's own license terms; details in `THIRD_PARTY_LICENSES.md`.
-
-This repo pins `remotion` / `@remotion/cli` to `4.0.529`. After upgrading to 5.0, Remotion's free tier requires passing a `licenseKey` in config (individuals / companies with ≤3 people / non-profits use `"free-license"`); see the [Remotion license page](https://www.remotion.dev/license). Read `THIRD_PARTY_LICENSES.md` before upgrading.
-
-</details>
-
-<details>
-<summary><b>Anything to watch for on Windows?</b></summary>
-
-- x64 only.
-- Set environment variables the PowerShell way, `$env:LLM_API_KEY="..."`, not `export`.
-- `--props` fails with "neither valid JSON": the Windows shell mangles quotes inside JSON strings, so storyboards are always passed as a file path (`--props=./storyboard.json`), never as an inline JSON string. `make.mjs` / `llm_make.py` already do it the file-path way.
-- `make.mjs` calls `python` on Windows and `python3` on macOS / Linux; if yours lives elsewhere, point the `PYTHON` environment variable at it.
-
-</details>
-
-<details>
-<summary><b>Do I need a system ffmpeg?</b></summary>
-
-Not for promo videos (the paragraph below is about them). **Talking-head B-roll needs a full ffmpeg**: `pip install imageio-ffmpeg` is the easiest, or put your own ffmpeg on PATH or point the `FFMPEG` environment variable at it; the slim build bundled with Remotion lacks filters, and the command says so at the start.
-
-For promo videos, the contact sheet `sheet.png` is rendered by Remotion's `Sheet` composition (one frame per second of the video, tiled as thumbnails, each labelled with its time and shot number). The check frames `check/*.png` are extracted from the video with Remotion's bundled ffmpeg, and any frame that can't be extracted is rendered as a Remotion still instead. If you set `FFMPEG=/path/to/ffmpeg`, the sheet is tiled with its `tile` filter first and falls back to the `Sheet` composition if that fails. A missing sheet never affects `video.mp4`; the terminal prints the reason.
-
-</details>
-
-<details>
-<summary><b>A platform's compositor package is missing from the lockfile after <code>npm install</code>?</b></summary>
-
-`package-lock.json` should contain 7 `@remotion/compositor-*` platform packages (win32-x64-msvc / darwin-arm64 / darwin-x64 / linux-x64-gnu / linux-x64-musl / linux-arm64-gnu / linux-arm64-musl). If the one for your platform is missing, delete `template/node_modules` and `template/package-lock.json` and re-run `npm install` on the target platform (a known npm optional-dependency quirk: a lockfile generated on one platform doesn't always include every platform).
-
-</details>
-
-<details>
-<summary><b>Chrome Headless Shell fails to download?</b></summary>
-
-`npx remotion browser ensure` may not be able to reach Google's download endpoint from some networks. Download it manually and point to a local Chrome / Chromium install with `--browser-executable` (rendering results may differ slightly between Chrome versions).
-
-</details>
-
-<details>
-<summary><b>Why the name "BrewReel" (精酿, "craft brew")?</b></summary>
-
-Good beer comes from a good recipe, even with ordinary ingredients. A strong model first gets a video style right; its layout, motion, pacing and rules are then written down as a "recipe": ready-made components plus a validator. A low-cost model like DeepSeek is the ordinary ingredient: it follows the recipe, fills in a storyboard, and brews a video at the same level. The docs use the same words throughout: style pack = recipe, deriving a style from a reference video = crafting a recipe, rendering = brewing. Paths and fields such as `styles/`, `meta.style` and `distill/` keep their original names.
-
-</details>

package/skills/brewreel/docs/faq.md

@@ -1,90 +0,0 @@
-# 常见问题
-
-[← 返回 README](../README.md) · [English](faq.en.md)
-
-
-<details>
-<summary><b>背景音乐是怎么来的？</b></summary>
-
-`scripts/make_bgm.py` 用 numpy / scipy 现场合成：乐器、和声、旋律、混音、母带都在脚本里，不用外部素材库。`make.mjs` 把每个镜头的起点传给它，一个镜头一个段落，按位置和情绪选写法（钩子、紧张、抬升、律动、高潮、收尾），音符落在整拍 / 半拍，镜头切换处有镲或加花，所以音乐是卡着镜头走的。响度用 ffmpeg 的 ebur128 校到 -16 LUFS（找不到 ffmpeg 就按 RMS 近似）。
-
-因为是现场合成的，没有版权问题。正式发抖音这类平台时，也可以换成平台曲库里的音乐：出片时加 `--no-bgm` 出静音版，再在平台里配乐。
-
-</details>
-
-<details>
-<summary><b>有没有配音？</b></summary>
-
-有。v0.5.0 起接了 MiniMax 语音合成，v0.5.1 起也可以用阿里云（百炼 CosyVoice）和火山引擎（豆包语音）。分镜里写 `meta.voice` 和每镜的 `vo`，设好对应的环境变量（`MINIMAX_API_KEY` / `DASHSCOPE_API_KEY` / `VOLCENGINE_TTS_API_KEY`）后出片即可；镜头时长跟着旁白走，字幕逐字点亮，配乐在人声处自动压低。没有 key 时加 `--voice-provider mock` 用占位音预览节奏。不写 `meta.voice` 就和以前一样没有配音。用 AI 配音发布时，记得按平台要求勾选 AI 生成内容声明。
-
-</details>
-
-<details>
-<summary><b>为什么画面是插画，不是实拍？</b></summary>
-
-项目不生成「看起来像真实拍摄」的拟真图片。没有商家素材时，画面用组件和简笔插画兜底，并明确标注，不冒充实拍。有商家实拍照片（门店、成品、价目表）时，用实拍镜头放进去，可信度会好很多；照片是否授权会列进「需人工复核」。
-
-</details>
-
-<details>
-<summary><b>成片能直接发吗？</b></summary>
-
-还不建议。把成片当初稿：看完整片，改掉别扭的文案和重复的卖点再发；行业片把画面上每个价格、条件、日期、营业时间、距离逐条和简报对一遍。实测情况见[「已知限制」](limitations.md)。
-
-</details>
-
-<details>
-<summary><b>要花钱吗？</b></summary>
-
-本项目免费开源。模型调用用你自己的 API Key，按用量在服务商那边结算，项目不经手任何费用；渲染在你自己的电脑上跑。渲染引擎 Remotion 对 4 人及以上的营利组织收费，见下一条。
-
-</details>
-
-<details>
-<summary><b>Remotion 要授权吗？升级到 5.0 要注意什么？</b></summary>
-
-Remotion 是源码可见、非开源的软件：个人、3 人及以下的营利公司、非营利组织可免费使用（含商用）；**4 人及以上的营利组织需要购买 Remotion 的 Company License**，详见 <https://www.remotion.dev/license>。本仓库的 Apache-2.0 不改变 Remotion 自己的许可条件，见 `THIRD_PARTY_LICENSES.md`。
-
-本仓库把 `remotion` / `@remotion/cli` 钉在 `4.0.529`。升级到 5.0 后，Remotion 免费层需要在配置里传 `licenseKey`（个人 / 3 人以下公司 / 非营利填 `"free-license"`），详见 [Remotion 许可证页](https://www.remotion.dev/license)。升级前请先看 `THIRD_PARTY_LICENSES.md`。
-
-</details>
-
-<details>
-<summary><b>Windows 上要注意什么？</b></summary>
-
-- 只支持 x64。
-- 环境变量用 PowerShell 写法 `$env:LLM_API_KEY="..."`，不是 `export`。
-- `--props` 报「neither valid JSON」：Windows shell 会吃掉 JSON 字符串里的引号，所以分镜一律用文件路径传入（`--props=./storyboard.json`），不要在命令行拼 JSON。`make.mjs` / `llm_make.py` 已经是文件传入方式。
-- `make.mjs` 在 Windows 上调 `python`，在 macOS / Linux 上调 `python3`；装在别处的，设环境变量 `PYTHON` 指过去。
-
-</details>
-
-<details>
-<summary><b>需要装系统 ffmpeg 吗？</b></summary>
-
-做宣传片不需要（下面这段说的是宣传片）。**口播配画面要完整版 ffmpeg**：`pip install imageio-ffmpeg` 最省事，或者自己装 ffmpeg 放进 PATH、设环境变量 `FFMPEG` 指向它；Remotion 自带的精简版缺滤镜，不够用，缺的话命令一开头就会说。
-
-宣传片这边：拼图 `sheet.png` 由 Remotion 的 `Sheet` 合成直接出图（整片每秒一帧，缩略排成网格，每格下面标时间和镜号），检查帧 `check/*.png` 用 Remotion 自带的 ffmpeg 从成片里抽，抽不出来会改用 Remotion 单帧补。设了环境变量 `FFMPEG=/path/to/ffmpeg` 时，拼图会先用它的 `tile` 滤镜拼，失败再退回 `Sheet` 合成。拼图没生成不影响成片 `video.mp4`，终端会打印原因。
-
-</details>
-
-<details>
-<summary><b><code>npm install</code> 后 lock 文件里缺某个平台的 compositor？</b></summary>
-
-`package-lock.json` 里应该有 7 个 `@remotion/compositor-*` 平台包（win32-x64-msvc / darwin-arm64 / darwin-x64 / linux-x64-gnu / linux-x64-musl / linux-arm64-gnu / linux-arm64-musl）。如果你的平台对应的条目缺失，删掉 `template/node_modules` 和 `template/package-lock.json` 后在目标平台上重新 `npm install`（这是 npm 已知的可选依赖解析问题，跨平台生成的 lock 文件不一定包含所有平台）。
-
-</details>
-
-<details>
-<summary><b>Chrome Headless Shell 下载失败？</b></summary>
-
-国内网络环境下 `npx remotion browser ensure` 可能连不上 Google 的下载地址。可以手动下载后用 `--browser-executable` 指定本地 Chrome / Chromium 路径（不同 Chrome 版本渲染结果可能有细微差异）。
-
-</details>
-
-<details>
-<summary><b>为什么叫「精酿」？</b></summary>
-
-好酒靠的是配方，原料普通也能酿好。先让强模型把一种视频风格做到位，再把它的版式、动效、节奏和规则写成一份「配方」——现成组件加校验脚本；之后 DeepSeek 这类便宜模型就是普通原料，照着配方填分镜，就能酿出同一水准的片子。文档里的说法随之统一：风格包 = 配方，从参考视频提炼风格 = 调配方，出片 = 开酿。`styles/`、`meta.style`、`distill/` 这些路径和字段沿用原名。
-
-</details>

package/skills/brewreel/docs/features.en.md

@@ -1,157 +0,0 @@
-# Features
-
-[← Back to README](../README.en.md) · [中文](features.md)
-
-
-## Three recipes
-
-A storyboard picks its recipe with `meta.style`; without it you get the default `cards`. Each recipe is a "style pack" with its own design tokens, shot components, validation rules and narrative templates: docs, rules and examples live in `styles/<id>/`, code in `template/src/styles/<id>/`.
-
-| Recipe | Aspect | Fits |
-|---|---|---|
-| `cards` (default) | 9:16 | A single selling point, a how-it-works flow, UI demos, physical products and stores, price lists |
-| `quiz` | 9:16 | A common misconception that can be framed as one multiple-choice question with exactly one right answer ("What does X actually mean?") |
-| `journey` | 4:5 (default) / 9:16 | 4–6 clear categories, features or stops worth touring along one route |
-
-How to choose: one multiple-choice question → `quiz`; 4–6 categories to tour → `journey`; everything else → `cards`, i.e. leave it out. `quiz` only supports 9:16. Its own shots (question, reveal, meaning card, etc.) can only be used in `quiz` and do not mix with the 19 `cards` shots; validation blocks a storyboard that mixes them.
-
-```json
-{ "meta": { "style": "quiz", "industry": "software", "lang": "zh" }, "shots": [ ... ] }
-```
-
-<details>
-<summary><b>What each recipe looks like</b></summary>
-
-**`cards`** (default, ready, 9:16): gradient background, a centered white card and bold outlined captions, one point per shot. Fits a single selling point, a how-it-works flow, UI demos, physical products and stores. Its 19 shots and samples for all six industries are in `examples/`. The six default themes keep the outlined captions: `warm-emotion`, `tech-dark`, `fresh-light`, `business-blue`, `festival-red`, `mono-premium`. Twelve optional palettes are listed in [`styles/cards/THEMES.md`](../styles/cards/THEMES.md).
-
-**`quiz`** (ready, 9:16): the skin is a "marked answer sheet": dot-grid paper with a vermilion margin line, a deep primary plus an amber highlighter and a vermilion marking pen, monospaced labels, and cards with small radii and solid hard shadows. Three palettes (`sage-pine` grey-green paper + pine, the default; `rice-soy` rice paper + soy brown, the food default; `ash-teal` grey paper + deep teal) and three voices (`phraseTitle.params.voice`: `exam`, `chat`, `show`). Two original characters, one short and one tall: a host in headphones (the ear cup lights up and sends out sound waves when it clicks) and a buddy in a backwards cap. The beat is: circle a common misconception in red pen → an A/B/C question on an answer sheet with a stopwatch counting 3 → reveal with a check mark → a dictionary card flips in (the misreading struck through with a wavy line, numbered meanings) → replay the clip and stamp it → a short scene acting it out → a comment prompt → a stamp wipe that lands on the closing card. Fits products with a common misconception that can be framed as one multiple-choice question: what a foreign phrase really means, a software feature people often misread (e.g. "Archive = deleted?"), or why a dish is made the way it is. See [`styles/quiz/`](../styles/quiz/README.md); templates and length limits are in `styles/quiz/recipes.md` (Chinese); five sample storyboards are in `styles/quiz/examples/`.
-
-**`journey`** (ready, 4:5 by default, 9:16 supported): the skin is "travel stationery" on a "layered paper-cut" city. Our own mascot, a red-panda cub on a hover board, crosses the city in one continuous take (far, middle and near layers and the character are each a sheet of paper with a hard offset paper shadow). It opens on split-flap station letters; a full-width ticket across the top shows the route, the stops and the current stop; each stop throws in an airmail-bordered postcard with a representative title, which is then "posted" into that stop on the ticket. Palette: postal green + neon lime + graphite outlines (`post-green`; `plum-ticket` is a wine-red alternative). One district per content category, three backdrops (modern city, low-rise street, or an old town with white walls, tile roofs and a stone bridge) and 14 district types picked by content (crossing, postbox, ticket check, photo booth, bus stop, night platform, phone, home, cafe, market, city gate, stone bridge, teahouse, lanterns), each with its own station- or mail-themed gag; the sky goes from day to night, and at the last stop the ticket from the top slides to the centre and unfolds, gets a hole punched in its stub and flips over to show the product name and slogan; at most one number, printed on the front beside the finish flag. Fits products with many clear categories: channels of a content platform, features of an app, stops on a sightseeing route. See [`styles/journey/`](../styles/journey/README.md); per-beat fields and length limits are in `styles/journey/recipes.md` (Chinese); three sample storyboards are in `styles/journey/examples/`.
-
-</details>
-
-## Voice-over (MiniMax / Alibaba Cloud / Volcengine)
-
-Write `meta.voice` in the storyboard and one `vo` line (narration) per shot, and the video comes out with a voice-over. Without them you get a video with no voice, exactly as before.
-
-- **Three providers**: set `meta.voice.provider` to `minimax` (MiniMax), `aliyun` (Alibaba Cloud Model Studio CosyVoice) or `volcengine` (Volcengine Doubao speech); each defaults to a male announcer voice, and voices can be changed. MiniMax has been rendered with a real key; Alibaba Cloud and Volcengine follow the official docs and are tested with fake responses, **not yet with a real key**.
-- **The voice is the timeline**: `make.mjs` first has text-to-speech read each line and gets per-word timing, then sets each shot with narration to "0.15 s + narration + 0.35 s", aligned to the beat. The cheap model doesn't have to work out durations.
-- **Word-by-word subtitles**: subtitles light up word by word with the voice, in all three recipes. You can also show the whole line at once (`line`) or skip narration subtitles (`off`). In `cards`, a shot with `vo` and no `caption` gets its subtitle from the narration.
-- **Music makes room**: the background music ducks by about 10 dB under speech and comes back after.
-- **Cached to save money**: the same line (same voice, speed, emotion and text) is synthesized only once and cached in the user folder `~/.cache/brewreel/tts` (change it with the `BREWREEL_TTS_CACHE` environment variable). Changing visuals or re-rendering costs nothing more; `manifest.json` records the characters billed for the run.
-- **Preview without a key**: add `--voice-provider mock` when rendering to check rhythm with an offline placeholder voice (not a deliverable); add `--no-voice` for a version without narration.
-- **Keys**: read only from environment variables, never written to any file or log: `MINIMAX_API_KEY` for MiniMax, `DASHSCOPE_API_KEY` for Alibaba Cloud, `VOLCENGINE_TTS_API_KEY` (or the legacy `VOLCENGINE_TTS_APP_ID` + `VOLCENGINE_TTS_ACCESS_TOKEN`) for Volcengine; optional hosts and workspace IDs are in `SKILL.en.md`. Custom endpoints must be https.
-
-Which provider to pick, how to write it, recommended voices and length limits are in the "Voice-over" section of `SKILL.en.md`; each recipe has a sample storyboard with voice-over (`styles/*/examples/*voice*.json`).
-
-## Six industry packs
-
-| `meta.industry` | Description |
-|---|---|
-| `software` (default) | Apps, SaaS, tools |
-| `food` | Restaurants, stores, single-item launches |
-| `ecommerce` | Physical goods |
-| `education` | Adult vocational training (not K12) |
-| `beauty` | Non-medical beauty services |
-| `travel` | Tourism, lodging, homestays |
-
-If the product is outside these six, do not invent `meta.industry`. Tell the user there is no dedicated compliance pack; you may borrow the closest row for layout and must say which one; set `general` to run only the shared advertising-law rules (see `SKILL.en.md`, "When the product is not in the six industries"). `general` is not a seventh pack, and a passing validation is not industry compliance. Driver-assistance wording, drugs, medical devices, and finance need a qualified human review.
-
-In `rules.json`, `enabledShots` is a recommendation and the list that reopens shots `_base` disabled. It is not a whitelist. Validation blocks only the merged `disabledShots`.
-
-`meta.lang`: `zh` (default, Chinese captions) / `en` (English captions).
-
-- 12 general-purpose shots plus 7 industry shots (real photo, price card, store card, review card, before/after, fact sheet, credential card). `shots.en.md` is the parameter reference for all 19 shots; `docs/shots/*.en.md` has one detailed doc per shot.
-- Each industry pack has compliance rules, a recommended shot structure, a brief template for merchants and a regression-test brief.
-- **Not supported** (the compliance rules don't cover these, so results aren't guaranteed compliant even if you force them): medical aesthetics, prescription drugs / medicine, K12 academic tutoring, dietary-supplement efficacy claims, tobacco.
-
-<details>
-<summary><b>Adding a new industry</b></summary>
-
-Each industry lives under `industries/<id>/`:
-
-```
-industries/<id>/
-  rules.json            Compliance rules (only disabledShots bans a shot; enabledShots is a recommendation and a reopen list, not a whitelist; checks, mediaPolicy, etc.), merged with industries/_base/rules.json
-  recipe.md             Recommended shot structure and writing notes (required reading before writing a storyboard)
-  recipe.en.md          English version
-  brief-template.md     The brief template to hand to a merchant
-  brief-template.en.md  English version
-  test-brief.md         One full sample brief
-  expected.md           Which rules that sample brief should trigger (used for regression tests)
-```
-
-Copy an existing industry folder and rename it, fill in `rules.json` following the field documentation in `industries/_base/rules.json`, write `recipe.md` describing which shots this industry should use and its common compliance pitfalls, then write a `test-brief.md` + `expected.md` and run `node scripts/test-rules.mjs` to confirm the rules actually fire.
-
-</details>
-
-## Tweaks (`meta.tweak`)
-
-A promo storyboard may set `meta.tweak` for pace, type size, and the heading face. A shot may also set `bg` for its own background. Leaving these fields out keeps the picture pixel-identical to today.
-
-| Field | Values | Effect |
-|---|---|---|
-| `meta.tweak.pace` | `slow` / `normal` / `fast` | Multiply each shot's length by 1.15 / 1 / 0.88, then snap to a beat. `normal` is the same as omitting it: no multiply. Each shot's min and max duration still apply |
-| `meta.tweak.textScale` | 0.9–1.15 | Scales the caption band, the end-card big type, and the hook hero type. Omitting it, or writing `1`, does not scale. Crop, safe-area, and overlap checks still block. On quiz and journey, writing this warns: these two fields currently apply to cards only |
-| `meta.tweak.headingFont` | `sans` / `serif` / `kai` | Face for titles and big type. `sans` is Noto Sans SC (the default), `serif` is BrewReel Serif, `kai` is BrewReel Kai. `sans` matches omitting the field. On quiz and journey, writing this warns: these two fields currently apply to cards only |
-| shot `bg` | png / jpg / webp, relative path | Drawn over that shot's gradient and darkened so type stays readable. The file must sit in the project folder and be at least 1KB. Copy must not say 「实拍」 unless the file is registered in `meta.assets` with `source: merchant` |
-
-`textScale` and `headingFont` currently apply to cards only. Quiz and journey draw their own titles, so those two fields do not change them. Writing either field on those styles produces a warning and does not block. `pace` and per-shot `bg` still apply to quiz and journey.
-
-When the user says "a bit faster", "larger type", "kai for the title", or "a new background on this shot", a cheap model writes the fields above and does not edit the template. To change one thing in a film that already exists, run `node scripts/revise.mjs <storyboard.json> "one sentence"` instead of rewriting the file.
-
-## Talking-head B-roll
-
-Use this when a real talking-head video needs explanatory pictures on the lines about a step, a number or a comparison. v0.8 was the experiment; since v0.9 one command can render from `talk.mp4` alone. The limits remain, and the pages still treat it as experimental. Put only `talk.mp4` in a project folder and run one command, `node scripts/talk.mjs <project> --out <dir-outside-the-repo>`: it transcribes the talk on your machine, a cheap model writes `broll.json`, and scripts compute timing, price, generation and the final cut.
-
-Two kinds of picture. Motion clips are free: five card types (keyword, checklist, steps, counter, compare) whose words are all copied from the speech and checked character by character. AI clips are generated with MiniMax H3, at most 2 per film by default, in three regular styles, wood blocks (the default), clay stop-motion and layered paper (ink sketch is experimental and only works as a stand-in preview), with the same robot throughout; cost is estimated before anything is spent, and a person approves them on the review page.
-
-Three placements: `full` covers the frame, `pip` keeps the face in a circle at the bottom right, `split` puts the picture on the top 60% and the face on the bottom 40% (vertical only). Captions can be added, omitted, or already burned into the talk (burned captions require `split`). Steps, price and limits: [`docs/broll.en.md`](broll.en.md).
-
-## Style factory
-
-Added in v0.10. One sentence describes an AI-picture style for a talk. `node scripts/broll/new-style.mjs --id <id> --name <name> --desc <one sentence>` writes `style.json`, generates reference images, scores them and makes one still. The vision check is only the first screen. After a person looks at `broll/styles/_drafts/<id>/review.html`, they run `node scripts/broll/approve-style.mjs <id>` themselves in a terminal, and only then does the style join the catalog. An AI assistant must not run that command. Pull requests that bring back an approved style are welcome; those files are not the promo-recipe remix described below. Steps, cost and the vision rules: [Style factory](style-factory.en.md).
-
-## Validation and compliance
-
-The rules in `scripts/validate.mjs` are compiled from publicly available advertising-law superlative-word lists and public platform rules, in three tiers:
-
-- **error (block)**: must be fixed, or `make.mjs` refuses to render.
-- **warning (warn)**: recommended to fix, not enforced.
-- **human review**: things a machine can't judge (e.g. "is this photo actually authorized", "is this review genuine"), listed as a checklist for the publisher to confirm after rendering.
-
-It also checks line wrapping, safe areas, fields written in the wrong place, and whether numbers on screen have a source (`meta.facts`); pass `--brief <brief>` when rendering and it checks that every number and quote in the facts comes from the brief. **These rules are for self-checking only and are not legal advice**; see ["Copyright and license"](../README.en.md#copyright-and-license).
-
-## Originality check and recipe crafting
-
-Just as homebrewers share recipes, the project welcomes three ways to play:
-
-- **Use**: brew with an existing recipe.
-- **Remix (reskin)**: `node scripts/gen-styles.mjs --new <id>` scaffolds a new style; copy over the shots and `tokens.json` you want to keep from the source style, then change palette, fonts and characters. For a color-only variant, editing `themes` in `tokens.json` is enough.
-- **Create (craft a recipe)**: follow the six-step process in [`distill/`](../distill/README.en.md) to break down a reference video: extract frames and find cuts (`scripts/extract-frames.mjs`) → nine-layer breakdown + signature list → replicate (local only, never committed) → **redesign** → componentize → cheap-model test → review. `distill/prompts/` has a ready-to-use prompt for each step, and `styles/_template/` is the blank template for a new style. `quiz` and `journey` were made this way, with a quiz-style short video and a character-journey short video as their references.
-
-**Redesign + originality check is a required step**, for remixes and new styles alike. We want a style of the same genre, not a one-to-one copy.
-
-<details>
-<summary><b>What the originality check requires</b></summary>
-
-- **The skeleton may stay**: narrative structure, rhythm, motion techniques, camera language and layout principles belong to the genre and anyone can use them.
-- **The skin must be redone**: palette, type treatment, character design and character colors, signature details, fixed copy lines, and branded endings.
-- **Prove the distance with numbers**: run `node scripts/check-originality.mjs --style <id> --ref <reference tokens.json> --signatures <signatures.md>` (keep the reference tokens and signature list outside the repo). Theme chromatic colors must be at least ΔE2000 20 from every reference chromatic color, backgrounds at least 8, the primary + accent pair must not match any reference pair, outlines must differ, no color may be copied, and every signature item needs a `Replaced with: …` (or `已替换为：…`) or `Removed` entry in `styles/<id>/originality.md`.
-- **Review scores confusability** from 1 to 10 and only ≤ 3 passes: someone who knows the reference should see at a glance that it is not the same maker.
-- In v0.2.1 `quiz` and `journey` went through this step and replaced the v0.2.0 skin wholesale; see each style's `originality.md` for the item-by-item record.
-
-Reference videos and their stills, extracted frames, characters, brand names, URLs, film clips and palette tokens never go into the repo. Keep the raw breakdown and the signature list outside the repo; the repo only holds breakdown text you wrote, characters and scenes you drew, and sample content you wrote. Steps and the PR checklist: [CONTRIBUTING.en.md](../CONTRIBUTING.en.md).
-
-</details>
-
-## Render QA and the delivery manifest
-
-`node scripts/make.mjs <storyboard> --out <dir>` runs the whole chain: validate → machine self-check → copy assets → music → render (with layout and Chinese-character probes) → contact sheet + check frames → `manifest.json`.
-
-- **A ✗ means no delivery**: if the machine self-check, text layout, layout check, the Chinese-character scan for English videos, or the blank-frame check (more than 95% of the screen one color for more than 6 frames in a row) has any ✗, the video is renamed `video.rejected.mp4` and kept only so a person can see what broke.
-- **Check frames on the beat that matters**: for one-take recipes like journey, the check frame is taken on the beat set by the shot spec's `checkBeat`, when the content card has settled; its title and category name must be visible.
-- **Delivery manifest**: on success it writes `manifest.json` with the storyboard's sha256, the video's sha256 / duration / size, the generation time and every check's result; the last terminal line is `交付：<mp4 path>` ("delivered"). No such line means no video.
-- **Verify a video**: `--verify` checks that the video in a folder still matches the current storyboard; if the storyboard changed, render again.
-- **Concurrency**: several instances can run at once; the render step queues automatically, and a lock held by a dead process is reclaimed.
-- Keep test output outside the repo with `--round <round>`.

package/skills/brewreel/docs/features.md

@@ -1,157 +0,0 @@
-# 功能
-
-[← 返回 README](../README.md) · [English](features.en.md)
-
-
-## 三种配方
-
-分镜里写 `meta.style` 选配方，不写就是默认的 `cards`。每种配方是一个「风格包」：自带设计令牌、镜头组件、校验规则和叙事模板，文档、规则、样例放在 `styles/<id>/`，代码放在 `template/src/styles/<id>/`。
-
-| 配方 | 画幅 | 适合 |
-|---|---|---|
-| `cards` 卡片信息流（默认） | 9:16 | 单一卖点、使用流程、界面演示、实物和门店、价目表 |
-| `quiz` 答题互动 | 9:16 | 有一个常见误解、能出一道唯一正确答案的选择题（「X 到底是什么意思？」） |
-| `journey` 角色漫游 | 4:5（默认）/ 9:16 | 有 4–6 个清楚的类别、功能或站点，想按一条路线挨站逛一遍 |
-
-怎么选：能出一道选择题 → `quiz`；有 4–6 个类别想挨站逛 → `journey`；其余情况 → `cards`，也就是不写。`quiz` 只接受 9:16，它的专属镜头（出题、揭晓、释义卡等）只能在 `quiz` 里用，和 `cards` 的 19 个镜头不混用，写错会被校验拦下。
-
-```json
-{ "meta": { "style": "quiz", "industry": "software", "lang": "zh" }, "shots": [ ... ] }
-```
-
-<details>
-<summary><b>每种配方长什么样</b></summary>
-
-**`cards` 卡片信息流**（默认，可用，9:16）：渐变底 + 居中白卡片 + 描边大字幕，一镜讲一件事。适合单一卖点、讲使用流程、演示界面、实物和门店。19 个镜头、六个行业的样例都在 `examples/`。默认 6 套主题仍是描边大字幕：`warm-emotion`、`tech-dark`、`fresh-light`、`business-blue`、`festival-red`、`mono-premium`。另有 12 套可选配色，见 [`styles/cards/THEMES.md`](../styles/cards/THEMES.md)。
-
-**`quiz` 答题互动**（可用，9:16）：皮肤是「批改纸」——点阵答题纸底加一条朱红页边线，深色主色 + 杏黄马克笔 + 朱红批改笔，标签用等宽字，卡片是小圆角 + 实色硬投影；三套配色（`sage-pine` 灰绿纸 + 松绿，默认；`rice-soy` 米纸 + 酱色，餐饮默认；`ash-teal` 灰纸 + 深青），三套口吻（`phraseTitle.params.voice`：`exam` 考场腔 / `chat` 闲聊腔 / `show` 挑战腔）。两个原创角色一矮一高：戴耳机的主讲人（懂了耳罩亮起、发出声波）和反戴棒球帽的搭档。节奏是「红笔圈出一个常见误解 → 答题卡出一道 A/B/C 题、秒表倒数 3 → 揭晓打勾 → 词条卡翻正（误解被波浪线划掉 + 编号释义）→ 回放原片段、盖一枚「懂了」章 → 小剧场演一遍 → 评论区互动 → 印章擦除落到结业卡」。适合有常见误解、能出一道选择题的产品：一句外语的真正意思、软件里一个常被误会的功能（如「归档 = 删掉了？」）、一道菜为什么要这么做。说明见 [`styles/quiz/`](../styles/quiz/README.md)，写法和字数见 `styles/quiz/recipes.md`，5 份样例分镜在 `styles/quiz/examples/`。
-
-**`journey` 角色漫游**（可用，默认 4:5，也支持 9:16）：皮肤是「旅行文具」+「剪纸分层」——原创吉祥物「橘团」踩悬浮滑板，一镜到底横穿一座剪纸风城市（远、中、近景和角色各是一张纸，身后投硬边纸影）；开场是车站翻牌大字，顶部一整条车票标着路线、站点和当前站名，每站甩进来一张航空信封边的明信片写代表内容、看完「寄出」到车票上这一站；配色是邮政绿 + 荧光青柠 + 石墨描边（`post-green`，另有 `plum-ticket` 酒红车票）。背景三选一（现代城市 / 低层街巷 / 古城的白墙黛瓦和石桥河道），14 种街区按内容挑（道口、邮筒、检票口、大头贴、搭车站、夜站台、手机、住家、咖啡、集市、城门、石桥、茶馆、灯会），每站一个车站 / 邮路题材的小笑点，天色从白天走到夜晚；到终点，顶部那张车票滑到画面中间展开，检票钳在票根上打个孔，车票翻面露出产品名和口号；最多一个数字，印在正面终点旗旁。适合内容多、类别清楚的产品：内容平台的栏目、软件的功能、一条游线的景点。说明见 [`styles/journey/`](../styles/journey/README.md)，写法和字数见 `styles/journey/recipes.md`，3 份样例分镜在 `styles/journey/examples/`。
-
-</details>
-
-## 配音（MiniMax / 阿里云 / 火山引擎）
-
-分镜里写 `meta.voice`、每镜写一句 `vo`（旁白），出片时就带配音；不写就是无配音片，和以前一样。
-
-- **三家可选**：`meta.voice.provider` 写 `minimax`（MiniMax）、`aliyun`（阿里云百炼 CosyVoice）或 `volcengine`（火山引擎豆包语音），默认都是男声播音，音色可换。MiniMax 已用真实 key 出过片；阿里云和火山引擎按官方文档接入、用假响应测过，**还没用真实 key 实测**。
-- **声音就是时间轴**：`make.mjs` 先用语音合成念出每句旁白、拿到逐字时间，再把有旁白的镜头时长改成「0.15 秒 + 旁白 + 0.35 秒」并对齐到拍，便宜模型不用自己算时长。
-- **逐字字幕**：字幕跟着声音一个字一个字点亮，三种配方都支持；也可以选整句出现（`line`）或不出旁白字幕（`off`）。cards 里有 `vo`、没写 `caption` 的镜头，字幕由旁白自动生成。
-- **配乐自动让路**：人声处背景音乐压低约 10 dB，说完再回来。
-- **缓存省钱**：同一句（音色、语速、情绪、文字都一样）只合成一次，缓存在用户目录 `~/.cache/brewreel/tts`（环境变量 `BREWREEL_TTS_CACHE` 可改）；改画面、重渲染不再计费，`manifest.json` 记下这次计费的字符数。
-- **没有 key 也能预览**：出片加 `--voice-provider mock`，用不联网的占位音看节奏（不能当成片交付）；加 `--no-voice` 出无配音版。
-- **密钥**：只从环境变量读，不写进任何文件和日志：MiniMax 用 `MINIMAX_API_KEY`，阿里云用 `DASHSCOPE_API_KEY`，火山引擎用 `VOLCENGINE_TTS_API_KEY`（或旧版 `VOLCENGINE_TTS_APP_ID` + `VOLCENGINE_TTS_ACCESS_TOKEN`）；可选的地址、空间 ID 等见 `SKILL.md`。自定义接口地址只接受 https。
-
-三家怎么选、写法、推荐音色、字数上限见 `SKILL.md` 的「配音」一节；三种配方各有一份带配音的样例分镜（`styles/*/examples/*voice*.json`）。
-
-## 六个行业包
-
-| `meta.industry` | 说明 |
-|---|---|
-| `software`（默认） | App / SaaS / 工具类产品 |
-| `food` | 餐饮、门店、单品上新 |
-| `ecommerce` | 电商实物商品 |
-| `education` | 成人职业培训（不含 K12） |
-| `beauty` | 生活美容（不含医美） |
-| `travel` | 文旅、住宿、民宿 |
-
-产品不在这六类里时，不要编 `meta.industry`。先告诉用户没有专门合规规则包；版式可借最接近的一行并写明借了哪个；只跑通用广告法就写 `general`（见 `SKILL.md`「产品不在六个行业里」）。`general` 不是第七个行业包，校验通过不等于已经合规。汽车辅助驾驶用语、药品、医疗器械、金融必须人工审。
-
-`rules.json` 里的 `enabledShots` 是推荐，并用来重新开放 `_base` 禁掉的镜头，不是白名单。校验只拦合并后的 `disabledShots`。
-
-`meta.lang`：`zh`（默认，中文字幕）/ `en`（英文字幕）。
-
-- 12 个通用镜头 + 7 个行业镜头（实拍、价目表、门店卡、评价、前后对比、参数表、资历卡）。`shots.md` 是 19 个镜头的参数总表，`docs/shots/*.md` 是每个镜头的详细文档。
-- 每个行业包带合规规则、推荐的镜头结构、给商家的简报模板和一份回归测试简报。
-- **不做**（校验规则没覆盖，硬要做也不保证合规）：医疗美容、处方药 / 药品、K12 学科培训、保健品功效宣称、烟草。
-
-<details>
-<summary><b>怎么新增一个行业</b></summary>
-
-每个行业是 `industries/<id>/` 下的一套文件：
-
-```
-industries/<id>/
-  rules.json            合规规则（disabledShots 才禁用镜头；enabledShots 是推荐并用来重新开放，不是白名单；checks、mediaPolicy 等，和 industries/_base/rules.json 合并）
-  recipe.md             推荐的镜头结构、写作要点（AI 助手写分镜前必读）
-  recipe.en.md          英文版
-  brief-template.md     给商家的简报模板
-  brief-template.en.md  英文版
-  test-brief.md         一份完整测试简报
-  expected.md           这份测试简报应该触发哪些规则（回归测试用）
-```
-
-复制一份现有行业目录改名，参照 `industries/_base/rules.json` 的字段说明填 `rules.json`，写 `recipe.md` 说清楚这个行业该用哪些镜头、常见合规红线，再写一份 `test-brief.md` + `expected.md`，跑 `node scripts/test-rules.mjs` 验证规则生效。
-
-</details>
-
-## 可调项（meta.tweak）
-
-宣传片分镜可以选写 `meta.tweak`，改整片节奏、字号和标题字体。每一镜还可以写 `bg` 换一张背景图。不写这些字段时，出片和现在逐像素一致。
-
-| 字段 | 取值 | 效果 |
-|---|---|---|
-| `meta.tweak.pace` | `slow` / `normal` / `fast` | 每一镜的拍长乘 1.15 / 1 / 0.88，再吸附到整拍。`normal` 和不写一样，不乘。每种镜头的最短、最长时长照样生效 |
-| `meta.tweak.textScale` | 0.9–1.15 | 缩放字幕带、片尾大字和 hook 主视觉大字。不写或 `1` 等于不缩放。裁切、出安全区、压字这些版式检查照样拦。quiz、journey 写了这项会提醒：这两项目前只对 cards 生效 |
-| `meta.tweak.headingFont` | `sans` / `serif` / `kai` | 标题和大字的字体。`sans` 是 Noto Sans SC（默认），`serif` 是 BrewReel Serif，`kai` 是 BrewReel Kai。`sans` 和不写一样。quiz、journey 写了这项会提醒：这两项目前只对 cards 生效 |
-| 镜头 `bg` | 相对路径的 png / jpg / webp | 盖在这一镜的渐变上，并自动压暗，保证字看得清。文件必须在项目目录里、至少 1KB。没在 `meta.assets` 登记为 `source: merchant` 的图，画面文字不能写「实拍」 |
-
-`textScale` 和 `headingFont` 目前只对 cards 生效。quiz、journey 的标题是风格自己画的，不跟着这两项变。这两种风格写了这两项，校验提醒「这两项目前只对 cards 生效」，不拦截。`pace` 和镜头 `bg` 对 quiz、journey 仍然生效。
-
-用户说「节奏快一点 / 字大一点 / 标题换楷体 / 这一镜换张背景图」时，便宜模型按上表写，不要改模板代码。已经写好的片子要改一处，用 `node scripts/revise.mjs <storyboard.json> "一句话"`，不要整份重写。
-
-## 口播配画面
-
-已经有真人口播、想在讲步骤、讲数字、讲对比的句子上加解释画面时用。v0.8 先做实验；v0.9 起可以只放 `talk.mp4`，一条命令出片。限制还在，页面仍按实验来写。项目目录里只放 `talk.mp4`，一条命令 `node scripts/talk.mjs <项目目录> --out <仓库外目录>`：本机转写出字幕，便宜模型写 `broll.json`，脚本算时间、估价、生成和合成。
-
-画面两种。动效画面免费：关键词、清单、步骤、数字、对比五种卡片，字全部照抄原话，脚本逐字核对。AI 画面用 MiniMax H3 生成，默认一部片最多 2 段，正式风格有积木风（默认，木头积木）、黏土定格、分层纸艺三种，手绘线稿是实验风格、只能占位预览，全片同一个机器人；花钱前先估价，要人看过审片页再批准。
-
-三种放法：`full` 盖满，`pip` 右下角圆窗露脸，`split` 上 60% 是画面、下 40% 是脸（只给竖版）。字幕可以后加、不加，或口播里已经烧好（烧好的只能用 `split`）。步骤、费用和限制见 [`docs/broll.md`](broll.md)。
-
-## 风格工厂
-
-v0.10 新增。一句话描述一种口播用的 AI 画面风格，`node scripts/broll/new-style.mjs --id <id> --name <名字> --desc <一句话>` 写 `style.json`、出参考图、看图打分、做一张静态试拍。看图只是第一道筛。人看完 `broll/styles/_drafts/<id>/review.html` 之后，自己在终端运行 `node scripts/broll/approve-style.mjs <id>`，风格才收进清单。AI 助手不许替人运行这条命令。欢迎把批准后的风格用 Pull Request 贡献回来；这和下面的宣传片配方二创不是同一套文件。步骤、花费和看图规则见 [风格工厂](style-factory.md)。
-
-## 校验与合规
-
-`scripts/validate.mjs` 里的规则整理自公开的《广告法》极限词清单和各平台公开规则，分三档：
-
-- **错误（block）**：必须改，否则 `make.mjs` 拒绝出片。
-- **警告（warn）**：建议改，不强制拦截。
-- **人工复核（human）**：机器判断不了的（比如「照片是否真的授权」「评价是否真实」），出片后列成清单给发布者自己确认。
-
-另外还查断词换行、安全区、字段写错位置、画面上的数字有没有出处（`meta.facts`）；出片时传 `--brief <简报>`，会逐条核对 facts 里的数字和引语是否出自简报。**这些规则只用于辅助自查，不构成法律意见**，见[「版权与许可」](../README.md#版权与许可)。
-
-## 原创性检查与「调配方」
-
-像自酿爱好者分享配方一样，这个项目欢迎三种玩法：
-
-- **一创**：直接拿现成配方开酿。
-- **二创（换皮）**：`node scripts/gen-styles.mjs --new <id>` 从脚手架生成一个新风格，把源风格要沿用的镜头和 `tokens.json` 复制过来，换配色、字体、角色。只换配色的话，改 `tokens.json` 的 `themes` 就行。
-- **三创（调配方）**：按 [`distill/`](../distill/README.md) 的六步流程拆一支参考视频：抽帧找切点（`scripts/extract-frames.mjs`）→ 九层拆解 + 招牌清单 → 复刻（只在本地、不提交）→ **再设计** → 组件化 → 便宜模型测试 → 评审。`distill/prompts/` 里有每一步可直接用的提示词，`styles/_template/` 是新风格的空白模板。`quiz` 和 `journey` 就是这样做出来的，分别参考了一支答题互动类短视频和一支角色漫游类短视频。
-
-**「再设计 + 原创性检查」是必经步骤**，二创、三创都一样。我们要的是「同一类型的风格」，不是一比一复制。
-
-<details>
-<summary><b>原创性检查的具体要求</b></summary>
-
-- **骨架可以保留**：叙事结构、节奏、动效手法、镜头语言、版式原则，这些是类型，谁都能用。
-- **皮肤必须重做**：配色、字体处理、角色造型与人设配色、招牌细节、固定文案句式、品牌化的收尾方式。
-- **用数字证明拉开了**：跑 `node scripts/check-originality.mjs --style <id> --ref <参考 tokens.json> --signatures <招牌清单.md>`（参考的令牌和清单放仓库外），要求主题有彩色与参考有彩色 ΔE2000 ≥ 20、背景 ≥ 8、主色 + 强调色不能同时撞上参考的一组、描边不同、没有照搬的颜色，招牌清单每一条都在 `styles/<id>/originality.md` 里写了「已替换为：…」或「已删除」。
-- **评审打混淆度**：1–10 分，≤ 3 才算过——熟悉参考的人要能一眼看出不是同一家。
-- v0.2.1 的 `quiz` 和 `journey` 就是按这一步把 v0.2.0 的皮肤整套换掉的，逐条对照见各自的 `originality.md`。
-
-参考视频和它的截图、抽帧、角色、品牌名、网址、影视片段、配色令牌一律不进仓库；拆解原稿和招牌清单放在仓库外，仓库里只放自己写的拆解文字、自己画的角色和场景、自己写的示例内容。步骤和 PR 自查清单见 [CONTRIBUTING.md](../CONTRIBUTING.md)。
-
-</details>
-
-## 出片质检与交付清单
-
-`node scripts/make.mjs <分镜> --out <目录>` 一条命令走完：校验 → 机器自查 → 复制素材 → 配乐 → 渲染（带版式和汉字探针）→ 拼图 + 检查帧 → `manifest.json`。
-
-- **有 ✗ 不交付**：机器自查、文字排版、版式自查、英文片汉字扫描、空帧检查（整屏 95% 以上一个颜色、连续超过 6 帧）任何一项有 ✗，成片改名 `video.rejected.mp4`，留着给人看哪里坏了。
-- **抽帧抽在关键那一拍**：journey 这类一镜到底的配方，按镜头规格的 `checkBeat` 在内容卡停稳时抽检查帧，标题和类别名必须看得见。
-- **交付清单**：成功时写 `manifest.json`，记分镜 sha256、成片 sha256 / 时长 / 大小、生成时间和各项检查结论；终端最后一行「交付：<mp4 路径>」。没有这一行就是没出片。
-- **核对成片**：`--verify` 检查目录里的成片是否还对应当前分镜，改过分镜就要重出。
-- **可以并发**：多个实例同时开，渲染这一步自动排队，持锁进程死了会被回收。
-- 测试产物用 `--round <轮次>` 放到仓库外。