video-use

An MIT agent skill from browser-use that edits video by conversation: drop raw footage in a folder, tell Claude Code “edit these into a launch video”, get edit/final.mp4 back. 18.6k stars, 2.3k forks, 45 open PRs at ingest — the most-starred single skill in this spoke by an order of magnitude. Installed by symlinking the repo into ~/.claude/skills/video-use; needs ffmpeg, an ElevenLabs API key, optionally yt-dlp.

Tier: T2 — first-party README, primary on itself, no independent evaluation. The star count measures attention, not production use (openclaw is the standing reminder).

Text-first: the whole design in one decision

video-use does not look at the video. One ElevenLabs Scribe call per source returns word-level timestamps, speaker diarization and audio events; that becomes a ~12KB markdown file (takes_packed.md) the model reasons over. Cuts land on silence gaps and speech boundaries, which are decisions transcripts can express precisely. Visual composites — filmstrip plus waveform plus word labels — get rendered only at decision points, not swept across the timeline.

Compare watch-skill on the same modality: it also refuses fixed-interval frame dumping, but spends a budget (≤100 frames) on scene-detected, hash-deduped frames plus OCR, because its job is answering questions about what a video shows. video-use is editing what a video says, so the transcript is the whole ballgame and frames are a spot-check. Same discipline, different derived representation, chosen by what the task actually needs. See text-first-agent-design.

Pipeline

Transcribe → pack → LLM reasoning → EDL → render → self-evaluate, looping up to 3 times.

Around it: filler and dead-air removal (“umm”, false starts, silence), per-segment color grading (cinematic warm, neutral punch, or a hand-written ffmpeg chain), 30ms audio fades at every cut so splices don’t pop, burned-in subtitles (2-word uppercase chunks by default), and overlay animations farmed out to parallel sub-agents driving HyperFrames, Remotion, Manim or PIL — agent-orchestration as a rendering detail. Session state persists in project.md, the same markdown-memory shape as recall and claude-md.

The user approves a cutting strategy before anything renders. That checkpoint is agent-guardrails‘s reversibility bound in its cheapest form: gate the plan, not each action.

The self-evaluation loop closes what mobilerun leaves open

After rendering, the agent inspects its own output at the cut boundaries and re-renders if the cut is wrong, up to three passes. This is loop-engineering‘s feedback signal on a rendered artifact — the same move as watch-skill‘s THE LOOP and ai-job-search reading its compiled PDF, now on video the agent itself produced.

Worth putting next to mobilerun, ingested the same day: a tap on a phone has no artifact to inspect, so its 8.6% failure rate is unverifiable from inside the loop. A video render does. Same week, same spoke, two agents driving media-heavy surfaces, and only one of them can check its work. The dividing line is whether the action produces something inspectable before it counts, not how sophisticated the agent is.

The claim it doesn’t back

The README asserts this beats presets and traditional NLEs for any content type from talking heads to tutorials. Nothing measures that — no comparison against a human editor, no rubric on the finished cuts, no failure taxonomy. The self-evaluation loop judges cut boundaries, which is a mechanical check, not a judgment about whether the edit is any good. Filed under the spoke’s standing Benchmarks gap.

A second answer arrived four days later

open-edit (VEED, 2026-08-07) does the same job and splits from this design at the authoring step: both read a per-word transcript, but OpenEdit has the agent write HTML and CSS for the caption design rather than reasoning in text to the finished cut. It also swaps ffmpeg for a closed renderer that skips the headless browser, abstracts the transcription provider instead of fixing on one, and runs only on Apple Silicon. What it lacks is this page’s self-evaluation loop — nothing inspects the render and re-cuts. The two together make agent video editing a real comparison rather than one repo’s opinion, and the star gap (18.6k vs 209) is a distribution fact, not a quality one.

text-first-agent-design · browser-use · open-edit · watch-skill · mobilerun · agent-skills · loop-engineering · agent-orchestration · agent-guardrails · recall · pxpipe