claude-watch

claude-watch

Turn any tutorial or lecture video into structured study notes. Paste a URL, walk away, come back to a markdown file with embedded screenshots, timestamped transcript, and Claude's synthesis — saved to a persistent library.

/claude-watch https://youtu.be/<lecture> backprop intuition

Install

Surface	Command
Claude Code	/plugin marketplace add devinilabs/claude-watch then /plugin install claude-watch@claude-watch
claude.ai (web)	Download claude-watch.skill from the latest release → Settings → Capabilities → Skills → +
Codex	git clone https://github.com/devinilabs/claude-watch ~/.codex/skills/claude-watch

What it does

1. Downloads via yt-dlp (or accepts a local file).
2. Detects scene changes with ffmpeg. Inserts coverage-floor frames every 45s across long static gaps so a lecture with one slide for 5 minutes still gets ~7 frames, not 1.
3. Pulls a timestamped transcript — captions first (free), Whisper API (Groq preferred, OpenAI alt) only when missing.
4. Hands frames + transcript to Claude. Claude Reads every frame as an image and writes notes.md to a strict template:
   • ## TLDR — 3-4 sentence synthesis
   • ## Key Concepts — bulleted with timestamps
   • ## Notes — one section per scene with embedded screenshot, on-screen text, what was said, Claude's synthesis
   • ## Code & Commands — every code-on-screen frame transcribed into a runnable fenced block
   • ## Diagrams Referenced, ## Open Questions
5. Saves everything to ~/claude-watch/library/<slug>/ — re-running the same URL is a cache hit.

Why this exists

claude-video's uniform frame sampling spends the budget poorly on long lectures with slow-changing slides. And answers live in chat, so you can't go back to "the notes from that video." claude-watch is opinionated for the tutorial workflow: scene-aware frames, persistent library, structured notes file.

Usage

/claude-watch <url-or-path> [topic]
/claude-watch ~/Lectures/cs231n.mp4 backpropagation derivation
/claude-watch https://youtu.be/<long> --start 5:00 --end 25:00
/claude-watch <url> --resolution 1024            # for slides with tiny code text

Flags: --start/--end, --max-frames, --resolution, --scene-threshold, --max-gap, --whisper groq|openai, --no-whisper, --out-dir.

Bring your own keys

Captions cover the majority of public videos for free. Whisper only kicks in when a video has no caption track.

Need	Cost
Download + native captions	free (yt-dlp + ffmpeg)
Whisper fallback (preferred)	Groq whisper-large-v3 — cheap, fast
Whisper fallback (alt)	OpenAI whisper-1
Disable Whisper	--no-whisper (frames-only when no captions)

Keys go in ~/.config/claude-watch/.env (mode 0600).

Re-running the same video

The library is keyed on slug = YYYY-MM-DD-<title>-<short-hash> where the short hash is sha1(source + focus_range)[:4]. Re-running the same URL with the same focus range hits the cache — no re-download, no re-transcribe, only frames + notes regenerate. Different focus range = different slug = a separate notes file.

To force a fresh run, delete the meta.json in the library dir.

Limits

• Best accuracy: under ~30 minutes for a single notes pass. Past that, use --start/--end to focus.
• Hard frame cap: 80 by default. Bump with --max-frames (token cost grows linearly).
• Whisper upload limit: 25 MB (~50 min mono 16 kHz). Longer videos need captions.
• No private platforms. Public URLs and local files only.

Develop

git clone https://github.com/devinilabs/claude-watch
cd claude-watch
python3 -m pytest                         # full suite
bash scripts/build-skill.sh               # → dist/claude-watch.skill (claude.ai bundle)

Releasing: tag vX.Y.Z, push the tag — CI builds and attaches claude-watch.skill.

License

MIT. Built on yt-dlp, ffmpeg, and Claude's multimodal Read tool. Whisper transcription via Groq or OpenAI.