hellyee: Skill for Claude Code

Remix & Reference

Both workflows stand on three tools: analyze_audio_file (offline: key, tempo, levels, band energies), separate_stems (demucs), and import_audio (file → User Library → session clip, with the view/indexing dance handled).

Reference channel

Producers keep a finished commercial track on a muted channel and check their mix against it. The hellyee version:

1. analyze_audio_file(reference) first — key, tempo, band_energy_pct and crest ratio are the comparison targets.
2. import_audio(reference, track_name="REF"), then immediately set_mixer(ref_track, mute=True). A reference must never play in the mix or reach a render.
3. A/B by solo: to audition, set_mixer(ref_track, mute=False, solo=True) and fire its clip; to return, solo off + mute on. Warn the user: the reference passes through the master chain too, so they hear it limited — for honest comparison offer to temporarily disable master devices ("Device On" = 0) during the A/B, and restore after.
4. Level-match before judging anything: play the reference solo, measure the master meter; play the drop, measure again. Louder always sounds "better" — match within ~0.02 before comparing character.
5. Spectral comparison is offline, not live. The API cannot render or spectrum-analyze the project; compare the reference's band_energy_pct against what you know of the mix (which tracks occupy which bands, from their filters/EQ) and against per-track meters. State plainly this is analysis + inference, not a measured spectrum of the user's master.
6. Typical use of the numbers: reference has 25% sub+bass and your bass region reads thin → raise bass/kick region or tame lowmid; reference crest_ratio 6 vs your obviously-dynamic master → more limiting for that target, and vice versa.

Remix workflow

The user provides a song and instructions ("vocals only over a new techno beat", "halftime flip of the drop"). The pipeline:

1. Analyze the source — analyze_audio_file: its key, tempo, and where its energy lives. Report findings before touching anything.
2. Separate — separate_stems(path) → vocals / drums / bass / other (or two_stems=True for vocals + instrumental). Takes 30–90 s for a full song; say so.
3. Import the stems the plan needs — one import_audio per stem, each on its own named track (VOCAL STEM, DRUM STEM…). Mute the ones not in use rather than skipping them; the user may change their mind.
4. Decide the tempo relationship and say it out loud:
   • Project follows source (set_tempo to the source's tempo) — safest, stems play natural.
   • Source follows project — needs warping; auto-warp guesses wrong tempos on stems, so leave clips unwarped for rubato material (vocals) and prefer tempo-matching the project instead. Never toggle warp on→off (region clamp — see the hellyee skill).
5. Key: new musical material follows the stem's key (get_scale_notes), or transpose the stem ±(≤3) semitones with set_audio_clip(pitch_coarse=…) — more sounds artifacted. The same tool's gain sits a stem into the mix without re-exporting it. An autotune plugin locked to the target scale (MAutoPitch, per-note switches) covers vocal drift.
6. Rebuild with the genre-blueprints skill for the target style, treat imported stems as first-class layers: EQ them into their slot (high-pass a vocal stem at ~120 Hz; a drum stem fights your kick — carve or use only its top via steep high-pass), give vocals reverb/delay via sends.
7. Arrange — stems are long clips; place segments by bar using place_in_arrangement, silence sections of a stem with gap clips over the unwanted bars.
8. Balance and master by measurement as always. If the remix should sit against the original, the original is the reference channel — run the reference workflow with it.