Music gen skill

Music Generation

This skill generates music by composing it as MIDI (where Claude's music-theory knowledge is strongest), then rendering MIDI to real instrument sounds with a SoundFont. It falls back to direct waveform synthesis when no MIDI toolchain is available.

Decision: which route?

1. Melodies, chords, arrangements, songs → MIDI route (default). Compose with mido, render with scripts/render.sh. Best quality, real instruments.
2. Tones, beeps, sirens, sweeps, simple SFX → direct synth with sox/ffmpeg. No SoundFont needed. See "Direct synthesis" below.
3. "Generate a track that sounds like <artist/genre>" (AI text-to-music) → this skill does NOT do that. Tell the user it needs a model (local MusicGen via audiocraft, or a cloud service like Suno/Replicate behind an MCP server), not system utilities, and offer to help set that up separately.

Step 0 — Detect the toolchain (always do this first)

Run the probe before composing so you know which route is viable:

bash scripts/render.sh --check

It reports which of fluidsynth, timidity, sox, ffmpeg, python3 + mido are present, and whether a SoundFont was found. Do not assume tools exist — Claude Code runs on many machines. If something needed is missing, tell the user the exact install command for their OS instead of failing silently:

• macOS (Homebrew): brew install fluid-synth sox ffmpeg && pip install mido
• Debian/Ubuntu: sudo apt install fluidsynth fluid-soundfont-gm sox ffmpeg && pip install mido
• The GM SoundFont (FluidR3_GM.sf2 or default-GM.sf2) is the one part people most often lack. render.sh looks in common locations; if none found, point the user to the fluid-soundfont-gm package (Linux) or a free download.

Step 1 — Compose the MIDI

Write a Python script using mido that builds the MIDI programmatically. Keep real musical structure in mind: pick a key and tempo, build chord progressions, voice them across tracks (melody / chords / bass / drums on channel 9), use note velocities for dynamics, and respect note-off timing so notes don't hang.

scripts/melody_example.py is a working reference — a short multi-track piece in C major. Read it for the patterns (track setup, program changes, note on/off deltas, the channel-9 drum convention) and adapt rather than copying blindly. General MIDI program numbers matter: e.g. 0 = Acoustic Grand Piano, 24 = Nylon Guitar, 32 = Acoustic Bass, 48 = Strings. Drums always go on channel 9.

Save the generated MIDI to the working directory (e.g. song.mid).

Step 2 — Render to audio

bash scripts/render.sh song.mid song.wav          # WAV via best available renderer
bash scripts/render.sh song.mid song.mp3          # auto-converts to MP3 if ffmpeg present

render.sh tries fluidsynth first (best), then timidity, picks up a SoundFont automatically, renders at 44.1 kHz, and converts to MP3 with ffmpeg if the output name ends in .mp3. It exits non-zero with a clear message if no renderer or no SoundFont is available — read its stderr and relay the fix to the user.

Direct synthesis (route 2, no MIDI)

For tones and effects, skip MIDI entirely:

sox -n tone.wav synth 3 sine 440                              # 440 Hz, 3 s
sox -n chord.wav synth 4 sine 261 sine 329 sine 392 fade 0 4 1  # Cmaj chord, fade out
sox -n sweep.wav synth 2 sine 200-2000                        # rising sweep
ffmpeg -f lavfi -i "sine=frequency=440:duration=3" tone.wav   # ffmpeg equivalent

For melodies without a SoundFont, you can also generate raw samples in pure Python with wave + numpy (additive synthesis / simple envelopes) — usable but sounds synthetic; prefer the MIDI route when a SoundFont exists.

Output conventions

• Write deliverables the user keeps to the outputs directory if one exists; tell them the path and present the file.
• Default to WAV for fidelity; offer MP3 if they want something smaller/shareable.
• Mention tempo, key, and instrumentation in your summary so the result is editable.
• Keep generated MIDI alongside the audio — it's the editable source.

Notes

• This skill is deterministic: same script → same audio. There is no randomness unless you add it. If the user wants variation, seed and expose it.
• Do not claim "AI-generated music" — this is algorithmic composition + sampled instruments, which is a different (and honest) thing.