The CLI pulls the video with yt-dlp, samples frames with ffmpeg, then uses Claude’s vision to locate the tab region in the frame and again to read the bar number printed at the start of each line — that number is what lets it collapse dozens of near-duplicate frames of the same line sweeping past into a single crop. The distinct lines get stitched into an A4 PDF titled after the video. It’s a free, self-hosted script rather than a hosted service: it needs Node 20+, yt-dlp and ffmpeg on your PATH, and your own Anthropic API key to run.