ableton-describe-to-midi

name	Ableton Describe To MIDI
description	This skill should be used when the user describes music in natural language and wants it written into Ableton Live — "turn this description into MIDI", "generate a [mood] [instrument] motif", "write me an 8-bar [genre] [part] in [key]", "make a melancholic piano line in D dorian", or any free-form "make music that sounds like…" request. Translates a prose brief into a MIDI clip in the Set.
version	0.1.0

Ableton Describe To MIDI

Translate a natural-language brief ("a melancholic 8-bar piano motif in D dorian, sparse, building") into MIDI written into the Live Set. This is the "language → music in your project" magic. Two delivery modes: (A) Claude composes the notes and writes them via the runner; (B) the extension calls an LLM at runtime from Node (it has network access) and writes the returned notes.

When to use

Triggers: "turn this into MIDI", "generate a [mood/genre] [part] in [key]", "make music like…". For a known song's melody use ableton-genre-cover; for a melody over given chords use ableton-topline.

Process

1. Use ableton-live-runner for build/run/kill + scaffold.
2. Parse the brief into musical parameters — key/scale, tempo, length (bars), part(s) (chords/bass/melody/drums), density, register, mood. Ask 1–2 questions only if a critical parameter is missing (key or length).
3. Set the project context — song.tempo, and optionally song.rootNote/scale so the Set stays coherent and Ableton's scale-aware tools agree.
4. Compose — for each requested part, reuse the relevant skill's logic (chords → ableton-reharmonize voicings, melody → ableton-topline principles, bass → ableton-bassline, drums → ableton-drum-groove). Stay in scale; anchor chord tones on strong beats; phrase with rests.
5. Write tracks with makeTrack (sensible instrument per part + light FX) and report what was generated (key, bars, parts).

Mode B — generate at runtime via an LLM (optional, advanced)

Because the extension is a full Node process it has network access, so compose() can call an LLM API and parse structured notes. Caveats: verify fetch exists in Live's bundled Node (the runtime Node version is not guaranteed — polyfill or use node-fetch if it is undefined); load the API key from environment.storageDirectory, not process.env; and pin the model in one constant so it is easy to update. Force JSON output and validate before writing.

const MODEL = "claude-opus-4-8"; // update as models change
async function llmNotes(apiKey, brief, bars) { // apiKey read from environment.storageDirectory
  const res = await fetch("https://api.anthropic.com/v1/messages", {
    method: "POST",
    headers: { "x-api-key": apiKey, "anthropic-version": "2023-06-01", "content-type": "application/json" },
    body: JSON.stringify({
      model: MODEL, max_tokens: 2000,
      messages: [{ role: "user", content:
        `Return ONLY JSON: {"notes":[{"pitch":0-127,"startTime":beats,"duration":beats,"velocity":1-127}]}.
         ${bars} bars, 4/4. Brief: ${brief}` }],
    }),
  });
  const txt = (await res.json()).content[0].text;
  return JSON.parse(txt.slice(txt.indexOf("{"), txt.lastIndexOf("}") + 1)).notes;
}
// then: clip.notes = validate(await llmNotes(apiKey, brief, bars));

Validation (always, especially Mode B)

Clamp pitch 0–127, velocity 1–127, startTime ≥ 0, duration > 0; drop out-of-scale notes if a scale was requested; cap total to the clip length. Never write unvalidated model output.

Limits

• Mode B needs an API key and network; interactively-authenticated services may be unavailable in headless runs — prefer Mode A (Claude composes) for reliability.
• No automation/real-time — generates notes, not performance moves.

See ableton-topline, ableton-reharmonize, ableton-bassline, ableton-drum-groove.