youtube-extract

Turn one YouTube URL into verified, durable knowledge. Requires yt-dlp + ffmpeg. Work in $CLAUDE_JOB_DIR/tmp (fallback mktemp -d). URL="<video>".

1. Metadata + links (no download)

yt-dlp --skip-download --dump-json "$URL" > meta.json
python3 -c "
import json,re; m=json.load(open('meta.json'))
print(m['title'],'|',m['uploader'],'|',m['upload_date'],'|',m['duration_string'])
for c in (m.get('chapters') or []): print(f\"{int(c['start_time'])//60}:{int(c['start_time'])%60:02d}\", c['title'])
print(*sorted(set(re.findall(r'https?://[^\s)\\\"]+', m['description']))), sep='\n')"

Classify links: github / arxiv-whitepaper / docs / promo-noise. Keep the first three kinds.

2. Transcript