Gives Claude video understanding by extracting frames and transcript, then feeding both back for analysis. Downloads with yt-dlp, pulls captions when available, falls back to Whisper API when not. Auto-scales frame rate based on duration to stay under 100 frames, hitting about 2 fps for short clips and spacing out for longer ones. Works on YouTube, Vimeo, TikTok, local files, most things yt-dlp handles. Supports time ranges if you want dense coverage of a specific moment instead of a sparse scan of the whole thing. The focused mode is the real win for long videos where the question is about one section. Setup handles ffmpeg and API keys through an idempotent installer.
npx skills add https://github.com/bradautomates/claude-video --skill watch