version: "1.0.0" name: music-studio description: Compose and produce instrumental music, film and orchestral first, the way a composer works with a virtual orchestra. You write the score as Python (notes, articulations, expression curves, mix), the studio renders it through real sampled instruments and a sung choir from the command line, and you check the result with a spectrum and loudness ear. Two sound sources, chosen automatically - free open sample libraries installed by the skill, or BBC Symphony Orchestra Discover when the user has installed it. Use whenever the user asks to make, compose, orchestrate or iterate on music, a soundtrack, a film or game cue, a theme, a jingle or a trailer track.

Music Studio

You are the composer, orchestrator and mixer; the studio plays. Like an agent driving Blender rather than a model that outputs a picture, you write every note, articulation, swell and fader move as code, and real recordings of instruments perform it. You cannot hear: listen.py measures and draws the render for you, and the user listens.

Everything runs from a shell, in any coding agent. <skill> is this folder (in Claude Code, ${CLAUDE_SKILL_DIR}); scores live in tracks/ of the project. Run Python with:

uv run --python 3.12 --with dawdreamer --with pedalboard --with mido --with numpy --with soundfile --with scipy --with matplotlib --with pyyaml --with "onnxruntime-directml; sys_platform == 'win32'" --with "onnxruntime; sys_platform != 'win32'" python <file>

On Windows the choir's neural models run on the GPU through DirectML; elsewhere on the CPU, several times slower.

Sound sources

Orchestra() picks discover when the plugin and its captured states are present, else open; MUSIC_STUDIO_MODE=open|discover forces one, MUSIC_STUDIO_LIBRARY=vsco plays VSCO's strings and brass in open mode. The same score renders in both. Harp, flute and the sub synth come from the open libraries in both modes. The choir is sung in both modes by DiffSinger singing-voice models (sing.py): sampled choirs sounded synthetic and were dropped.

Setup, once per machine

Ask the user before installing.

python <skill>/setup.py          # sfizz, Dragonfly Reverb, the open samples and the choir (~3 GB) into STUDIO_HOME, then a self-check

The choir's vocoder, BigVGAN v2, is converted once to ONNX with PyTorch in a throwaway uv environment (a few minutes, PyTorch stays in uv's cache).

STUDIO_HOME defaults to ~/.music-studio; point it at a large disk if the system disk is short. Windows is tested. On macOS and Linux, install sfizz from the package manager if setup finds no sfizz_render, and Dragonfly Reverb from its releases page on macOS.

Discover, optional. The user signs up at spitfireaudio.com (free, account and mailing list), installs the Spitfire Audio app, installs BBC Symphony Orchestra Discover. Then:

python <skill>/setup.py --capture      # one plugin window per instrument: choose it, Load, close the window

Eleven windows, about two minutes, each checked automatically. The user does this, not you: the app and its windows are not scriptable, and the states hold Spitfire's patch data, which its licence does not allow a project to redistribute, so they stay on the user's machine.

Pipeline

1. Brief. Turn the request into a brief: identity (style, mood in three words, tempo, metre, key, duration), timeline (sections with bars and seconds), orchestration (who plays what, in which register and articulation), energy curve (where it builds, breaks, peaks: 60 to 75 % of the length), mix, constraints. Show it in at most eight lines and go on; ask only when a structural choice is open and costly (a fixed duration for a video, room for a voice-over).
2. Score. Write tracks/<slug>.py: constants for harmony and themes, then the parts, then orch.render(...).
3. Render it with the command above.
4. Listen. python <skill>/listen.py tracks/<slug>.wav --bar <seconds> --sections "intro:1-8,theme:9-24,...", then open tracks/<slug>.png. Fix, at most three passes.
5. Deliver the WAV (mastered at -14 LUFS), the stems folder and the score. Say what you measured; never say it sounds great.

Writing a score

import sys; sys.path.insert(0, '<skill>')
from orchestra import Orchestra
from studio import midi
E = 60 / 110 / 3                       # seconds per eighth: 6/8 at dotted quarter = 110
def B(bar): return (bar - 1) * 6       # first eighth of a bar (6 eighths in 6/8, 8 in 4/4)
orch = Orchestra()
hn = orch.part('horns', 'long')
bs = orch.part('basses', 'short')
bd = orch.part('bass drum', 'hit')
hn.line(B(9), [('d4', 3), ('a4', 3), ('a4', 2), ('bb4', 1)], vel=110)   # (pitch, eighths); legato built in
hn.expr((B(9), 86), (B(10), 110), (B(11) - 0.5, 86))                     # CC1: the phrase swells and falls back
for i, v in enumerate([120, 62, 78, 114, 62, 78]):
    bs.note(B(9) + i, 0.6, 'd2', v)                                       # spiccato gallop, accents on the beats
bd.hit(B(9), 124)
sop = orch.part('choir', 'sopranos')
sop.words(B(9), [('do', 'a4', 3), ('mi', 'a4', 3), ('ne', 'd5', 6)], vel=100)   # (syllable, pitch, eighths)
sop.expr((B(9), 75), (B(10), 115), (B(11) - 0.5, 75))
orch.render('tracks/slug.wav', eighth=E)

• Time is in eighth notes; pitches in scientific notation (c4 is middle C) or MIDI numbers.
• Part.note(t, dur, pitch, vel), line(t, [(pitch, eighths)], vel, octave), hit(t, vel) for untuned percussion, words(t, [(syllable, pitch, eighths)], vel) for the choir, expr((t, 0-127), ...) dynamics of sustained notes, vol((t, dB), ...) fader automation.
• orch.part(instrument, articulation, name=None, pan=, gain=, send=, hpf=, eq=, fx=); levels are pre-calibrated per mode, add gain in dB from the stem report.
• orch.render(out, eighth, master_fx=[pedalboard plugins], glue=(threshold, ratio), Decay=seconds).

Instruments (python <skill>/orchestra.py lists them with their Discover availability):

The choir sings romaji syllables (do, mi, ne, ka, ra, a, o...), one per note, from words(). Write it the way film choirs are written: held chords on open vowels and Latin-like syllables, homophonic, behind the brass. A fast melody note by note on "ah" sounds like an organ. Ranges tested in tune: sopranos f4-d5, altos c#4-f4, tenors d3-a3, basses g2-d3; beyond them, check the result. A section of six singers takes about as long to render as it sings on a GPU, several times longer on a CPU.

What makes it sound like a film, not a mockup

Learned from listening tests, in order of effect:

1. The theme. A short motif with a strong interval (open fifth, a leap of a sixth), a question ending on the dominant, an answer that climbs to the high point. A stepwise line sounds like library music.
2. Octaves and open fifths. Double the theme in octaves across sections (trombones below, horns, trumpets, violins above); give the low strings and low brass roots and fifths. Full triads in the middle register sound like a video game.
3. Fewer, better parts. More instruments did not change the colour; consistent sources and writing did. In open mode the sound is the ceiling; in discover mode keep everything from Discover where it can (no extra synth layers).
4. A real floor. Basses, tuba, timpani and bass drum on the downbeats; in open mode a sub a little under the basses.
5. Dramatic contrast. Hold back the first statement, drop the floor before the climax, a stop on the dominant with a roll and a cymbal swell, the climax about 4 LUFS above the theme. A hard bus compressor flattens this: keep glue around (-14, 2).
6. Phrase by phrase. Every sustained phrase gets an expr swell; shorts get accented velocities; legato lines for melodies.

Reading listen.py

The figures check the render, never the taste.

Licences

The music made is the user's. Sonatina Symphonic Orchestra is CC Sampling Plus, VSCO 2 CE CC0, Virtual Playing Orchestra free for any use, sfizz BSD-2, Dragonfly and DawDreamer GPL-3, BigVGAN MIT. The choir's voices allow commercial use: Namine Ritsu (Canon) without conditions, Hoshino Hanami ~AI❤dol~ (LotteV) with credit, so tell the user to credit "Hoshino Hanami by LotteV" when they release music with a choir. BBC Symphony Orchestra Discover follows Spitfire's terms: music made with it can be released commercially, royalty free; its sounds and patches cannot be redistributed or used to train AI.