video-use edits video from transcripts, not frames

video-use edits raw footage by feeding Claude Code word-level transcripts instead of video frames, per its GitHub repo. That matters because most models are still weak at reasoning over raw video, so video-use routes around the problem instead of solving it.

Drop clips in a folder and chat with Claude Code about the edit; it assembles a final.mp4 without touching a single frame. It reads ElevenLabs Scribe transcripts with word-level timestamps as the primary signal and pulls in filmstrip-and-waveform composites only when it needs visual context. Before touching anything, it proposes a cut plan for approval.

Once you approve the plan, video-use cuts at audio boundaries and silence gaps. It also strips filler words, adds 30ms fades between cuts, and burns in subtitles. It runs a self-evaluation pass on the finished render before showing you the result.

The repo doesn't say how it handles footage where the story is visual, silent b-roll with no dialogue to time cuts against. Transcripts are a good interface for AI video editing right up until the footage stops talking.