browser-use/video-use

Treat this as an experiment for spoken-word footage, not a replacement for a timeline editor yet. It is most interesting if you already work inside Claude Code or Codex and want rough cuts produced from a folder of takes.

The problem

A solo creator finishes recording several takes and wants a usable first cut: filler removed, dead air tightened, color adjusted, subtitles burned in. Today that means importing clips, scanning waveforms, scrubbing for false starts and hand-placing every caption. With this project, the raw files stay in a folder, the agent reads a transcript with word timestamps and silence gaps, proposes an edit strategy, and renders final.mp4 next to the sources after approval. The result is a reviewable cut rather than a hand-assembled timeline.

What's different

The central design choice is that the language model never watches frames. One ElevenLabs Scribe transcription supplies word-level timestamps, speaker labels and audio events, packed into roughly 12KB of text per set of takes. A separate timeline_view composite of filmstrip, waveform and word labels is generated only at ambiguous moments, such as a pause or a retake comparison. That keeps token use bounded without hiding the visuals entirely, and it is why cuts follow speech boundaries rather than frame-by-frame visual judgement.

The pipeline runs transcribe, pack, reason, edit decision list, render, then self-evaluation: the rendered output is checked at each cut boundary, with up to three fix-and-rerender cycles before you see a preview. The project gives the agent 12 hard production rules, covering items such as 30ms audio fades, while leaving taste to the user. The trade-off is dependence on a cloud transcription service and on the agent's judgement: the text view makes it good at dialogue-driven material and gives it little to say about fast visual action, so montages and heavy motion are a weaker fit despite the content-agnostic pitch.

Who should try it

Good fit: a podcaster, tutorial creator or founder making talking-head launch videos who already has shell-capable agents installed and wants to iterate by describing changes. The confirm-before-cut and project.md session memory suit multi-day edits of the same footage. Poor fit: an editor who needs frame-accurate compositing, or who cannot use a paid transcription API. In that case the visual composite cannot substitute for the transcript, so the workflow loses its main input.

Requirements and limitations

You need ffmpeg, Python dependencies via uv or pip, and an ElevenLabs API key; the setup instructions ask for the key once during installation. An agent with shell access is needed for the skill route, but the manual path of cloning, symlinking into the agent's skills directory, installing dependencies and adding the key documents an alternative to the paste-in setup prompt. yt-dlp is optional for downloading online sources.

Ongoing tasks include keeping the agent registration current, managing API usage for transcription, and reviewing the self-evaluation report rather than trusting it blindly. The self-eval checks cut boundaries for jumps, pops and hidden subtitles; it does not establish that the creative edit matches your intent. The output lands in an edit/ folder next to sources, so storage and organization are your responsibility.

HotRepo's take

HotRepo recorded 29,770 stars and +120 on the daily board on 2026-10-25, one appearance, with no earlier history. There is not enough comparable board data to judge whether interest persists; the repository is about six months old at that snapshot. Before investing, read the production rules in SKILL.md and confirm the transcription dependency fits your footage and budget.