Claude Video: letting your agent read a video without burning your to…
A Claude Code skill (MIT) that makes a video legible to an agent. You paste a URL or a local path along with your question, the command downloads the video, grabs existing captions when they exist (free, through yt-dlp) and falls back to a Whisper transcript otherwise. Frame extraction then runs on one of three strategies: fast keyframes, scene-change detection, or exhaustive coverage, with a screenshot budget that scales with duration.
Strengths
- The transcript drives the screenshots, so token cost tracks content rather than duration
- Existing captions are fetched first, so free whenever the source publishes them
- Three extraction strategies, so you arbitrate cost against thoroughness yourself
- One-command plugin install, and available outside Claude Code through NPX
Limitations
- Whatever is not spoken out loud escapes it: gestures, editing, visual intent
- With no published captions, transcription goes through a Whisper key billed per source
- Depends on yt-dlp and ffmpeg installed locally, and on updating them when a platform shifts
Best for
- Stripping a long talk down to quotable points without watching it
- Auditing a product demo: what the tool actually does on screen, not what the voice promises
- Turning a recorded meeting into written notes without replaying the tape