From idea to final cut, a video goes through seven stages. This skill is themaster coordinator: determine which step you are at now, assign the work to the corresponding sub-skill, and wrap up with a retrospective at the end.
Full Pipeline Map
参考/灵感 ──► ① 选题 space-video-topic 选题·标题·正文·下载参考·逐字稿
► ② 脚本 space-video-script 去 AI 味口播稿·精确到秒
► ③ 剪辑 space-video-edit 剪口播·删前保后·成片
► ④ B-roll space-video-broll 动效导演·HTML→30s+ MP4
► ⑤ 字幕 space-video-subtitle 转写·断句·AI 校对·烧录
► ⑥ 配音配乐 space-video-audio TTS 旁白·找 BGM·压音混音
► ⑦ 封面 space-video-cover 封面家族气质
成片 ──────► ⑧ 复盘（本技能收口）
Routing Rules
First determine what the user has and what they want, then decide the entry point:
| What the user says / what they have | Assign to |
|---|---|
| "I don't know what to shoot", "help me think of a topic", "break down this viral hit", "come up with a title" | space-video-topic |
| Send a YouTube/Xiaoyuzhou/Bilibili/Douyin/Xiaohongshu link, requesting "extract subtitles/verbatim transcript/download video" | space-video-transcript |
| "Write the script", "talking-head script", "this part sounds too AI", "second-by-second storyboard" | space-video-script |
| "Edit this screen recording", "cut the slip-ups", "make the final cut", "timeline" | space-video-edit |
| "Add some animation", "B-roll", "make a motion intro", "HTML to video" | space-video-broll(code rendering, deterministic, free) |
| "Turn an article into an illustrated video", "hand-drawn whiteboard animation", "Seedance/libtv to generate B-roll" | space-video-broll-sketch(calls video models for generation) |
| Say to a local final cut "add subtitles", "line breaking", "subtitle proofreading", "burn in subtitles" | space-video-subtitle |
| "Voiceover", "soundtrack", "add BGM", "find music", "TTS narration", "mixing" | space-video-audio |
| "Make a cover", "thumbnail", "hero image" | space-video-cover |
| "Help me make one from scratch" | Start from ①, go stage by stage |
If unsure, ask one question first: Is this video a talking-head/screen recording/mashup? Which platform (vertical 3:4 / horizontal 16:9)? Any references?
Handoff Convention for the Full Pipeline
Each stage hands off via files; do not leave the previous step's result only in the conversation:
项目目录/
├── 00_参考/ 下载的参考视频、逐字稿
├── 01_选题/ 选题.md、标题候选.md、正文.md
├── 02_脚本/ script.md（精确到秒）
├── 03_剪辑/ source_cut.mp4、cutlist.json
├── 04_broll/ *.mp4 B-roll 片段
├── 05_字幕/ subtitles.srt
└── 06_封面/ cover.png
After each step's output is saved to disk, proceed to the next step. Sub-skills do not call each other directly; this skill orchestrates them all.
⑦ Retrospective (wrap-up after the final cut)
After the final cut is published, help the user consolidate:
- Data: completion rate, like/comment ratio, follower growth, which stage might be the bottleneck
- Hook review: did the first 3 seconds hold people
- Reusable assets: which B-roll / cover template / topic formula from this one can be reused next time
- Append one entry to the user's video retrospective record (if any)
Core Principles
- Single responsibility: this skill only does coordination and retrospectives; it does not edit or write itself; specific work goes to sub-skills
- Save-to-disk handoff: each step's output is written to a file, traceable and reversible
- Platform first: decide the aspect ratio and platform before starting, to avoid rework between vertical and horizontal