The Problem

Editing video with precision—cutting filler words, syncing cuts to word boundaries, color grading, and burning subtitles—normally requires a timeline editor and hours of manual work. video-use replaces that workflow with a coding agent: drop raw footage in a folder, describe the edit, and get a finished final.mp4.

What This Does

video-use is a skill package for agentic coding tools (Claude Code, Codex). It gives the agent the scripts and reference material to transcribe footage, identify cut points, grade segments, and assemble the output. The heavy lifting lives in six Python modules under helpers/, with the agent orchestration documented in SKILL.md and install.md.

The skills/manim-video/ directory provides 15 reference documents plus a setup script for Manim-based animation overlays, spawned as parallel sub-agents.

How It Is Wired

Execution starts at main in helpers/grade.py:294, which reaches 37 functions. The pipeline is linear: main → auto_grade_for_clip → _sample_frame_stats, then main → apply_grade. The same main calls build_final_composite and load_api_key twice each.

The code touches the outside world directly: 14 functions run external commands, 14 read or write files, and 1 makes a network call. The shortest path from entry to process exit is main → auto_grade_for_clip, which shells out via subprocess.check_output(probe_cmd). The transcribe module (helpers/transcribe.py) makes the outbound network call to ElevenLabs.

File-by-file responsibility:

  • helpers/render.py — 17 functions; the hub. Owns get_preset, auto_grade_for_clip, resolve_grade_filter, resolve_path (called from 3 places, the widest blast radius).
  • helpers/grade.py — entry point; 6 functions, calls into 5 other modules.
  • helpers/timeline_view.py — 8 functions; frame extraction, silence detection, envelope computation.
  • helpers/pack_transcripts.py — 7 functions; transcript packing and phrase grouping.
  • helpers/transcribe.py — 5 functions; audio extraction, API calls, transcription.
  • helpers/transcribe_batch.py — 2 functions; video discovery and batch entry.

All six modules have zero import edges between them—a flat structure with no cycles, so changing one helper doesn't ripple through others. The cost is that shared logic (like get_preset and resolve_path) is duplicated rather than centralized.

How To Use It

Setup (from the README, verified against pyproject.toml):

git clone https://github.com/browser-use/video-use ~/Developer/video-use
ln -sfn ~/Developer/video-use ~/.claude/skills/video-use
cd ~/Developer/video-use
uv sync
brew install ffmpeg
cp .env.example .env
$EDITOR .env   # set ELEVENLABS_API_KEY

Running: Start an agent in your footage directory and describe the edit. The agent reads SKILL.md, inventories sources, proposes a strategy, and writes edit/final.mp4 next to your sources.

Real-World Use

For a talking-head interview with filler words and dead air: drop the raw takes in a folder, tell the agent to "edit this into a launch video." It transcribes via ElevenLabs, identifies filler and silence gaps, applies 30ms audio fades at each cut, color-grades segments, and burns subtitles in your chosen style. All outputs land in <videos_dir>/edit/.

Code Health & Issues

Static analysis found 6 medium findings across 4 kinds:

  • Medium (resource safety) — helpers/timeline_view.py opens files without a context manager; handle may leak on error.
  • Medium (clarity) — duplicated 6-line blocks in helpers/grade.py and helpers/timeline_view.py.
  • Medium (resilience) — broad exception handling in grade.py, render.py, and timeline_view.py swallows errors indiscriminately.
  • Medium (cognitive load) — helpers/grade.py has max indentation depth of 6; hard to follow.

SDLC gaps: no test files, no CI/CD pipeline, no lockfile. The lockfile absence is the highest-risk item—pyproject.toml declares dependencies without pinning transitive versions, so what you test may differ from what ships.

The Bottom Line

A practical, well-scoped tool that replaces a manual editing workflow with agent-driven automation. The flat module structure is easy to extend, but the lack of tests and a lockfile makes production use risky. Best suited for teams already running coding agents who want a repeatable, scriptable editing pipeline.