Video Use
Video Use is a Browser Use sibling repo that turns raw footage into edited video through an agent skill: transcript-first reasoning, EDL generation, FFmpeg rendering, subtitle/overlay passes, and rendered-output self-evaluation.
Source Snapshot
The reviewed bookmark image is a GitHub repo card for browser-use/video-use: "Edit videos with coding agents", Python, 11,658 stars in the screenshot, and 976 stars that day. Live source review on 2026-07-04 found an MIT-licensed repo created 2026-04-12, default branch main at 92c2b34e44c205cbc2acae7f6ca7c1c219d5dd66, 14,576 stars, 1,744 forks, no public tags, and no PyPI package for video-use. The repository's pyproject.toml reports version = "0.1.0", Python >=3.10, and dependencies on requests, librosa, matplotlib, pillow, and numpy, with optional manim support. Source: GitHub browser-use/video-use, 2026-07-04
What It Adds
Video Use is useful because it encodes video-editing operations as an agent skill rather than as a normal CLI or nonlinear editor. The workflow is: put raw footage in a folder, ask an agent for the desired edit, confirm the plan, generate an edit decision list, render final.mp4, inspect the rendered output, and persist project memory in project.md. The README describes a pipeline of transcribe, pack context, LLM reasoning, EDL generation, rendering, and self-evaluation. Source: browser-use/video-use README, 2026-07-04
The important design choice is transcript-first editing with on-demand visual composites. Audio drives most cuts because words, silence, filler, and pacing are cheaper to reason about than frames. Visual review still exists, but it is requested when composition, overlays, or cut-boundary checks require it. That makes the tool a practical cousin of HyperFrames and Remotion: HyperFrames/Remotion are code-rendered video routes, while Video Use is an agent-operated editing route over existing footage.
Production Rules Worth Keeping
The upstream skill contains several rules that are worth copying into any local video-editing agent:
- Never cut inside a word.
- Add small padding at cut edges.
- Apply subtitles last.
- Use per-segment lossless concatenation.
- Add short fades around edits.
- Shift overlay timestamps after cuts.
- Cache transcripts and preserve source media.
- Confirm strategy before destructive edits.
- Self-evaluate rendered output at cut boundaries.
Those rules are the real asset. They are an Agentic Quality System for video: explicit invariants, deterministic intermediate files, rendered proof, and memory for follow-up edits.
Adoption Decision
Keep Video Use as a candidate/reference tool, not a default installed skill. It requires FFmpeg and optionally yt-dlp, Manim, and an ElevenLabs key for Scribe transcription. The install guide recommends cloning to a stable local path and symlinking the whole repo into the harness skills directory, for example:
ln -s ~/Developer/video-use "${CODEX_HOME:-$HOME/.codex}/skills/video-use"
That shape fits Kevin's runtime skill-index pattern, but the source should be reviewed again before installing because the repo is fast-moving, untagged, and not package-published. If adopted, pin a commit and add a small local video fixture so the first run proves transcription, EDL, render, and self-eval without spending real project media.
Timeline
- 2026-07-04 | Created from @gregpr07's traction bookmark after local artifact review and source audit. Decision: candidate/reference tool for agent-operated video editing, not installed by default. Source: X/@gregpr07, 2026-06-29; Source: GitHub
browser-use/video-use, 2026-07-04