whistlegraph toolchain #

Pull a whistlegraph TikTok and read its musicality — the front of the pipeline that turns a whistled clip into a /pop track.

Usage #

cd toolchain/whistlegraph

node grab.mjs <tiktok-url>          # download + analyze one clip
node grab.mjs --latest 3           # newest 3 from @whistlegraph
node grab.mjs --account @someone --latest
node grab.mjs <url> --redo         # refetch + re-analyze
node grab.mjs --list               # summarize downloads/INDEX.json

Each run:

1. yt-dlp → downloads/<account>-<id>.mp4
2. ffmpeg → mono 44.1k .wav
3. analyze.py (pop .venv librosa) → tempo, key, whistled-melody note sequence + onsets → <id>.analysis.json
4. updates downloads/INDEX.json

What's tracked #

*.mp4 / *.wav are gitignored (reproducible from the URL). The small *.analysis.json and INDEX.json are committed so the melody read survives without re-pulling from TikTok.

→ /pop #

Take the key, tempoBPM, and melody[] from the analysis JSON and compose bottom-up from AC instruments (never Suno end-to-end). See pop/SCORE.md and pop/lib/.

⚠️ Re-posting these clips to Instagram #

grab.mjs pulls via yt-dlp, so the .mp4 usually carries the TikTok watermark. As of Instagram's April 2026 originality expansion, visible cross-platform watermarks make a Reel ineligible for recommendation to non-followers — so a watermarked TikTok export posted to IG gets no cold-start reach. If funnelling @whistlegraph clips to IG, re-render from a clean (watermark-free) source, not the yt-dlp download. Background + sources: social/research/instagram-official-guidance.md.