Video-use: Claude Code edits your video without watching it

browser-use, the team behind the web-automation agent everyone forked, just dropped video-use, and it's a clever one. Drop raw footage in a folder, chat with Claude Code, get final.mp4 back. It cuts filler words and dead air, color grades every segment, adds a 30ms audio fade at every cut, burns in subtitles, and generates animation overlays through Manim, Remotion or PIL.

Here's the trick that makes it work. The model never actually watches the video. It reads two layers instead. Layer one is an ElevenLabs transcript with word-level timestamps and speaker labels. Layer two is a timeline-view tool that renders filmstrip-plus-waveform PNGs, but only for the exact time range it's deciding a cut on, called on demand at decision points. So editing a 20-minute video turns into a text problem instead of a vision problem, which is cheap and fast, and that's the entire reason it's usable at all.

It ships as a Claude Code skill, a SKILL.md you symlink into your skills folder, and it runs in Codex, Cursor and OpenClaw too. This is the OpenMontage thesis showing up again: creative production is turning into something an agent does with a skill file, not something you do by hand in a timeline. Honest caveat, it's brand new with no releases cut yet, so treat it as a sharp demo of where this is going rather than a finished product. Link: github.com/browser-use/video-use