video-use

video-use is an open-source project that enables video editing through coding agents such as Claude Code, Codex, or Hermes. Users drop raw footage into a folder, chat with the agent, and receive a finished `final.mp4`. It works across content types including talking heads, montages, tutorials, travel, and interviews without requiring presets or menus. Key capabilities include: - Cutting filler words (umm, uh, false starts) and dead space between takes - Automatic color grading with warm cinematic, neutral punch, or custom ffmpeg chains - 30ms audio fades at every cut to prevent audio pops - Burning subtitles in customizable styles (2-word uppercase chunks by default) - Generating animation overlays via HyperFrames, Remotion, Manim, or PIL, spawned in parallel sub-agents - Self-evaluating rendered output at every cut boundary before showing results - Persisting session memory in `project.md` for continuity across sessions The system works by having the LLM read video through two layers instead of watching it. Layer 1 is an audio transcript from ElevenLabs Scribe providing word-level timestamps, speaker diarization, and audio events, packed into a ~12KB text file. Layer 2 is an on-demand visual composite (filmstrip + waveform + word labels PNG) generated only at decision points. This approach avoids processing millions of tokens of raw frames. The pipeline flows: Transcribe -> Pack -> LLM Reasons -> EDL -> Render -> Self-Eval, with a self-evaluation loop that checks rendered output at cut boundaries and can fix issues up to 3 times before presenting the preview. Setup can be done via a paste-in prompt for automated installation or manually through cloning, dependency installation (uv/pip, ffmpeg, yt-dlp), and adding an ElevenLabs API key. Design principles emphasize text and on-demand visuals, audio as primary with visuals following, strategy approval before execution, zero content assumptions, and 12 hard production rules with artistic freedom elsewhere.