claude-video teaches Claude to actually watch the video, not read the title

claude-video is a skill with one command, /watch, and it does the unglamorous plumbing that makes video legible to an agent. Give it a URL, it downloads with yt-dlp, pulls frames with ffmpeg picking scene changes and keyframes, transcribes with native captions or Whisper, and hands Claude the frames as images alongside a timestamped transcript. Then you can ask what happens at the 30 second mark and get an answer grounded in what's actually on screen, not scraped metadata. v0.2.0 shipped July 1, it's at 4.1k stars, and it jumped up the GitHub trending list today.

The reason this is useful is that video has been a blind spot for coding and research agents. Your bug repro is a screen recording. The competitor's hook is 40 seconds into a TikTok. The one thing you need from a two-hour talk is a single slide. Pasting a YouTube link into a chat gets you a summary of the description; this gets you the actual content, frames and audio, in a form the model can reason over.

Like OfficeCLI, it's distributed as a portable skill that drops into 50-plus agents, Claude Code, Cursor, Codex, Copilot. That's the pattern worth noticing: capabilities are shipping as SKILL.md packages you bolt onto whatever agent you already run, not as standalone apps. The agent stays the same, the senses keep multiplying. This week alone it learned to read your spreadsheets and watch your videos. Link: github.com/bradautomates/claude-video