claude-video

Claude Video (/watch) is an open-source skill that gives AI coding assistants the ability to analyze video content. Out of the box, these agents can read webpages and run scripts but cannot process video. This tool bridges that gap by orchestrating download, frame extraction, and transcription, then handing the results to the AI's multimodal context. **Supported Hosts:** Claude Code, Codex, Cursor, Copilot, Gemini CLI, claude.ai web, and 50+ other Agent Skills hosts. **How it works:** 1. Accepts a URL (YouTube, TikTok, Loom, Vimeo, etc.) or local file path (.mp4, .mov, .mkv, .webm). 2. Uses yt-dlp to fetch native captions first (free, no download required). If unavailable, it downloads only the necessary audio/video. 3. Uses ffmpeg to extract frames based on the chosen detail mode: efficient (fast keyframes), balanced (scene-change detection), or token-burner (uncapped scene candidates). 4. Transcribes audio via Groq's whisper-large-v3 (preferred) or OpenAI's whisper-1 as a fallback. 5. The AI reads each frame as an image alongside the timestamped transcript to answer questions grounded in actual visual and audio content. **Key features:** - Frame deduplication: Drops near-identical frames (e.g., static slides) using a pure-stdlib Python mean absolute difference algorithm, saving token budget. - Detail modes: Trades speed/token cost for visual fidelity, with caps to prevent context budget exhaustion on long videos. - Focused analysis: Supports --start/--end flags for dense per-second frame budgets on specific video sections. - Zero-config setup: Auto-installs ffmpeg and yt-dlp on macOS via brew; prints exact commands for Linux/Windows. **Common use cases:** Analyzing viral content structure, diagnosing bugs from screen recordings, summarizing long videos, extracting substance from launch updates, and turning playlists into searchable notes.