CLI that turns a YouTube guitar-lesson video into a PDF of the guitar tab.

It downloads the video, samples frames, uses Claude vision to locate the tab region, crops every frame to that region, de-duplicates the crops by the bar number printed on each line of the score, and stitches the distinct tab lines vertically into a PDF. It works out of the box with no configuration — the PDF is written to out/<video-title>.pdf, with the video title as a heading on the first page and in the document metadata.

• Node.js ≥ 20
• yt-dlp and ffmpeg on your PATH (brew install yt-dlp ffmpeg)
• An Anthropic API key

npm install
npm run build
cp .env.example .env   # then put your ANTHROPIC_API_KEY in it

# with a .env file
node --env-file=.env dist/cli.js "https://www.youtube.com/watch?v=WgU5tDGC-Vc"

# or with the key already exported
export ANTHROPIC_API_KEY=sk-ant-...
node dist/cli.js "https://www.youtube.com/watch?v=WgU5tDGC-Vc"

During development you can skip the build step:

node --env-file=.env --import tsx src/cli.ts "<url>"

The result is written to out/<video-title>.pdf (its path is also printed to stdout); progress goes to stderr.

Everything has a sensible default; you normally only need the URL.

-i, --interval <seconds>   screenshot interval (default 2)
--model <id>               Claude vision model (default claude-sonnet-5)
--sample <n>               frames sampled for tab-region detection (default 6)
--dedup-threshold <n>      pre-dedup Hamming distance, cost control (default 12)
--max-height <px>          cap download resolution (default 720)
--keep-temp                keep intermediate frames/crops

1. Download — yt-dlp fetches the video (≤ --max-height).
2. Frames — ffmpeg extracts one frame every --interval seconds.
3. Detect — a labeled horizontal-band grid is drawn on --sample frames and Claude vision reports which bands contain sheet music. The median first/last music band (robust to an outlier) gives the vertical extent; width is kept full. (Coarse labeled bands are far more reliable than asking a vision model for precise pixel coordinates.)
4. Crop — sharp crops every frame to that region.
5. Pre-dedup — a dHash perceptual hash drops near-identical consecutive crops. This is only a cost control to reduce the number of vision calls in the next step.
6. Bar-number dedup — Claude reads the measure/bar number printed at the start of each line and whether the crop is real sheet music. The tool keeps exactly one crop per distinct bar number (first appearance wins) and drops non-tab crops (title cards, intros/outros). Because the bar number is constant while the playback cursor sweeps a line and only changes when the score advances, this collapses all the near-identical cursor frames of a line into a single page.
7. PDF — pdf-lib stacks the distinct tab lines vertically down A4 pages, in the order they appear in the video. The video title (read from yt-dlp) becomes the file name, a heading on the first page, and the document metadata title. Output: out/<video-title>.pdf.