Give a topic, produce an MP4:Visuals 100% drawn by Remotion + React/SVG code(precise, parameterizable, auto-reflow on data changes),Voiceover generated by TTS and driving the timeline by measured duration(narration is never cut off by visuals; default edge-tts neural voice,--engine clone switchable to custom cloned voices for more realistic speech), three-layer audio mixing, deterministic rendering, dual-version delivery.
Boundaries with other video skills
| Content characteristics | Which to use |
|---|---|
| Diagrams/data/structure are the star (matrices, connections, charts, formulas, flows) | This skill |
| Narration-script-driven, want 12 style quick cuts, karaoke word-by-word subtitles | voice-to-video |
| Industry conference talk videos (HTML PPT + digital human + lip sync) | talk-video-studio |
| Replicating real product pages (screenshots + 2.5D camera moves) / live-action presenter | This skill is not suitable |
Templates are based on 1920×1080 landscapedesign; portrait requires changing the Composition dimensions and reflowing the layout yourself.
Iron rules (to be followed every time)
- Visuals grow from code: all graphics are drawn with React/SVG components, parameters come from data;Prohibitedscreenshots as animation, prohibited AI video model-generated visuals (formulas will garble, connections will drift).
- One narration segment = one scene; scene duration =
max(visualMin, 配音实测帧数 + tail(默认40f)), automatically rebuilt into a timeline bybuild-timeline.mjs—changing the script/speech rate requires no manual adjustment of any shot, visual out-points (fade(frame, D)) follow automatically. - Deterministic rendering: prohibited
Date.now()/Math.random()/ network/environment variable dependencies; pseudorandomness alwaysmulberry32(seed). The same input must render the same frames — this is the prerequisite for regression and iteration. - Three audio layers independently verifiable: VO (main) + BGM (
bgminputProp switch, producing a no-BGM version for post-production scoring) + SFX pin-frame table (from: SHOTS.x.from + offsetrelative frame expressions, automatically following timeline shifts). - QA still frame upon completion of each shot(
npx remotion still --frame=...), don't save it all for the end; after the full render, extract frames to review key shots + ffprobe/volumedetect to validate audio tracks. src/timeline.tsandsrc/types.tsare generated by scripts, do not hand-edit; hand-editing data only touchesVOICEOVER_ZH.md、build/scenes.json、DESIGN.mdthree places.
Eight-step pipeline
0 环境自检 node scripts/check-env.mjs (Node/ffmpeg/ffprobe/Python/edge-tts)
1 简报 写 DESIGN.md 头部:主题/观众/时长/风格/渠道
2 旁白+分镜 VOICEOVER_ZH.md(一段=一场景,单段≤120字)
+ DESIGN.md 分镜表(id/画面/动画模式/visualMin/SFX 计划)
3 搭工程 Copy-Item -Recurse assets\template-project → my-video
npm install && npx tsc --noEmit (模板自带静音占位,开箱可渲)
4 配音 python scripts\tts.py --script VOICEOVER_ZH.md --out . --voice yunyang
→ public/vo/sceneN.mp3 + build/durations.json(ffprobe 实测)
要真实人声:加 --engine clone --ref-audio 参考音频 --ref-text 逐字稿
(voice_id 自动存档复用;详见 workflow.md §4)
5 时间线 手写 build/scenes.json(id + visualMin)
node scripts/build-timeline.mjs → timeline.ts + types.ts + SRT
6 场景实现 src/Video.tsx 每场景一个组件;逐镜头仍帧 QA
7 声音设计 SFX 钉帧表 + BGM;规则见 references/sound-design.md
8 渲染交付 node scripts/render.mjs --mode both --concurrency 8
→ final.mp4 + final-nobgm.mp4 + subtitles.srt + out/qa 静帧
Operational details, outputs and known pitfalls for each step:Read references/workflow.md first before starting.
File map
SKILL.md 本入口
references/workflow.md 八步流程细节与坑(执行前必读)
references/animation-vocabulary.md 8 种动画模式 + 代码(矩阵点亮/连线生长/数字滚动/弹簧卡片…)
references/sound-design.md 三层音频 + 钉帧规则 + 混音验收
scripts/check-env.mjs 环境一键自检(跨平台)
scripts/tts.py 旁白 → 分段配音 + durations.json(edge-tts / --engine clone 克隆音色双引擎;增量缓存/限流重试)
scripts/build-timeline.mjs durations + scenes.json → timeline.ts + types.ts + SRT
scripts/render.mjs 单次渲染 + FFmpeg 派生双版本
scripts/selftest.mjs 回归自测(解析/校验/时间线/字幕,无需网络)
assets/template-project/ 可直接拷贝的 Remotion 工程骨架(lib/theme/primitives/audio)
examples/case-study-long-context.md 实战复盘:《1M 上下文的秘密》125s 成片
Revision cost comparison (explain to users according to this)
- Change narration → rerun §4 (tts.py) + §5 (build-timeline.mjs) + re-render; scene durations are automatically rebuilt
- Change visuals → edit Video.tsx components + per-shot QA + re-render
- Change BGM → change
public/bgm/bgm.mp3+ re-render (the visuals do not need to be re-rendered, see render.mjs--forcesemantics)
Deliverables (must be complete every time)
final.mp4(with BGM) + final-nobgm.mp4(for adding your own music in post) + build/subtitles.srt(sentence-by-sentence subtitles) + out/qa/ still frames; and report to the user: the final video path, duration, the differences between the two versions, and which step to rerun for script changes vs. visual changes.