Making Claude watch videos so you don't have to

What it does Claude Video is essentially well-engineered glue: a portable skill for Claude Code, claude.ai, and Codex that orchestrates yt-dlp, ffmpeg, and optional Whisper to close the gap between Claude’s text-and-code prowess and the video format it otherwise can’t ingest. Feed it a public URL or local file and it produces timestamped frames and a transcript so Claude can Read the visuals in parallel and answer questions grounded in actual on-screen action and audio rather than guessing from titles.

The interesting bit The craft is in the frame budget logic. The skill auto-scales extraction rates based on duration, hard-capping at 100 frames and 2 fps so a 30-minute video doesn’t burn your token limit. Focused --start/--end windows swap coverage for density, which is far more useful than a sparse skim of an entire livestream.

Key highlights

• Supports hundreds of public platforms via yt-dlp plus local .mp4, .mov, .mkv, and .webm files.
• Native captions are free and preferred; Whisper via Groq or OpenAI is a fallback for caption-less sources like TikToks or screen recordings.
• macOS auto-installs dependencies on first run; other platforms get exact manual commands printed to the terminal.
• Useful for debugging UI glitches from screen recordings, analyzing competitor video hooks, or summarizing talks without pressing play.

Caveats

• Videos over 10 minutes trigger a “sparse scan” warning; meaningful analysis requires narrowing the window with --start/--end.
• No authenticated platforms: if yt-dlp can’t fetch it without logging in, neither can this skill.
• Whisper audio uploads are limited to 25 MB (~50 minutes at mono 16 kHz), so long caption-less content may need to be chunked.

Verdict Worth it if you want Claude to audit visual content, debug screen recordings, or summarize short-to-medium videos without watching them yourself. Less useful if you need deep, frame-dense analysis of long-form content without slicing it into segments first.