@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/talking-head-video-pipeline/examples/prompts.md

@@ -1,217 +0,0 @@
-# Example Prompts
-
-## First Project
-
-```text
-用 $talking-head-video-pipeline 初始化视频项目。
-主视频：<path>
-我是第一次用，你帮我按默认流程走。
-```
-
-## Full Pipeline With Assets
-
-```text
-用 $talking-head-video-pipeline 制作成片。
-主视频：<path>
-字幕：<srt-path>
-
-12.5 秒插入截图 <image-path>，停留 2 秒。
-35 秒全屏播放 <demo-path>，素材静音，人物放左下角画中画。
-
-先给我官方预览，确认后再导出。
-```
-
-## Raw Footage With Optional Rough Cut
-
-```text
-用 $talking-head-video-pipeline 处理这批原片：<paths>
-我不会粗剪。请先告诉我可以删哪些口误、重复和明显停顿，等我确认策略后再执行。
-如果要转写，请说明云端词级转写的费用可能性和本地词级转写的差别，不要替我强制选择。
-```
-
-## Semantic Asset Placement
-
-```text
-这些截图和录屏需要放进视频：<paths>
-我不会看时间轴，请根据字幕语义决定每个素材的飞入、停留和飞出，并把依据写进 EDIT_PLAN.md。
-```
-
-## Exact Asset Placement
-
-```text
-说到“这里就是最终效果”时插入 <image-path>，停留 2 秒后飞出。
-72.16 秒左右同时插入 <image-a> 和 <image-b>。
-请用字幕轴解析准确时间；如果原话重复出现，先问我是哪一次。
-```
-
-## Hybrid Asset Placement
-
-```text
-关键演示视频固定在 28.09 秒全屏播放，其他截图你按口播语义安排。
-告诉我：我指定的时间会更精准，自动安排的部分会先在官方预览里让我确认。
-```
-
-## Captions Only
-
-```text
-只抽字幕轴，不动原视频。
-视频：<path>
-输出 SRT、CSV 和 transcript JSON。
-```
-
-## Existing Captions
-
-```text
-我已经有字幕轴，不需要重新转写。
-视频：<path>
-SRT：<path>
-CSV：<path>
-请进入风格设计和预览阶段。
-```
-
-## Content Logic Before Motion
-
-```text
-先读取现有字幕和词级转写，把口播按观众能理解的完整论证单元拆成 CONTENT_LOGIC.json。
-字幕保留我说了什么；每个逻辑组只保留观众要记住的结论。
-输入、关系、结果要按口播逐层累积，结论后整组退出。
-如果没有词级时间，请明确写 cue-level，不要用整句开始时间冒充关键词时间。
-先让我确认逻辑组，再制作动效。
-```
-
-## Fast Style Setup
-
-```text
-风格你帮我定，干净专业、蓝色强调、动效克制。
-截图正文和人脸绝对不能挡。
-```
-
-## Custom Style
-
-```text
-卡片用深色高对比，强调色 #FF4D4F。
-字幕用粗白字，关键词用强调色。
-动效 energetic，但展示截图时收敛。
-```
-
-## Screenshot Style Reference
-
-```text
-参考这张截图设计视频风格：<image-path>
-按这张图的视觉语言做，但不要复制其中的品牌和内容。
-我是第一次用，请使用推荐的 medium 强度，并先告诉我你提取了哪些设计规则。
-```
-
-## Semantic GSAP Motion
-
-```text
-根据现有字幕自动匹配 GSAP 语义动效。
-数字、列举、对比、警告、流程、确认和结果要用不同的本地动效配方。
-动效密度 energetic，但不能遮挡人脸、字幕或截图；先检查 MOTION_PLAN.json，再做官方预览。
-```
-
-## Automatic ShotCraft References
-
-```text
-使用新项目默认的 ShotCraft 精选镜头包，动效密度 balanced，最多选 3 个真正有价值的语义节点。
-只借镜头语言，不复制品牌、文案或完整 UI；截图、演示和产品证据区间禁用。
-优先用 HyperFrames + GSAP 适配，本机没装 ShotCraft 也要走原生 fallback，不要自动安装。
-先检查 MOTION_PLAN.json，再给我官方预览。
-```
-
-## Optional Complex Remotion Shot
-
-```text
-这个复杂镜头如果原生 GSAP 很难做好，先告诉我为什么需要 Remotion、会下载什么、装到哪里，以及许可证和渲染成本。
-先运行 remotion_runtime.py --plan，不要安装。等我明确确认后，才允许项目内安装；子片段必须静音并保留原生 fallback。
-```
-
-## Selected ShotCraft Cards
-
-```text
-我想参考 list-reveal、spotlight-hero-card 和 counter-confetti。
-请先检查它们是否与当前字幕语义匹配，只保留合适的；配色沿用当前 DESIGN.md。
-不要直接复制上游模板，优先 gsap-adapted 或 hyperframes-custom，并保留主音频时钟。
-```
-
-## Offline GSAP Package
-
-```text
-我已经下载了官方 GSAP 文件夹：<gsap-directory>
-先用 gsap_check.py 检查版本和可用插件，不要把整个下载目录复制进公开 Skill。
-这个视频只挑真正有语义价值的插件，并给每个插件动效准备 core fallback。
-```
-
-## Brand-Neutral Visual Quality
-
-```text
-先为这个项目写清字体、真实字重、字幕最多一行还是两行、换行规则和组件类型。
-正文和按钮文字必须水平对齐，不要细体大字、手绘感箭头或歪斜文字。
-需要状态提醒时可以用玻璃通知；Focus、Type 和 Split 动效必须由可 seek 的 GSAP 时间轴驱动。
-先运行 design_check.py，再做官方预览。
-```
-
-## Combined Beginner Setup
-
-```text
-用 $talking-head-video-pipeline 初始化这个口播项目：<video-path>
-我没有剪辑经验。请提醒我可以补一张喜欢的参考截图；如果我不提供，就用安全默认风格。
-设计完成后按字幕语义匹配 balanced GSAP 动效，并自动使用不依赖 ShotCraft/Remotion 的精选镜头包；先给官方预览，确认后再导出。
-```
-
-## Staged Keyframe Review
-
-```text
-这是第一次使用这个风格。不要马上渲染完整预览。
-先生成开场、每个主要逻辑组峰值、最复杂画面、截图或画中画、主要转场和结尾关键帧，写入 qa/KEYFRAME_REVIEW.md。
-我确认静态构图后，再给我官方动态预览。
-```
-
-## Scoped Revision
-
-```text
-只修改以下问题：
-12.4-14.0 秒，右侧结果卡遮住了截图正文，请缩小并移到左上安全区。
-其他已确认内容保持不变。
-先输出修改点关键帧，或前后各 3 秒的短片段；确认后再重建完整预览。
-```
-
-## Reusable Style Handoff
-
-```text
-这条视频已经确认完成。请根据最终 DESIGN.md、确认过的参考图、关键帧、VIDEO_LESSONS.md 和 VIDEO_RETRO.md 生成 STYLE_PROMPT.txt。
-删除本期台词、时间点、文件路径、人物和品牌素材，只保留下一条视频可复用的设计规则、确认节点、质量标准和禁止项。
-```
-
-## Resume
-
-```text
-用 $talking-head-video-pipeline 看一下这个视频项目做到哪一步，继续未完成的部分。
-```
-
-## Diagnose
-
-```text
-用 $talking-head-video-pipeline 只做体检，不要改文件，也不要重跑渲染。
-```
-
-## Explicit Learning
-
-```text
-这个圆形人物画中画是正确的。以后我的演示视频都用圆形，不要用矩形。
-请用 $talking-head-video-pipeline 记成长期风格规则。
-```
-
-## Direct Export
-
-```text
-我明确跳过预览，直接导出最终成片。
-保留原始主音频，只做文件参数检查，不抽帧。
-```
-
-## Delivery Retrospective
-
-```text
-用 $talking-head-video-pipeline 复盘这条成片。
-先区分具体故障、环境问题和我的长期风格偏好；只有我确认后，才把规则带到下一条视频。
-```

package/skills/talking-head-video-pipeline/references/asr-adapters.md

@@ -1,51 +0,0 @@
-# Transcription Adapters
-
-The skill does not require one proprietary transcription service. Reuse existing timing first and discuss provider choices only when transcription is actually needed.
-
-## Priority
-
-1. Existing SRT and CSV supplied by the user.
-2. Existing word-level transcript JSON.
-3. A transcription capability already installed in the user's environment.
-4. For precision rough cutting, installed `video-use` with its hosted word-level adapter when the user accepts possible service cost.
-5. A compatible local word-level ASR chosen by the user.
-6. Another user-configured cloud API.
-7. User-provided text as a correction guide while timing still comes from audio.
-
-For raw takes, hosted word-level transcription is the recommended best-fit route because retakes, filler words, and cut boundaries need fine timestamps. Say clearly that it may cost money and requires the user's own credentials. A local route remains optional and may be slower or less consistent depending on hardware and model. Never force either route.
-
-## Detection
-
-Check without installing anything silently. Treat these as examples, not mandatory dependencies:
-
-```bash
-command -v whisper
-command -v faster-whisper
-python3 -c "import whisper" 2>/dev/null
-python3 -c "import faster_whisper" 2>/dev/null
-```
-
-Do not print or read API key values. Check only whether required environment variable names are present.
-
-If `video-use` is installed and the task is rough cutting, follow [video-use-integration.md](video-use-integration.md) instead of recreating its editing pipeline.
-
-## Output Contract
-
-Any adapter must produce word-level timing or enough timing detail to create:
-
-```text
-edit/script-aligned.srt
-edit/caption-table.csv
-edit/transcripts/<video_name>.json
-```
-
-Keep the raw adapter output cached so later text corrections do not require retranscription.
-
-## Missing Adapter
-
-If no adapter exists:
-
-1. Say exactly what is missing.
-2. Offer the shortest compatible installation path for the current OS.
-3. Offer to continue immediately from an existing SRT, CSV, JSON, or transcript.
-4. Do not claim caption files were generated.

package/skills/talking-head-video-pipeline/references/caption-workflow.md

@@ -1,53 +0,0 @@
-# Caption Workflow Reference
-
-## Output Files
-
-Create:
-
-```text
-edit/script-aligned.srt
-edit/caption-table.csv
-edit/transcripts/<video_name>.json
-```
-
-When the adapter exposes trustworthy token timing, also normalize:
-
-```text
-edit/word-timeline.json
-```
-
-Each token entry must contain `text`, `startMs`, and `endMs`. Do not derive fake word timestamps by assigning every word the cue start.
-
-## Caption Grouping Rules
-
-- Prefer semantic sentence groups over mechanical line breaks.
-- Keep each caption readable at normal playback speed.
-- Avoid one very long line plus one tiny line.
-- Avoid captions shorter than roughly 0.5s unless the speech is genuinely short.
-- Avoid zero-duration or near-zero-duration captions.
-- Keep timestamps on the output timeline when cutting is involved.
-- Record whether downstream timing precision is `word-level`, `cue-level`, or `manual`.
-- Use word timing for keyword triggers when available; use cue boundaries or exact user anchors otherwise.
-
-## Text Correction Rules
-
-- Use a user script as a correction guide, not as absolute truth.
-- Prefer the spoken words when the script and video differ.
-- Preserve tool names and brand names exactly.
-- When replacing a word globally, update SRT and CSV together.
-- Do not change timings during text-only revisions.
-
-## Suggested CSV Fields
-
-```csv
-index,start,end,duration,text
-```
-
-## Common Self Checks
-
-- Count captions.
-- Report shortest and longest duration.
-- Search for requested replaced words.
-- Search for known ASR mistakes in proper nouns and tool names.
-- Run `python3 tools/caption_check.py <srt> [<csv>]` before marking captions ready.
-- If the CSV and SRT disagree, fix both from the same timing source rather than patching them independently.

package/skills/talking-head-video-pipeline/references/export-recipes.md

@@ -1,78 +0,0 @@
-# Export Recipes
-
-These are patterns, not mandatory commands. Adapt paths, codecs, and bitrates to the user's machine and target platform.
-
-## Inspect Source
-
-```bash
-ffprobe -v error \
-  -show_entries format=duration,size \
-  -show_entries stream=index,codec_type,codec_name,width,height,r_frame_rate,duration,pix_fmt \
-  -of json "<timeline_video>"
-```
-
-Resolve `<timeline_video>` to approved `working_video` when present, otherwise `main_video`. Never delete or overwrite `main_video`.
-
-## Browser Preview Proxy
-
-Use when the source video is HEVC, high bitrate, or stutters in browser preview.
-
-```bash
-ffmpeg -hide_banner -y \
-  -i "<timeline_video>" \
-  -vf "scale=1920:1080,fps=30,format=yuv420p" \
-  -c:v libx264 -preset veryfast -crf 23 \
-  -c:a aac -b:a 128k \
-  "<work_dir>/assets/main-preview.mp4"
-```
-
-## Final From Full-Resolution Render And Original Audio
-
-Use only when the rendered video is already at the requested final resolution. Do not upscale a low-resolution preview and call it a true high-resolution final.
-
-```bash
-ffmpeg -hide_banner -y \
-  -i "<preview_render.mp4>" \
-  -i "<timeline_video>" \
-  -map 0:v:0 -map 1:a:0 \
-  -vf "fps=30,format=yuv420p" \
-  -c:v h264_videotoolbox -b:v 42000k -maxrate 52000k -bufsize 84000k \
-  -c:a aac -b:a 192k -movflags +faststart \
-  "<output_4k.mp4>"
-```
-
-## Stable Overlay Pipeline
-
-Prefer this when the source is already high-quality 4K, HEVC/Main10, or the user explicitly wants the original video preserved as the base.
-
-1. Render motion graphics as transparent overlay frames or a transparent MOV/WebM if supported.
-2. Use the approved full-quality timeline video as input 0.
-3. Overlay graphics on top.
-4. Map original audio.
-
-Conceptual filter:
-
-```text
-[timeline_video][overlay_rgba] overlay=0:0:format=auto
-```
-
-When the timeline video is 4K HEVC/Main10, this is the preferred path: render only transparent graphics in the motion renderer, composite over the approved full-quality timeline base with ffmpeg, and map its audio.
-
-Do not add `-shortest` as a convenience when full master-audio preservation matters. Verify the intended timeline duration and audio packets instead of silently truncating one stream.
-
-## Render Log Rule
-
-Do not do this for long renders:
-
-```bash
-render-command | head -c 4000
-```
-
-That can terminate the render early. Instead:
-
-```bash
-render-command > render.log 2>&1
-tail -n 40 render.log
-```
-
-If a render is progressing, keep waiting. Do not switch methods merely because the frame count advances slowly.

package/skills/talking-head-video-pipeline/references/layout-safety.md

@@ -1,49 +0,0 @@
-# Layout Safety
-
-Use evidence-first composition. Screenshots and demos explain the claim; motion graphics support them.
-
-## Protected Regions
-
-Before adding overlays, mark:
-
-- face and hand-gesture region
-- existing subtitles
-- screenshot body text
-- product navigation and buttons
-- watermarks and legal labels
-- user-specified no-cover areas
-
-## Placement Order
-
-1. Evidence asset.
-2. Face or speaker PiP.
-3. Main captions.
-4. Semantic card in remaining empty space.
-5. Decorative motion only if space remains.
-
-## Screenshot Rules
-
-- Prefer contain over crop when text matters.
-- Do not blur, darken, or place a large translucent panel over screenshot text.
-- Do not add scan boxes or moving outlines unless the user explicitly asks.
-- Use short entrance and exit motion; keep the evidence still enough to read.
-- When two screenshots share the frame, size them for legibility rather than symmetry.
-
-## PiP Rules
-
-- Use a circle or rounded crop only when requested or established by the user's profile.
-- Keep PiP clear of captions and evidence.
-- Animate position, scale, and opacity; do not warp the face.
-- Preserve the main speaker audio even when the speaker image becomes PiP.
-
-## Collision Check
-
-At each insert boundary, verify:
-
-- no overlap between caption and evidence
-- no overlap between card and evidence
-- no face occlusion
-- text fits its container
-- exit animation completes before the next protected asset begins
-
-If the scene is crowded, remove the semantic card first.

package/skills/talking-head-video-pipeline/references/video-use-integration.md

@@ -1,71 +0,0 @@
-# Optional Rough-Cut Integration
-
-Use this branch only when the user has raw takes, repeated lines, filler words, long pauses, multiple clips, or explicitly asks for a rough cut. Do not interrupt an already rough-cut project with transcription choices.
-
-## Beginner Conversation
-
-Ask one question at a time.
-
-1. `这段视频已经粗剪好了吗？`
-2. If no, explain the visible result first: `我可以先帮你删掉口误、重复和明显空白，生成一个不覆盖原片的粗剪版。`
-3. Ask for strategy confirmation before cutting. Never silently choose what content to remove.
-4. Only when transcription is needed, offer the available timing sources in this order:
-   - existing SRT, CSV, or word-level transcript JSON;
-   - installed `video-use` with hosted word-level transcription;
-   - a compatible local word-level ASR adapter;
-   - user-provided timing/text when no adapter is available.
-
-## How To Explain Transcription
-
-Use plain language:
-
-- **Hosted word-level transcription — recommended for precise rough cutting.** It is normally the best fit for retakes, filler words, and exact cut boundaries, but the service may charge by usage and requires the user's own API key. The current public `video-use` setup uses ElevenLabs Scribe; always confirm the installed skill because its adapter may evolve.
-- **Local word-level transcription — optional.** It can avoid cloud transcription fees and keep audio local, but setup, speed, filler-word retention, and timestamp quality depend on the machine and model.
-- **Existing timing files — reuse first.** If usable SRT, CSV, or word-level JSON already exists, do not transcribe again.
-
-Do not frame either cloud or local transcription as mandatory. Do not claim that a sentence-level local transcript is equivalent to word-level timing for precise cuts.
-
-## `video-use` Route
-
-If the `video-use` skill is installed, route the rough-cut task to it and follow its own workflow. If it is not installed:
-
-1. Say that it is an optional companion skill, not a hidden dependency.
-2. Point the user to the public project: <https://github.com/browser-use/video-use>.
-3. Offer to help install it only after the user asks or approves.
-4. Never install packages, request an API key, or upload media without explicit approval.
-
-The rough-cut contract remains:
-
-- inventory sources before editing;
-- preserve every original file;
-- propose the keep/remove strategy and wait for confirmation;
-- cut only at verified word or silence boundaries;
-- avoid cutting inside a word;
-- pad verified cut edges by roughly 30–200 ms when needed so words are not clipped;
-- add 30 ms audio fades at new boundaries to prevent pops;
-- cache transcript data;
-- provide a reviewable rough cut and edit decision record;
-- make the accepted rough cut the `working_video` for captions, design, preview, and export.
-
-## State Handoff
-
-After the user approves the rough cut, record it without changing `main_video`:
-
-```bash
-python3 <skill_root>/tools/state.py \
-  --project <project_dir> \
-  working-video <rough_cut_path>
-```
-
-The original `main_video` remains the immutable source. Downstream stages use `working_video` when present and otherwise fall back to `main_video`.
-
-## Exit To Main Pipeline
-
-Return:
-
-- original source path;
-- rough-cut path;
-- transcription route used and whether it may incur cost;
-- strategy approval status;
-- unresolved ambiguous cuts;
-- next phrase: `粗剪确认，继续做字幕和画面设计。`

package/skills/talking-head-video-pipeline/references/workflows/captions.md

@@ -1,52 +0,0 @@
-# Naive Video Captions
-
-Produce editable caption files without changing the source video.
-
-Resolve `<skill_root>` by locating the installed `talking-head-video-pipeline/SKILL.md`; do not assume the user's current directory is the skill directory.
-
-## Inputs
-
-- `working_video` from state when an approved rough cut exists, otherwise `main_video` or the video from the user message
-- optional existing SRT, CSV, or transcript JSON
-- optional correction script
-- optional terminology list
-
-## Workflow
-
-1. Confirm the exact source and output directory.
-2. Reuse existing word-level transcript data before retranscribing.
-3. Select an adapter using `references/asr-adapters.md`.
-4. Use spoken audio as timing truth. Use a supplied script only as a correction guide.
-5. Group semantic sentences and create:
-
-   ```text
-   edit/script-aligned.srt
-   edit/caption-table.csv
-   edit/transcripts/<video_name>.json
-   ```
-
-   When trustworthy word timing exists, also write `edit/word-timeline.json` using `text`, `startMs`, and `endMs`. Preserve cue-level timing honestly when it does not.
-
-6. For text-only corrections, update SRT and CSV together without changing timestamps.
-7. Run:
-
-   ```bash
-   python3 <skill_root>/tools/caption_check.py \
-     <project_dir>/edit/script-aligned.srt \
-     <project_dir>/edit/caption-table.csv
-   ```
-
-8. Set state stage to `captions_ready` only after G1 passes.
-
-## Rules
-
-- Never modify source media.
-- Avoid zero-duration or near-zero flash captions.
-- Keep proper nouns consistent.
-- Do not pretend transcription succeeded when no adapter or usable caption source exists.
-- Do not ask users with valid existing captions to choose a cloud or local provider.
-- Do not claim word-level keyword synchronization from SRT cue starts alone.
-
-## Completion
-
-Return paths first, then cue count, shortest duration, longest duration, and any unresolved terminology. End with: `字幕确认后，我可以继续做风格和官方预览。`

package/skills/talking-head-video-pipeline/references/workflows/design.md

@@ -1,77 +0,0 @@
-# Naive Video Design
-
-Turn captions, assets, and spoken meaning into a render-ready `DESIGN.md` and `EDIT_PLAN.md`.
-
-Resolve `<skill_root>` by locating the installed `talking-head-video-pipeline/SKILL.md`; do not assume the user's current directory is the skill directory.
-
-## Inputs
-
-- project state
-- SRT/CSV or transcript
-- user-specified inserts and timing
-- available screenshots, logos, and demo videos
-- explicit style reference or private profile when present
-
-## Workflow
-
-1. Read `references/style-onboarding.md`, `references/asset-onboarding.md`, `references/content-logic-workflow.md`, `references/layout-safety.md`, `references/visual-quality-rules.md`, and `references/gsap-runtime.md`.
-2. Reuse active confirmed profile rules. Project-specific instructions override the profile for this video only.
-3. If style is still unknown, use quick defaults or ask one guided question at a time. Offer beginners an optional screenshot reference; do not require one.
-4. If visual assets exist and placement mode is not already known, ask whether the user wants semantic placement or exact seconds/spoken-sentence anchors. Recommend semantic placement for speed, but say exact anchors are more precise; accept a hybrid.
-5. Inspect image dimensions and video durations. Never guess how long a supplied demo runs.
-6. Fill `EDIT_PLAN.md` with the placement mode and every requested insert: original anchor, resolved start, end or duration, layout, source-audio behavior, entrance, exit, caption evidence, and protected regions.
-7. Build `CONTENT_LOGIC.json` from complete viewer-facing reasoning units rather than punctuation. Use word timing when available; otherwise declare cue-level or manual precision. Let input/support/relation/result beats accumulate only while they help explain one group, then exit the group together.
-8. Run `python3 <skill_root>/tools/content_logic_check.py <project_dir>/CONTENT_LOGIC.json`. For a first project, new style, or reasoning-heavy video, show the group summaries before motion authoring.
-9. Read `references/motion-recipes.md`. Match logic beats and caption evidence to recipes and write `MOTION_PLAN.json` from `templates/MOTION_PLAN.template.json`. Every derived node needs `logic_group_id`, `logic_beat_id`, caption evidence, start/end, region, visual role, and fallback behavior.
-10. For newly initialized projects, run the curated automatic ShotCraft reference branch after native recipes are selected unless the user chose `skip`. Existing approved projects keep their recorded setting. Record evidence intervals before selecting cards.
-11. Add semantic cards only in intervals without screenshot or demo conflicts. Evidence intervals reduce density and may suppress decorative nodes.
-12. Write `DESIGN.md` with CSS-friendly color variables, an explicit typography contract, caption maximum lines and wrap policy, component families, motion density, PiP/mask geometry, and safe zones. Keep creator-specific styling in the private profile rather than the public default.
-13. Check the intended font family and real weight, text baseline, longest-label fit, caption line policy, and glass-surface contrast.
-14. Run `design_check.py`, `content_logic_check.py`, and `motion_plan_check.py`, then G2 from `references/quality-gates.md`.
-15. Set stage to `design_ready` with `tools/state.py` only after G2 passes.
-
-## Screenshot Reference Branch
-
-Trigger on phrases such as `参考这张截图设计视频风格` or `按这张图的视觉语言做，但不要复制其中的品牌和内容`.
-
-1. Read `references/style-reference-workflow.md`.
-2. Confirm the image path and infer reference strength only when the user already states it; otherwise recommend `medium` in plain language.
-3. Inspect the image and write project-local `STYLE_REFERENCE.md` from `templates/STYLE_REFERENCE.template.md`.
-4. Extract colors, typography class and hierarchy, card geometry, spacing, information hierarchy, density, composition zones, reusable visual elements, and prohibited copied elements.
-5. Mark motion directions inferred from a static image as `inferred`; never claim exact reconstruction of the source animation.
-6. Translate the resulting visual language into `DESIGN.md`, then select compatible recipes from `references/motion-recipes.md`.
-
-Never copy a logo, watermark, person, brand UI, original wording, or a complete recognizable interface. Keep the reference image path in the user's project only.
-
-## ShotCraft Reference Branch
-
-Trigger for new projects whose `Shot references` setting is `automatic`, or when the user asks for ShotCraft, selects a card, or enables cinematic shot recommendations.
-
-1. Read `references/shotcraft-integration.md`, `references/shotcraft-default-pack.md`, and the local `references/shotcraft-mapping.json`.
-2. Keep the already selected native `recipe_id`; ShotCraft never becomes the semantic source of truth.
-3. For automatic mode, run `python3 <skill_root>/tools/shotcraft_default_plan.py <project_dir>/MOTION_PLAN.json --in-place`. This adds only safe, semantically matching references and never installs a provider. For manual browsing, query `shotcraft_catalog.py --defaults --semantic <tag> --density <density> --json`.
-4. Respect the density cap from the default pack. Prefer different semantic roles instead of repeating one card.
-5. Write screenshot, demo, and product-UI ranges to top-level `evidence_intervals`. Do not assign a mapped reference across those intervals.
-6. Add `reference.provider`, `reference.card`, `reference.implementation`, and `reference.provider_required_at_runtime=false` to the selected nodes.
-7. Use `gsap-adapted` or `hyperframes-custom` by default. If a selected complex card genuinely needs Remotion, explain the added runtime, download, license, and render cost first. Only after explicit approval may `remotion_runtime.py --install --yes` be used and the node changed to `remotion-subclip`.
-8. Continue normally when the provider is absent; `reference-only` recommendations do not require installation.
-
-## Rules
-
-- Evidence takes priority over decorative motion.
-- Do not invent a personal brand when the user did not provide one.
-- Keep colors configurable; do not bake one creator's palette into the public skill.
-- Explain semantic cards with short readable text, not full transcript paragraphs.
-- Keep captions and cards separate: captions preserve speech; logic beats preserve the takeaway.
-- Keep readable text level and baseline-aligned. Animate a wrapper instead of skewing or rotating labels, captions, or button text.
-- Treat GSAP as the motion layer, not a substitute for typography, spacing, surface, or component geometry.
-- Use focus-frame, seekable-type, split-reveal, and glass-notification only for their declared semantic roles and record a deterministic fallback.
-- Use only GSAP capabilities already available in the project. If a plan names an unavailable plugin, use the recorded no-plugin fallback.
-- Never install ShotCraft or Remotion during design. The automatic default pack requires neither. Offer the project-local Remotion setup only after explicit user consent for a specific complex node.
-- A ShotCraft reference enriches composition and pacing; it must not replace its node's native recipe or appear during readable evidence.
-- Do not vendor a personal GSAP download into the public skill. Use project-local npm installation or selected official browser files after `tools/gsap_check.py` passes.
-- `energetic` is a coverage goal, not permission to cover evidence or stack effects mechanically.
-
-## Completion
-
-Return the chosen preset, accent, protected regions, number of planned inserts, and next phrase: `按这个设计做官方预览。`