@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/draftcut/references/01-plan.md

@@ -1,30 +0,0 @@
-# 视频导演脚本与第二道审核
-
-先完成 [完整口播精剪](00-audio.md)，并通过第一道人工审核生成正式精炼音频。此后才在批准音频时间线上规划画面。
-
-明确授权的隔离模拟 run 使用两次独立的模拟用户确认，流程和变更失效要求相同，但不冒称 owner 批准，见[测试说明](../docs/simulation-testing.md)。
-
-逐步解决 layout、idea、materials、style、references；素材和参考片可选。依据内容组织论点、演示、比较与 MG 动画；不用每句标题卡代替解释。数字与事实必须有来源。
-
-用户已经给过的素材和风格不要重复询问。其余可选偏好由 Agent 根据内容提出默认方案，合并进本次导演脚本确认；不用先连续问一串参数问题。
-
-在工作项目写 `MG导演脚本.md` 与 `director-plan.json`。`source.audio/subtitles` 指向上一阶段返回版本目录的 `narration.wav/final.srt`，`duration_seconds` 取实际采样时长。结构例见 [plan.json](../examples/plan-validation/plan.json)。`workflow` 必须包含：
-
-```json
-{"schema":"draftcut.workflow.v2","project_dir":"项目绝对路径","director_script":"导演脚本绝对路径"}
-```
-
-`assets` 列出该计划要用的可选素材路径。`creative_brief` 解决版式、想法、素材和风格，`visual_direction/package_direction` 写具体表达，`render_spec` 指定画布与帧率。cues 在帧网格上从零连续覆盖至音频时长的帧舍入。不要设置 `approved=true`，它不是批准来源。
-
-```bash
-python3 <skill-root>/scripts/spoken_workflow.py director-review \
-  --plan "$work/project/director-plan.json" --script "$work/project/MG导演脚本.md"
-```
-
-向用户交付返回的 `review.md`，其中包含完整脚本、计划、批准音频和素材摘要。**停在第二道门，等待用户对这版视频脚本明确签名批准**。第一道批准、无人值守措辞或 Agent 自审都不能代替这次批准。
-
-```bash
-python3 <skill-root>/scripts/validate_plan.py --plan "$work/project/director-plan.json"
-```
-
-该入口同时检查结构和两道当前签名。脚本、计划、素材变更使本次批准失效；口播稿变更则两次批准都需更新。批准后按 [制作](02-build.md) 连续执行。

package/skills/draftcut/references/03-deliver.md

@@ -1,26 +0,0 @@
-# 导出与交付
-
-```bash
-node <skill-root>/scripts/render_master.mjs --project-dir "$output_dir/hyperframes" \
-  --plan "$plan" --output "$output_dir/final/video.mp4"
-node <skill-root>/scripts/validate_master.mjs --plan "$plan" \
-  --final "$output_dir/final/video.mp4" --output-dir "$output_dir"
-```
-
-production 导出要求 ready 和连续预览证据。默认本机完整 Chrome、单 worker、硬件 H.264；环境不支持时明确使用 --encoding software，不静默替换。不能自动下载浏览器或升级运行时。--purpose smoke 仅用于标为合成的技术测试，不绕过创作者视频的制作。
-
-用户明确授权且有独立 run 清单的模拟用户测试，必须用 `--purpose simulation`、run 内 `TEST-*.mp4` 和测试水印，见[测试说明](../docs/simulation-testing.md)。不能用 production 或 smoke 接受模拟批准；报告只作交互与技术链路证据。
-
-final/ 仅含最终 MP4，日志、计划、字幕和报告在工作目录。校验批准状态、尺寸/帧率、视频和音频时长、H.264/AAC、单文件交付；时长容差为三帧，不是感知同步证明。连续观看覆盖开头/中段/尾段，核对字幕与论点；试听可选，未试听如实记录，不构成出片门槛。
-
-从导出的 MP4 精确提取实际最后一帧（帧号 `nb_frames - 1`），检查批准结尾仍可见，模拟导出还须可见测试水印；不能只看 `duration - 0.1s`。最后一帧空白即修正同工程视觉覆盖并重新渲染，保留失败版本，禁止把重封装或元数据通过当作画面已修复。
-
-quality-report.v1.json scope=technical；视觉与听感单独记录，没做就 pending，不把技术通过写成创意或音质已确认。交付视频、导演计划、准备报告、技术报告、预览/试听结论和真实限制。失败保留日志，不以假样例替代。
-
-不自动上传 GitHub、Library 或视频平台；[非商业许可](../LICENSE.md) 与商业事先书面授权要求保留。
-
-## v2 发布门禁
-
-初始化、正式／合成导出和交付检查都会重验两道签名链。工程目录固定为该工作项目的 `hyperframes/`；不能从另一个工程借用批准。导出前检查唯一原声音轨、DOM 字幕内容／时间和批准素材；导出后比较 AAC 解码波形与批准 WAV，mono→stereo 显式复制声道，避免隐式下混增益干扰。
-
-`--purpose smoke` 仅接受经过两次 fixture 模拟批准的已知合成 PCM／词表。给真实录音加 `example_only` 标签、设置旧 approved 字段或拿旧审核文件重试都不能通过。导出临时文件先完成检查，再复制到交付目录；日志和导出批准记录保存在工程内。

package/skills/draftcut/references/04-reference-theme-prompt.md

@@ -1,7 +0,0 @@
-# 风格起点与参考分析
-
-主题是可选起点，不是必选模板。只有音频时做全屏解释图、屏幕演示、编辑式文字和 B-roll；另有已授权人物素材时才考虑画中画课程或人物叙事。
-
-询问受众、观点与视觉感觉，不要求主题编号。参考可访问则分析信息进入、重点切换、文字密度、节奏、留白、色彩和动效力度，再转化为当次创意建议。不能复制音乐、画面或字体；无参考也可依据 brief 继续。
-
-统一设计语言，让画面帮助理解内容。不用演示数据冒充事实，不靠持续弹跳、模板卡片或无意义装饰填满时长。

package/skills/draftcut/references/05-social-loudness.md

@@ -1,58 +0,0 @@
-> **v2 审核边界：**以下为历史母带工具的信号处理参考。当前完整口播流程的正式母版绑定逐采样原声；直接变速或母带处理后的文件不能替换批准音轨，也不能靠旧授权绕过两次批准。需要此类处理时先明确新的审核与信号变换方案，当前生产门禁会拒绝未绑定的替换。`master_audio.py` 只作为独立诊断工具保留，不代表该选项已接入 v2 生产链。
-
-# 倍速与口播母带
-
-在用户要求倍速、母带、高响度，或报告爆音、发糊、刺耳、泵动时读取。用户的听感问题优先于响度目标；LUFS、真峰值和解码通过不能证明没有可听失真。
-
-## 倍速只做一次
-
-- 分别记录画面、音频、外部字幕各自的当前速度和共同时间原点；每项实际处理倍率为 `target_speed/该项当前速度`。例如画面已1.15倍、干净音轨仍1倍时，画面倍率为1，音频才用1.15。不要对已经1.15倍的母带再做 `atempo=1.15`。
-- 整片变速时，保留已批准画面作为整体底片：HyperFrames `<video data-playback-rate="倍率">`，烧录字幕、人物和MG一起变速。外部SRT/章节时码按各自当前速度除以对应倍率，文字不变；用户明确纠正文字时只改指定词。
-- 声音从同剪辑时间线、未做母带的源音轨重建，以 `atempo` 保留音高；已变速的WAV放入工程后，音轨播放倍率为1。不要用改采样率的方式制造变速。
-- 画面保留批准的fps；时长按最终画面帧数确定。不能用 `-shortest` 掩盖音画速度或时长不匹配。
-- 用户提出修订时先记录要求；涉及已批准信号或脚本变化的，按当前 v2 流程重新审核相应版本。
-
-## 用户说爆音/发糊时
-
-1. 保留被拒绝版，回到未做母带的源音轨。只把坏母带调小不能撤销已产生的失真。
-2. 优先对照开头和用户指出的位置：源音轨、处理后的PCM、最终解码AAC。以相近播放响度比较，区分源录音失真、处理失真、编码问题与播放端过载。
-3. 先撤去重压缩、makeup增益、追目标的额外 `volume`、饱和器及多级限幅；根据证据再加入需要的处理，不默认套高频增强/降噪/去齿音链。
-4. 峰值低于0dBFS仍可能存在之前的削波、压缩/限幅失真、齿音或时间伸缩伪影。不要用“真峰值合格”否定用户的听感。
-5. 无法试听时，明确记为 `listening.performed=false`，可交付重建版及技术报告，但不能宣称“爆音已确认消失”或“听感验收通过”。用户反馈是进一步修订证据，不设置额外的统一审批门。
-
-## 响度目标按声音决定
-
-- 温和口播/失真修复可先从约 `-14 至 -12 LUFS-I` 开始，保留句子起伏。
-- 用户明确要“响度战争”时，可以尝试更高响度，`-10 至 -9 LUFS-I`是可选目标，不是必须达成的下限。出现失真、发糊、底噪抬升或持续限幅时退回更温和版本，并报告实际结果。
-- 处理峰值通常留在 `-2 至 -1.5 dBTP`，最终解码AAC应不高于 `-1.0 dBTP`；平台再编码需要余量。LRA仅作辅助，不为了范围压平自然语气。
-- 不把某一期的速度、LUFS、均衡和压缩参数变成全局固定配方。
-
-## 固定信号，再做两遍测量
-
-先把倍速、声道、必要的EQ/压缩等前处理写成无损预混文件，再测这个文件。第二遍必须使用同一信号及相同I/TP/LRA目标；初次测量若只是为确定LRA，确定目标后重新测量，不能沿用另一组目标的offset。前处理发生变化（包括新加压缩器、增益或变速）就重新测量，不能复用旧的 `measured_*`。
-
-优先只使用一层测量驱动的归一化，不默认再叠加补偿增益和末级限幅。`linear=true`并不保证线性模式：峰值/LRA条件不满足时，FFmpeg会回退动态模式，必须记录实际 `normalization_type`。目标未达到时先评估余量与失真，不能仅按差值加增益。
-
-可使用内置的保守入口；它不叠加独立压缩器/EQ，也不声称做过试听；归一化本身仍可能是动态处理：
-
-```bash
-python3 <skill-root>/scripts/master_audio.py \
-  --source-audio "$prepared_audio" \
-  --source-speed 1 --target-speed "$target_speed" \
-  --duration "$approved_duration" \
-  --target-lufs -13.5 --true-peak -2 \
-  --output-dir "$output_dir/audio-master"
-```
-
-输出无损倍速源、24bit/48kHz WAV、AAC及 `audio-master-report.v1.json`。输出目录必须是新的空目录。`--subtitles`须同时提供 `--subtitle-source-speed`，按字幕自己的倍率生成时码副本；只改时码行，正文中的时间字符串不变。此入口是温和起点，不保证每个录音都无需进一步处理。
-
-## 长度、封装与验收
-
-- 倍速在前，母带在后。每次从无损预混继续，不反复转码已经处理的AAC。
-- 延迟与长度分别检查：如果使用 `alimiter`，显式考虑 `latency=true`；使用任意有延迟的处理后，核对开头/中段/尾段同步。`asetpts`不能撤销真实波形延迟。
-- 末端重采样后先 `asetpts=PTS-STARTPTS`，再 `apad,atrim=end_sample=round(duration*sample_rate)`，最后重置PTS；按样本数核对WAV及AAC长度，避免滤镜PTS偏移造成约一个AAC帧的尾部缺口。仅允许补偿编码/舍入尾差，不能用补静音修复错误倍速。
-- 已批准画面只改音频时，`-map 0:v:0 -map 1:a:0 -c:v copy`替换音轨；母带AAC只编码一次，能直接复制封装时使用 `-c:a copy`。验证前后视频码流哈希不变。
-- 对**最终解码AAC**复测LUFS-I、真峰值、LRA、音视频起点、尾差及解码；记录目标未达成的情况。检查字幕边界和倍速映射，不能只查文件存在。
-- 试听覆盖开头、最高电平/齿音/爆破音、用户问题段；修复后用相近响度A/B比较。没有实际听过就保留待试听状态。
-- 报告将 `technical_checks` 与 `listening` 分开。脚本将未达到请求响度目标单独列为检查失败，不能把安全但未达目标的输出当作已达目标；优先保留声音质量并明确记录目标调整。默认LRA不低于测得的源LRA，避免无意压缩动态。原始音轨读取启用严格解码错误检查。报告还分别记录倍速阶段和归一化阶段补删前的PCM样本数、补删尾差及AAC解码填充；仅声明时长一致不足以证明同步。
-- 技术通过但未试听的总体状态为 `technical_passed_listening_pending`，不得用无条件 `passed` 暗示听感已合格。

package/skills/draftcut/references/06-modern-components.md

@@ -1,72 +0,0 @@
-# 现代组件：先复用，再编排
-
-主题决定“这一段为什么出现、出现多久”，组件提供已实现的运动与布局，`theme.css`
-统一字体、强调色、圆角、描边和节奏。组合这三层，避免每条视频重写所有效果。
-Hyperframes Catalog 是可编辑源码，不是只能调用的黑盒插件，也不是任意 React UI 库。
-
-## 找到合适的效果
-
-```bash
-node <skill-root>/scripts/hf.mjs catalog --query "two income curves comparison" --json
-node <skill-root>/scripts/hf.mjs catalog --query "highlight a key phrase" --json
-```
-
-用英文描述动作，屏幕内容仍写中文。先搜索当前段落需要的效果，不遍历整库或安装整组。
-结果中的 name 是实际安装名；查看条目说明和源码，确认是想要的效果再用。
-搜索结果较长时只向模型返回最相关几项。无需下载语义搜索模型即可先用词匹配。
-
-## 对 DraftCut 有用的起点
-
-| 表达需求 | 优先搜索/查看的组件 | 使用判断 |
-|---|---|---|
-| 两条路径或职业收入曲线 | `mk-line-graph`、`data-chart` | 改为自己的示意关系与图例，不保留演示数据 |
-| 关键词高亮与现代字幕 | `caption-highlight`、`caption-clip-wipe`、`caption-editorial-emphasis` | 保留最终 SRT，统一中文字体；短语高亮即可，不伪造逐词时间 |
-| 反问、反差、综艺强调 | `caption-kinetic-slam`、`caption-emoji-pop` | 只在语义落点介入，不把整片变成持续弹跳 |
-| 数字或结论解释 | `animated-bar-chart`、`chart-story` | 图形随讲述展开；只用实际内容支持的数据 |
-| 方法、步骤与因果 | `flowchart`、`flowchart-vertical` | 选择适合画布的方向，控制字量，避开人物 |
-| 身份条与段落标记 | `yt-lower-third`、`lt-mask-reveal` | 以创作者信息替换示例，不把每句都做成标题卡 |
-| B-roll 进出与视觉衔接 | 搜索 `subtle wipe transition` / `soft reveal` | 先找克制转场；强 shader 效果只在内容需要时使用 |
-
-条目可能随 Registry 更新；这些是查找起点，安装前以当前查询结果为准，不自动升级运行时。
-
-## 安装与接入
-
-```bash
-node <skill-root>/scripts/hf.mjs add mk-line-graph --dir "$output_dir/hyperframes" --no-clipboard --json
-node <skill-root>/scripts/hf.mjs add caption-highlight --dir "$output_dir/hyperframes" --no-clipboard --json
-```
-
-- **Block**：独立场景，通常位于 `compositions/<name>.html`。在主页面用
-  `data-composition-src` 引入，填写 id、开始时间、时长、画布尺寸。宿主 id 与子场景 id 对齐。
-  若把完整文档改为 template，style/script 必须放在 template 内，避免被装配器丢弃。
-- **Component**：小效果片段，通常位于 `compositions/components/<name>.html`。读取后把所需
-  HTML/CSS/JS 合入当前场景。不要把其整套示例页面、演示字幕和另一个主时间线一并复制。
-- 主页面注册 `window.__timelines.master`；子场景注册自己的 id，由框架自动装配。
-  给组件内部元素加唯一 id 前缀，避免多次使用时相互影响。
-- 安装后的文件留在项目中复用，不在每次导出时重新安装；已有自定义内容不要用 `--force` 覆盖。
-  将条目名、来源 URL、修改点简记到 `components-used.json`，方便下一条视频继承。
-
-## 统一到你的视觉体系
-
-以现有 `theme.css` 或用户的视觉规范为准。新项目可从少量变量开始：字体、主文字色、
-强调色、背景、圆角和描边；由实际设计决定值，不硬编码“诗人程序员 Jun”到所有项目。
-先统一这些变量，再调整尺寸、安全区和动效进出。现有主题中的布局关系继续有效。
-
-中文适配尤其注意：移除英文大写/过大字距，加载真实中文字体，长句按语义分页或调整字号。
-原 SRT 只有句级时间时使用句/短语级高亮，不为插件重跑转写。不改变口播原文来迁就样例。
-
-组件带远程字体、GSAP CDN 或演示图片时，按需本地化到 assets，使用初始化时已复制的
-`assets/gsap.min.js`；不要给同一页面加载多个 GSAP 版本。样例图表数据必须替换或明确标为示意。
-Lottie 适合现成动画资产；GSAP 适合花字、曲线和时间编排；Three.js/WebGL 仅在确需三维效果时使用。
-
-## 快速对 Agent 说
-
-> 用“竖屏个人 IP”作为视觉起点。先查 Catalog 里的曲线图和关键词高亮组件，统一成我的视觉风格，
-> 动态口播保持主体，B-roll 按语义穿插，MG 在人物旁边分步解释。复用好的实现，不保留演示文案。
-
-组件检索、安装、设计、render 都沿用 DraftCut 的批准状态。不要运行上游整套 talking-head-recut
-或迁移 skill 来重新转写、要求额外批准、逐场景截图。安装组件不代表上传素材；没有明确要求时
-不运行 publish、feedback、file-issue、云渲染或其他外部发布命令。
-
-参考：[Catalog](https://hyperframes.heygen.com/catalog)、
-[Registry 接线说明](https://github.com/heygen-com/hyperframes/blob/main/skills/hyperframes-registry/SKILL.md)。

package/skills/easyedit/SKILL.md

@@ -1,66 +0,0 @@
----
-name: easyedit
-description: "Make a movie fan edit with easyedit - the film's best speech with animated word-by-word captions, then a beat-cut montage and a black-and-white title shot. Use when the user says \"/easyedit <movie>\", \"make an edit of <movie>\", \"fan edit\", \"movie edit\", or wants another easyedit video. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "blixvip/easyedit"
----
-
-# easyedit
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-easyedit is installed at `{{ROOT}}`. Run it with `{{PY}}`.
-Read `{{ROOT}}/AGENTS.md` before your first edit in a session. It has the whole workflow, and it's short.
-
-## Make an edit
-
-1. `cd {{ROOT}}` and run `{{PY}} -m easyedit.doctor`. Fix anything marked XX.
-2. Build without rendering: `{{PY}} -m easyedit "<Movie>" --no-render`.
-3. Check it before you spend ~10 minutes on the render, as described in AGENTS.md:
-   - Run `{{PY}} -m easyedit.sheet "<Movie>"` and read `qa/report.txt` (text, cheap) - it covers the quote and
-     every shot with flags, plus a ready-made draft at `qa/suggest-curate.json`. If `source: heuristic`, the AI
-     writer was unavailable, so choose the quote yourself.
-   - Look at `qa/candidates.jpg` and `qa/footage.jpg` only when the text leaves doubt. Exclude text cards,
-     repeats and badly cropped shots. Pin a sequence that builds, and pick the hero shot in `curate.json`
-     (start from the suggested draft).
-4. Render: `{{PY}} -m easyedit "<Movie>"` (same flags as the build), then `{{PY}} -m easyedit.sheet "<Movie>" --final`
-   and look at the stills.
-5. Give the user the path `jobs/<slug>/<slug>.mp4`.
-
-If the easyedit web UI is running (`http://127.0.0.1:4331`), you can also start jobs with
-`POST /api/new {"movie": "..."}` and watch `GET /api/jobs`.
-
-## Source
-
-Rewritten for AICUT from [`blixvip/easyedit`](https://github.com/blixvip/easyedit)
-(MIT), skills/easyedit/SKILL.md. The method is that skill's own; the way it runs is AICUT's.

package/skills/erduo-broll-loop-engineering/SKILL.md

@@ -1,136 +0,0 @@
----
-name: erduo-broll-loop-engineering
-description: "Turn original SRT and design into editable B-roll through an independent Director, focused chapter creators, and visual review. Use for SRT-to-video production or revisions; preserve existing legacy projects and explicit AICUT choices. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: MIT
-source: "erduo1998-cell/erduo-broll-loop-engineering"
----
-
-# Erduo B-roll Loop Engineering
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-Make the spoken meaning visible, with considered composition and purposeful
-motion. **Picture quality comes first.** Preserve fresh creative contexts and
-independent visual review; save work through focused handoffs and deterministic
-scripts, never by removing needed creative work or returning an inferior film.
-
-## Route once
-
-- New projects and existing `broll-plan.json` use the creative relay below.
-- Existing Recipe/runtime-plan v1–v4, explicit Remotion/hybrid, or an explicit
-  five-shot version comparison use [legacy production`legacy-production.md`.
-  Preserve original inputs and approvals; do not silently migrate projects.
-- This relay uses pinned HyperFrames. Its role references replace the legacy
-  stage Skills for this route; do not load both sets or the whole Skill catalog.
-
-## Parent: prepare and dispatch
-
-Keep the complete original SRT and original design in the project. If design is
-missing, make a short brief from the user's request; preserve the original SRT.
-Prepare known local assets/fonts and `assets/index.md` with paths and sources.
-Add a material researcher only for real selection work; authorized search or
-generation remains available when needed for the picture.
-
-Read [production commands`lean-production.md`. Resolve the installed
-Skill root and runnable commands once. Use `scripts/prepare-creative-task.mjs`
-to prepare each role's short `TASK.md`, then start it in a **new context** using
-the host's delegation capability. Give the task file and any explicit user
-constraints, not the parent conversation, previous agent logs, or all Skills.
-Respect the user's model selection; do not silently substitute a cheaper model.
-The helper prepares inputs; the host must actually start fresh contexts and may
-still inject global instructions. Do not claim it enforces a context sandbox.
-
-## Creative relay
-
-1. **Independent Director** reads originals and chooses the visual world,
-   meaningful shot boundaries, material, action development and seams. It writes
-   one `broll-plan.json`, a short `direction.md`, and `direction/<id>.md` shot
-   cards. Its role is [creative direction`creative-director.md`.
-2. **Fresh creators** each own a continuous passage and its editable HTML.
-   Group shots by narrative continuity and complexity, not a fixed count.
-   Give originals, shared direction/assets, owned shot cards and adjacent seam
-   cards. Separate file ownership; run independent chapters in parallel where
-   their visual contract is clear. Use [chapter creation`creative-creator.md`.
-3. **Parent runs the scripts** for plan checks, renders, decode, sheets and
-   assembly. Return a concise failed-shot error to its owner. Environment
-   failures belong to the parent; do not make every creator reinstall or search
-   for the same executable. Creators can request a render and view their work.
-4. **Independent visual reviewer** sees the originals, direction, actual media
-   and relevant seams, without creator explanations or cost figures. Use
-   [visual review`creative-reviewer.md`. Technical success cannot
-   approve aesthetics. Send specific time/shot feedback to the original creator,
-   render only changed shots, and recheck affected shots and seams.
-
-Keep these creative contexts separate even for a short film; a short film can
-have one creator. If independent delegation is unavailable, disclose that the
-relay cannot be completed as specified; do not label a same-context self-review
-independent. Revisions of existing work retain its direction and shot ownership;
-reopen directing only when the requested change affects meaning or the visual world.
-
-## Quality and delivery
-
-Use [visual direction`lean-visual-direction.md` and a few matching
-[motion patterns`motion-patterns.md` to make the action concrete.
-For an uncertain style or complex shared transition, make a representative
-passage and view it before multiplying that treatment. Reuse accepted visual
-source or frames as a reference, while varying composition with meaning. There
-is no mandatory number of samples, material quota, or decorative element count.
-
-Respect requested user review stops; existing authorization also covers normal
-renders and repairs. Do not stop for a new approval form merely because a role
-changed. Repair visible quality problems before delivery. If repeated local
-repairs are not improving the picture, change the visual solution instead of
-repeating the same layout or adding more decoration.
-
-Deliver ordered shot files, editable source/assets, complete preview, actual
-output facts and remaining limitations. A partial preview must identify missing
-shots and cannot stand for the full film. State what was actually viewed; frame
-sampling is not continuous playback. Report tokens only from actual host usage,
-separating cached input and output. Do not turn file sizes into savings claims.
-Follow [safe execution`safe-execution.md` for external commands.
-
-## Optional author support
-
-When the user asks for author tutorials, human help, membership or custom work,
-or remains blocked after reasonable self-service troubleshooting, read
-[author support`author-support.md` and offer the relevant optional
-link once for this issue in this conversation. Continue authorized diagnosis;
-do not recommend on ordinary success, one recoverable error, or after refusal.
-Only the user-facing Parent offers the link, including across stage handoffs.
-
-## Source
-
-Rewritten for AICUT from [`erduo1998-cell/erduo-broll-loop-engineering`](https://github.com/erduo1998-cell/erduo-broll-loop-engineering)
-(MIT), erduo-broll-loop-engineering/SKILL.md. The method is that skill's own; the way it runs is AICUT's.
-
-The repository holds 14 skills; this is the one its entry stands for.

package/skills/erduo-broll-loop-engineering/references/animation-craft.md

@@ -1,208 +0,0 @@
-# Animation craft generation protocol
-
-## Purpose
-
-Use Disney's twelve animation principles as upstream thinking prompts for
-Direction and backend-native authoring. They generate motion decisions; they
-are not a shot checklist, runtime capability taxonomy, evidence schema, or
-still-frame review rubric.
-
-Use this protocol together with the shared visual system, compact Recipe v2
-`microBeats[]`, and progressively queried craft guidance. Do not duplicate its
-rules in every Recipe or expand it into a longer downstream prompt.
-
-Do not begin with an effect, easing curve, spring, transition, catalog card, or
-runtime primitive. Separate the content decision from the creative solution and
-generate each shot in this order:
-
-```text
-spoken meaning
-→ audience obstacle
-→ visual job
-→ freely chosen visual idea
-→ attention
-→ body and material
-→ causal action
-→ key states or continuous motion
-→ one expressive peak
-→ settle and readable stability
-→ native runtime implementation
-```
-
-Do not write principle names, scores, compliance tables, or twelve-item
-inventories into Shot Recipes, manifests, handoffs, or QA artifacts. Do not
-require every shot to exhibit every principle. Deliberate stillness can be the
-right generated result.
-
-## Compile the twelve principles into five decisions
-
-### Direct attention before action
-
-This compiles **Anticipation**, **Staging**, and **Secondary Action**.
-
-- Give the shot one first-read focus.
-- Before an important change, send attention to where the change will happen.
-- Let the primary action lead. Secondary action may clarify cause, scale,
-  reaction, or atmosphere, but it must derive from and remain subordinate to
-  the primary action.
-- Do not start every element together or animate decoration merely to keep the
-  frame busy.
-
-### Give every moving element a body
-
-This compiles **Squash and Stretch**, **Solid Drawing**, and the physical side
-of **Appeal**.
-
-- Decide material, mass, volume, rigidity, support, depth, and character before
-  choosing motion.
-- Paper, glass, metal, interface panels, text, photographs, people, and liquid
-  must not share one generic spring personality.
-- Deformation, perspective, shadow, balance, and center of mass must agree.
-- Preserve perceived volume when elastic deformation is appropriate. Do not
-  squash screenshots, body copy, logos, or serious data by default.
-
-### Make action causal and let it finish
-
-This compiles **Slow In and Slow Out**,
-**Follow Through and Overlapping Action**, and **Timing**.
-
-- Build important action as preparation → action → settle, not uniform travel
-  from A to B.
-- Choose speed and acceleration from mass, distance, emotion, and narrative
-  urgency. Do not give unrelated actions one default duration or ease.
-- Derive dependent layers from the primary action. They may lag, overshoot,
-  oscillate, or stop at different times, but must not run as unrelated ambient
-  loops.
-- Let the result become stable long enough to be understood. Stability is the
-  consequence of motion, not the absence of design.
-
-### Choose key states or continuous motion deliberately
-
-This compiles **Straight Ahead Action and Pose to Pose** plus **Arcs**.
-
-- Explanation, comparison, reveal, and argument shots normally begin with
-  clear initial, decisive, and result states; motion connects those states.
-- Organic growth, particles, crowds, fluids, procedural trails, and other
-  continuously evolving phenomena may be generated as continuous motion, but
-  their state must remain a deterministic function of time.
-- Organic or weighted motion normally follows an arc. Straight paths belong to
-  mechanical, interface, constrained, or intentionally severe motion.
-- The choice is an authoring strategy, not a label to expose downstream.
-
-### Spend exaggeration once
-
-This compiles **Exaggeration**, the expressive side of **Timing**, and
-**Appeal**.
-
-- Give each shot one primary expressive peak.
-- Push the property that carries the meaning: scale, speed, pause, path,
-  hierarchy, contrast, deformation, or density.
-- Keep other properties restrained so the peak remains legible.
-- Appeal means clear, specific, and worth watching; it does not mean cute,
-  bouncy, glossy, or maximally animated.
-
-## Director generation procedure
-
-Generate the shot without Shotcraft from these questions:
-
-1. What must the audience understand, what may be hard to grasp from speech
-   alone, and what job should the picture perform?
-2. What original concrete, abstract, metaphorical, diagrammatic, typographic,
-   media-led, or mixed visual idea best performs that job? Explore freely; no
-   route is the default and style must not decide meaning first.
-3. What is the one first-read focus, and how does attention reach it before the
-   decisive change?
-4. What are the subject's material, mass, depth, support, and motion character?
-5. What are the initial state, decisive action, settled result, and readable
-   stable interval?
-6. Which layers lead, which depend on them, and how do dependent layers finish?
-7. Does the idea need authored key states or deterministic continuous motion?
-8. What is the one expressive peak, and what remains restrained around it?
-9. How does this action inherit from or hand energy to the neighboring shot?
-
-Translate the answers into the compact runtime-neutral Recipe fields:
-`audienceUnderstanding`, `visualJob`, `focus`, `heroFrame`, `microBeats`,
-`readableHold`, material needs, and neighbor handoff. Describe observable motion
-and results, not principle names or backend APIs. Give every beat its resulting
-visible state and principal development, and cover intentional stability with a
-`deliberate-stillness` beat. If this closes the shot, stop: no Shotcraft query or
-no-pattern receipt is required. Query only for a named remaining technique
-question or an explicit user request. A selected card may refine the mechanism
-but must not originate the shot's meaning or movement character.
-
-## HyperFrames Builder generation procedure
-
-Read the Recipe and reconstruct its attention path, physical character, causal
-action, key states, expressive peak, and stable result before selecting an
-atomic rule, blueprint, adapter, transition, or ease.
-
-Author in this order:
-
-1. author the maximum visible hero frame first, including focus, depth layers,
-   supporting structure, media relationship, and stable readable result;
-2. place the initial, decisive, result, and stable states on the one paused,
-   seekable timeline;
-3. add only the preparation needed to direct attention;
-4. connect the primary action using motion consistent with the subject's body;
-5. derive lag, overlap, overshoot, and settling of dependent layers from that
-   primary action;
-6. add secondary action only when it strengthens meaning;
-7. remove ambient motion that has no informational or emotional cause.
-
-Use as many HyperFrames mechanisms and visual layers as the generated direction
-needs. Do not treat mechanism count as a quality target or collapse a rich idea
-into a caption-plus-image layout. For elastic motion pair axial scales coherently
-rather than tweening layout dimensions. Put entrance, continuing motion, and
-dependent motion on separate wrappers when their transform ownership would
-conflict. All motion remains deterministic and seek-safe under the official
-HyperFrames contract; continuous motion may run through a shot but cannot stand
-in for its planned finite developments.
-
-## Remotion Builder generation procedure
-
-Author the same maximum visible hero frame first, without creating a preview
-approval artifact. Read the generated direction before choosing
-`interpolate()`, `spring()`, `Sequence`, or a procedural function. Define a
-small number of semantic states
-and their frame windows first, then reconstruct every visible state from the
-current frame and fps.
-
-Use `spring()` only when the declared material, inertia, or landing genuinely
-calls for it. Do not spring every property. Derive follow-through with explicit
-offset functions from the primary action. Continuous motion such as particles
-or trails must be closed-form or fixed-seed and seekable at an arbitrary frame.
-Preserve the stable result instead of letting procedural motion obscure it.
-
-## Boundary of verification
-
-Technical checks may prove time coverage, deterministic seeking, file and font
-closure, rendering, and decoding. Runtime-captured sampled geometry, escalated
-to dense or per-frame evidence only for a bounded risk window, may also
-detect discontinuity, abrupt speed/acceleration changes, failure to settle,
-excessive reversal, synchronized starts, competing motion foci, short or moving
-readable holds, safe-area exits, off-canvas elements, and strong unplanned
-overlap. Only actual rendered DOM or scene geometry qualifies; static source
-regex, Recipe prose, and hand-authored estimates do not.
-
-When Recipes are supplied to the shared lint, it may additionally prove that
-each planned non-still beat is bound to a non-decorative rendered action and that
-the subject actually changes within that beat. Finite actions may transform and
-settle; progressive continuous subject action may fulfill a beat when its end
-state visibly advances. A background loop, particle drift, moving decorative
-line, unbound continuous motion, or a motion declaration with unchanged rendered
-state cannot satisfy a beat. For a long non-still beat, an opening change followed
-by an undeclared long still tail also cannot satisfy the planned window; describe
-intentional stability as `deliberate-stillness`. This is a technical risk filter,
-not a fixed motion interval, beat count, or judgment that the chosen composition,
-metaphor, rhythm, or animation is aesthetically correct.
-
-Technical checks must not claim to prove anticipation, weight, overlap, timing, arcs,
-exaggeration, or appeal. More precisely, these checks do not prove that
-anticipation communicates, mass feels right,
-follow-through is meaningful, arcs are pleasing, exaggeration serves the
-story, or the result has appeal. Do not create a twelve-principle score or
-multiply static review artifacts. Use the shared motion/layout lint as a risk
-filter and render diagnostics only for findings. Representative moving scenes
-provide the early visual-lock decision; the complete moving preview provides
-the final aesthetic decision. If generated motion is wrong, return
-to the Director or owning Builder.