Open-source music-skill lets Claude compose note by note — 7 pieces, 5 models, one prompt each

BandicootNo3152 · reddit · 2026-11-12

A developer released music-skill (v0.6.0, MIT): instead of an audio-generation model, an agent composes every note itself and renders it headlessly. Output includes MP3, MIDI, per-part stems and an editable project file, with no usage restrictions. Seven example pieces — game battle theme, orchestral overture, city pop, organ toccata, lo-fi, baroque fughetta, 8-bit — each came from a single prompt across five models (Claude Fable 5.2, Opus 5.5, Sonnet 5.5, GPT-6 Astra/Luna).

Workflow: nearly all code was written by Claude Code; the arrangement/style guides the AI reads were compiled by Claude too; before each release another model reviews the code and Claude must reproduce a bug before fixing it.
Can't hear: since Claude can't listen, the author had it build its own measurement tools checking loudness, spectrum and part balance after rendering; the piano roll in the video was drawn by Claude from project files.
Origin & upsides: born from making a visual-novel skill and wanting on-demand scoring; strengths are headless generation, freely editable output and seamless loops that fit video/game agent workflows.
Limits: heavy disk usage (synthesis, not generation), no vocals, long render times at high effort, below commercial libraries (synth side includes open-sourced Surge XT).

Repo: github.com/owoevan-dev/music-skill; example prompts, composing scripts and project files are coming to the examples folder.

Original post →

Open-source music-skill lets Claude compose note by note — 7 pieces, 5 models, one prompt each

More from coding & agent