YouTube Watch — Communication skill for Claude Code

youtube-watch

A skill that lets Claude watch a video. Version in [`VERSION`](VERSION), changes in [CHANGELOG.md](CHANGELOG.md).

Watch any YouTube video the way a human would — see the frames, hear the audio, cross-reference what was shown with what was said.

https://github.com/user-attachments/assets/31727c53-8940-40c7-a324-eae3213855a1

A Claude Code (or any Anthropic-SDK / MCP-compatible) skill that:

1. Downloads the video and its transcript via yt-dlp.
2. Extracts frames with ffmpeg, smart-sampled based on what the speaker is actually referencing (zone-based extraction, not blind 1-frame-per-N-seconds).
3. Builds understanding.md: the author's title/description/chapters, then a timeline where each frame is followed by what was said until the next frame — so the AI never has to cross-reference by hand. Long videos are split into sections with a contents list, overview.jpg (one picture of ~20 frames) shows the whole video at a glance, and sheets/ packs every frame, 16 per labelled picture, so the AI looks at all of them in a few reads.

Why content-aware frames

Sampling blindly (one frame every N seconds) wastes most frames on a speaker's face, and sampling only at scene changes misses a static slide the speaker keeps pointing at. This skill reads the transcript first, classifies the video (tutorial-code / demo-ui / slide-presentation / livestream-chat / talking-head / vlog / mixed / no-speech) and samples **densely** (1 frame / 2-3 s) around the moments the speaker says "look at this", "click here", "next slide", "let me code", and **normally** (1 frame / 8-15 s) elsewhere, plus a frame at every scene change — hard cuts and gradual ones (a fade, a motion graphic building up) once the new picture holds, while a handheld or panning camera does not count.

How it compares to claude-real-video

[claude-real-video](https://github.com/HUANGCHIHHUNGLeo/claude-real-video) (crv) does a similar job and is worth a look. The differences, going b