An AI post-production suite hiding inside Claude Code

What it does

This project is a bundle of agent skills for Claude Code, Codex CLI, and similar tools that turns a single natural-language sentence into a full video-recap pipeline. Point it at a local video, ask for a ten-minute Chinese narration clip, and the agent analyzes scenes, writes a script, slices the footage, synthesizes voiceover via Xiaomi MiMo or Fish Audio, mixes audio, and burns in subtitles. The final product is a rendered MP4, plus an optional CapCut draft if you want to tweak the timeline by hand.

The interesting bit

Instead of one monolithic script, the work is split into six file-system-based skills—video-understanding, video-script, video-cut, video-voiceover, video-assemble, and video-recap—so the agent can pause, resume, and reuse prior analysis without a database. Scene breakdowns, narration scripts, and clip plans are plain JSON and Markdown stored on disk, which means you can grep them, diff them, or hand-edit them between runs. It treats post-production like a stateful, versionable workflow rather than a black-box API call.

Key highlights

• Multi-source editing: feed the agent several videos and it will weave them into a single narrative timeline rather than producing separate summaries.
• Cut-first mode edits the visual timeline before writing narration, so the voiceover naturally locks to the picture without awkward remapping.
• Optional CapCut export produces a schema-driven multi-track draft with original footage, narration, BGM, and subtitles for human refinement.
• You can inject background research as JSON to help the vision model recognize characters or plot context before it starts cutting.
• Advisory-only quality review: MiMo can critique the cut or script, but a flaky API or rate limit will not block the final render.

Caveats

• English-to-Chinese voice replacement currently supports only a single speaker and does not separate background music, so translated tracks may lose their original score.
• Hard-subtitle detection requires square-pixel video with bottom-aligned subtitles, a constraint the README explicitly flags as narrow.
• Every AI stage—ASR, vision-language analysis, and TTS—depends on the Xiaomi MiMo cloud API, so local execution still requires an active key and internet connection; there is no offline fallback.

Verdict

Ideal for developers or creators who need fast Chinese-language recaps and dubs without opening a traditional NLE. If you demand pixel-perfect manual control or need to work offline, this agent will frustrate you.