music-skill 0.3.0

Music that keeps changing and moves, guidance from professional sources only, and the full Surge XT patch set.

• Music that develops. Seven pieces that agents made with 0.2.0 were heard as not wrong, but at times dull or stiff. Their scores showed why: loops held for 24 to 48 bars, choruses thinner than their pre-choruses, chord parts struck as blocks. development-and-groove.md plans energy and variation section by section, review asks about each of those causes, and musiclib.voicing.voice_lead voices a progression with its common tones held.
• review follows the voice leading. It names notes outside the written chord as embellishing tones and asks about those not resolved by step (Open Music Theory). It lists parallel fifths and octaves between independent lines, and close intervals under Hoffmann's low interval limits.
• Guidance from professional sources only. Every guide was checked line by line: each statement names a source, or is marked as the skill's own measurement, an example plan or a starting value. New: harmony.md from Open Music Theory, sound-selection.md, a method for writing a melody, and instruments for a scene's character.
• Surge XT for synth parts. All 3,008 patches of Surge XT 1.3.4 are measured in surge-catalog.json; studio.py sounds searches them and studio.py audition plays candidates on a part. Surge parts load in about half a second instead of minutes, and Surge XT now installs by itself on Linux (x86-64) and macOS, as on Windows.
• New examples, each made by Claude (Opus 5.5, xhigh effort) in one run from a one-line subject: Fughetta on a Lantern Theme, Platform in the Snowmelt, Sodium Lights and Sprinkle Satellite. Their recordings are on examples-2, and fetch_soundbanks.py --only examples installs the sound banks they use. The earlier examples' recordings stay on examples-1.

Two changes break earlier use: score_review.json's foreground_outside_chord is replaced by outside_chord, which covers every pitched part; and bootstrap.py --system no longer selects a system install of Surge XT on Windows (--surge-vst3 <bundle> does, on every system).

The changelog has the details.

music-skill 0.2.0

More styles, each with a guide written from its sources, a nylon-string guitar, and a way to see what a revision changed.

• Guides for more styles, each with its sources listed at its end: film score (underscore, scoring to picture, suspense and horror), pop and ballad, rock and metal, and baroque, plus boom bap and trap, techno, blues and bossa nova in the existing guides.

• A nylon-string classical guitar, iowa_guitar, cut from the University of Iowa's 2011 recordings: the thirtieth bank. fetch_soundbanks.py --only iowa_guitar installs it from soundbanks-1.

• studio.py compare shows what a revision changed between two renders:

  • the project's edits and the parts synthesized again;
  • for every stem, bus, reverb and delay return, the premix and the master: identical, or where it differs, in bars.

  --clip makes an old-then-new clip at equal loudness, and --differences writes each change on its own. Before and after shows it on a one-note revision.

• Reverb and delay returns as files on the bus chain. With the buses they add up to the premix, so a mix can be rebuilt in another program.

The example recordings on examples-1 are unchanged. The changelog has the details and the known issues.

Example recordings 2

The example recordings of music-skill 0.3.0: MP3 files at 256 kbps and 48 kHz, tagged. Claude (Opus 5.5, xhigh effort) made each piece with the skill's version 0.3.0, in one run from a one-line subject and the same rules for every piece, and nobody edited the result. Each example folder in the repository holds the prompt, the agent's script that writes the piece and the project file it renders. Like the compositions, they are dedicated to the public domain under CC0 1.0. The recordings of the earlier examples stay in examples-1.

File	Piece	Length	Size	SHA-256
fughetta-on-a-lantern-theme.mp3	Fughetta on a Lantern Theme, a baroque miniature for harpsichord and strings	1:55	3.7 MB	f0c07d66ffd1de3882e95c73f506423f3dd0b6d99bb2e434baa565ee9c4f7b4f
platform-in-the-snowmelt.mp3	Platform in the Snowmelt, a Japanese ballad for piano and strings	4:37	8.9 MB	a87458ee0e2f08a762e9214404bdf1cb28a40dac1f1898b00c091fd82f0c1489
sodium-lights.mp3	Sodium Lights, Japanese city pop	4:28	8.6 MB	567e58a05d80e4fbe30a6dd04c7ccad1ef7607322a56c93932f4a67c6ba84f2d
sprinkle-satellite.mp3	Sprinkle Satellite, kawaii future bass	2:30	4.8 MB	d0bea89800d1b33797844e971d57d30ff7b5383e052f419a0a1a9ba565fea54a

music-skill 0.1.1

Packaging for the Claude plugin directory, and a fix for the chorus effect. The sound banks are unchanged.

• The chorus effect no longer crackles. FFmpeg's chorus, which it ran on, stepped its moving delay at whole samples and read a voice 61.5 ms late whenever that voice's delay came down to zero. That was heard as a buzz with a catch every few seconds under sustained pads. It now runs in NumPy with the same voices, read between samples. Midnight Coastline's string pads use it, so its recording was rendered again.
• The example recordings moved to a release. They are now assets of the release examples-1 instead of repository files, and the READMEs link there. No file in the repository is 5 MiB or larger.
• What the skill runs, fetches and sends. The README now says what the skill runs, every host its setup downloads from, and that it sends nothing.
• CI. Each pull request is tested once.

Listen:

• Above the Clouds, a concert overture for orchestra and organ (MP3).
• Signal Fire, festival house (MP3).
• Five more styles in the README.

The changelog has the details and the known issues.

music-skill 0.1.0

First public release.

music-skill lets an AI agent write original instrumental music as notes, then plays them on sampled instruments, mixes and masters the result. No audio-generation model is involved: you get an MP3 and a WAV, plus the MIDI, a stem per part and a project file where any note can be changed and rendered again.

Listen: Above the Clouds, a concert overture for orchestra and organ (listen); Signal Fire, festival house (listen); and five more styles in the README. Claude (Opus 5.5) composed, arranged and mixed each piece with the skill.

What's in it:

• A project format for score, performance and mix, with MIDI export and a score review that flags clashing voices and melody notes outside the chords.
• A performance model: per-player timing and dynamics, swing, rubato, staggered chords, legato and bends, dynamics that move through a bank's recorded layers, and drum parts that step through their kit's round robins.
• Rendering with FluidSynth on twenty-nine optional sound banks fetched or built from their publishers' samples (pianos, guitars, basses, drum kits, a full orchestra, saxophones, organ and world instruments; all CC0, public domain or free for any use), built-in synthesizers and drum machines, and Surge XT with fifty curated presets, optional and installed by default.
• A bus mix with generated rooms, plates, halls and churches and production tools for electronic styles, mastered to −14 LUFS (louder for club music, or with full dynamics for orchestra) with true-peak and MP3 checks, and analyze for levels, spectrum, stereo, how far the melody sits above the band, and the master's tonal balance against published measurements and genre references.
• Guides for electronic music, beats, jazz and funk, anime openings, game and scene music, orchestra, organ, and piano and chamber music; game loops, also in layers.
• Setup on Windows (pinned portable tools), Linux and macOS (system packages), every download checked by SHA-256.

Install in Claude Code with /plugin marketplace add owoevan-dev/music-skill, or anywhere with npx skills add owoevan-dev/music-skill. The README covers the renderer setup.

Known issues are listed in the changelog. The agent can't listen: its measurements point at problems, and a person still has to hear the result.

Example recordings 1

The example recordings of music-skill: MP3 files at 256 kbps and 48 kHz, tagged. Claude (Opus 5.5) composed, arranged and mixed each piece with the skill; each example folder in the repository holds the script that writes its piece and the project file it renders. They are the recordings the repository held up to v0.1.0, byte for byte, except Midnight Coastline, rendered again for 0.1.1: its string pads use the chorus effect, which no longer crackles. Like the compositions, they are dedicated to the public domain under CC0 1.0.

File	Piece	Length	Size	SHA-256
above_the_clouds.mp3	Above the Clouds, concert overture for orchestra and organ	4:28	8.6 MB	609c42a4f8f4d9e6d635fb597d3402f89340c6128c23364d9a7b1de935b26e24
signal_fire.mp3	Signal Fire, festival house	1:51	3.6 MB	6efbedb9610f52646599669d31b26fd93e3772bf15b0be3c5482b54cf59bac8f
late_set.mp3	Late Set at the Corner, jazz blues	1:40	3.2 MB	73d1828f7592cbfeee05c1c599e8975d6f989816c8571cef96b728c619fc4e45
midnight_coastline.mp3	Midnight Coastline, synthwave	2:01	3.9 MB	78d18bf0a75ab55d14114b3f9a2ca33d3caae2eca8256eebed5bf213a20a0efa
tram_line_rain.mp3	Rain on the Tram Line, lo-fi hip-hop	1:43	3.3 MB	44bcced6ec6b0f82a199455180928c134fb2b10b83ea6ab0a74ece1c727cf389
paper_satellites.mp3	Paper Satellites, future bass	1:09	2.2 MB	00b4c4e27239356aa506ac097412991deb79adc3e94bb167610037a0e3da5d88
pixel_lantern.mp3	Pixel Lantern, chiptune loop	0:53	1.7 MB	f3e2de96db1fe44bc3652dadea9c62970bf83d9b8d56b9237c7b6cdb0cbb61f2
daybreak_signal_piano.mp3	Daybreak Signal (piano), anime-opening-style, TV size	1:33	3.0 MB	aa7ad9f54ba54d64ad1908ab2af4003140c47cdbf3640ea3b4a9b1233fb688ec

Sound banks 1

Sound banks for music-skill, as the project builds them.

Each asset is one SF2 bank compressed with xz. fetch_soundbanks.py downloads the ones a setup asks for, checks each against skills/music-production/references/soundbank-lock.json by size and SHA-256, unpacks it and checks the bank against the hash a build from the publishers' own samples gives, so every bank here is byte for byte what --from-source builds with NumPy 2.1.2, the version the skill pins for Python 3.10–3.13. Nothing needs to be downloaded by hand.

25 banks, 2.9 GB to download, 7.0 GB unpacked. One was added on 6 November: iowa_guitar (b2), a nylon-string classical guitar from the Iowa recordings. Seven were rebuilt on 3 November (tuba_war, vsco_colors, vcsl_winds and the four Iowa banks): a dynamic borrowed from a louder recording now plays as soft as its measured level (FluidSynth applies zone attenuation at 0.4 of its value), and the loop search uses each recording's own pitch. Three were rebuilt on 4 November: piano_osiris (b3), where many pedal takes had recorded a short knock from the piano's action while the note rang on, which the build now repairs; vcsl_keys (b2) and vsco_colors (b3), whose Kawai grand and Yamaha, Knight and VSCO uprights now have every dynamic of a note at one level, so their soft playing no longer sits 25-44 dB under their loud playing and the velocity curve sets loudness, as on the other pianos. Earlier builds stay attached for reference; the lock names the current ones.

Asset	Bank	Licence	Download	Unpacked
bass_darkblack-b1.sf2.xz	bass_darkblack	CC0-1.0 source samples: Black And Blue Basses by Karoryfer Samples; built by music-production soundbank.py	109 MB	355 MB
bass_sneaky-b1.sf2.xz	bass_sneaky	CC0-1.0 source samples (sfzinstruments/karoryfer.sneakybass); built by music-production soundbank.py	81 MB	253 MB
cello_bigcat-b2.sf2.xz	cello_bigcat	CC0-1.0 source samples: Karoryfer x bigcat cello by Karoryfer Lecolds and bigcat instruments; built by music-production soundbank.py	57 MB	82 MB
erhu-b1.sf2.xz	erhu	CC0-1.0 source samples (sfzinstruments/aliexpress-erhu); built by music-production soundbank.py	35 MB	48 MB
guitar_shiny-b1.sf2.xz	guitar_shiny	CC0-1.0 source samples (sfzinstruments/karoryfer.shinyguitar); built by music-production soundbank.py	90 MB	273 MB
iowa_brass-b2.sf2.xz	iowa_brass	University of Iowa Electronic Music Studios, Musical Instrument Samples: "Since 1997, these recordings have been freely available on this website and may be downloaded and used for any projects, without restrictions." (https://theremin.music.uiowa.edu/MIS.html, retrieved 2026-11-02)	21 MB	42 MB
iowa_guitar-b2.sf2.xz	iowa_guitar	University of Iowa Electronic Music Studios, Musical Instrument Samples: "Since 1997, these recordings have been freely available on this website and may be downloaded and used for any projects, without restrictions." (https://theremin.music.uiowa.edu/MIS.html, retrieved 2026-11-02)	54 MB	149 MB
iowa_perc-b3.sf2.xz	iowa_perc	University of Iowa Electronic Music Studios, Musical Instrument Samples: "Since 1997, these recordings have been freely available on this website and may be downloaded and used for any projects, without restrictions." (https://theremin.music.uiowa.edu/MIS.html, retrieved 2026-11-02)	132 MB	465 MB
iowa_strings-b4.sf2.xz	iowa_strings	University of Iowa Electronic Music Studios, Musical Instrument Samples: "Since 1997, these recordings have been freely available on this website and may be downloaded and used for any projects, without restrictions." (https://theremin.music.uiowa.edu/MIS.html, retrieved 2026-11-02)	101 MB	172 MB
iowa_winds-b4.sf2.xz	iowa_winds	University of Iowa Electronic Music Studios, Musical Instrument Samples: "Since 1997, these recordings have been freely available on this website and may be downloaded and used for any projects, without restrictions." (https://theremin.music.uiowa.edu/MIS.html, retrieved 2026-11-02)	47 MB	73 MB
organ_vcsl-b2.sf2.xz	organ_vcsl	CC0-1.0 source samples: VCSL by Versilian Studios, pipe organ sampled by Simon Dalzell (Ivy Audio); built by music-production soundbank.py	57 MB	83 MB
piano_osiris-b3.sf2.xz	piano_osiris	CC0-1.0 source samples: Osiris Piano by Karoryfer Samples; built by music-production soundbank.py	101 MB	398 MB
rusty-b1.sf2.xz	rusty	CC0-1.0 source samples: Big Rusty Drums by Karoryfer Samples; built by music-production soundbank.py	133 MB	513 MB
sax_bear-b1.sf2.xz	sax_bear	CC0-1.0 source samples (sfzinstruments/karoryfer.bear-sax); built by music-production soundbank.py	36 MB	58 MB
steel_drum-b1.sf2.xz	steel_drum	Unlicense (public-domain dedication) source samples (sfzinstruments/jlearman.SteelDrum); built by music-production soundbank.py	47 MB	117 MB
tuba_war-b2.sf2.xz	tuba_war	CC0-1.0 source samples (sfzinstruments/karoryfer.war-tuba); built by music-production soundbank.py	25 MB	80 MB
vcsl-b4.sf2.xz	vcsl	CC0-1.0 source samples: VCSL by Versilian Studios; built by music-production soundbank.py	233 MB	679 MB
vcsl_keys-b2.sf2.xz	vcsl_keys	CC0-1.0 source samples: VCSL by Versilian Studios; built by music-production soundbank.py	263 MB	799 MB
vcsl_winds-b2.sf2.xz	vcsl_winds	CC0-1.0 source samples: VCSL by Versilian Studios; built by music-production soundbank.py	236 MB	363 MB
vcsl_world-b1.sf2.xz	vcsl_world	CC0-1.0 source samples: VCSL by Versilian Studios; built by music-production soundbank.py	115 MB	296 MB
vocal_ah-b1.sf2.xz	vocal_ah	CC0-1.0 source samples (sfzinstruments/legato_vocal_tutorial); built by music-production soundbank.py	17 MB	24 MB
vsco-b5.sf2.xz	vsco	CC0-1.0 source samples: VSCO 2 Community Edition by Versilian Studios; built by music-production soundbank.py	339 MB	608 MB
vsco_colors-b3.sf2.xz	vsco_colors	CC0-1.0 source samples: VSCO 2 Community Edition by Versilian Studios (the library also asks that the samples are not sold as such and encourages credit and a link to its homepage); built by music-production soundbank.py	178 MB	374 MB
vsco_extra-b2.sf2.xz	vsco_extra	CC0-1.0 source samples: VSCO 2 Community Edition by Versilian Studios (the library also asks that the samples are not sold as such and encourages credit and a link to its homepage); built by music-production soundbank.py	157 MB	318 MB
vsco_strings-b1.sf2.xz	vsco_strings	CC0-1.0 source samples: VSCO 2 Community Edition by Versilian Studios (the library also asks that the samples are not sold as such and encourages credit and a link to its homepage); built by music-production soundbank.py	200 MB	336 MB

The five banks FreePats publishes as SF2 are attached too, their bank, licence and readme files repacked unchanged from FreePats' .7z releases as tar.xz (0.6 GB), so every bank installs the same way; fetch_soundbanks.py checks each bank against the hash of FreePats' own file and falls back to FreePats' download:

Asset	Bank	Licence	Download
EGuitarFSBS-dist1-SF2-20220911.tar.xz	fsbs_dist1	CC0-1.0: FreePats Electric Guitar FSBS (dist1)	351.5 MB
EGuitarFSBS-dist2-SF2-20220911.tar.xz	fsbs_dist2	CC0-1.0: FreePats Electric Guitar FSBS (dist2)	131.3 MB
EGuitarFSBS-clean-SF2-20260807.tar.xz	fsbs_clean	CC0-1.0: FreePats Electric Guitar FSBS (clean)	148.5 MB
PickedBassYR-SF2-20190930.tar.xz	bass_pick	CC0-1.0: Picked Bass YR, samples by Andrea Biasior; FreePats	2.4 MB
FingerBassYR-SF2-20190930.tar.xz	bass_finger	CC0-1.0: Finger Bass YR, samples by Andrea Biasior; FreePats	2.6 MB

Credits

Music rendered with these banks needs no credit. The sample libraries they are built from, with thanks:

• VSCO 2 Community Edition and VCSL by Versilian Studios: VSCO recorded by Sam Gossner and Simon Dalzell (Ivy Audio), the VCSL pipe organ sampled by Simon Dalzell. CC0 1.0. The VSCO readme asks that the samples are not sold directly, that improvements stay open, and for credit with a link to the VSCO: CE homepage.
• Karoryfer Samples: Big Rusty Drums, Black And Blue Basses, Osiris Piano, Sneakybass, Shinyguitar, War Tuba, Bear Sax and the Hadzi-Fia "ah" from the legato vocal tutorial; the Karoryfer x bigcat cello with bigcat instruments, played by Kamila Borowiak. CC0 1.0.
• The aliexpress-erhu samples of the sfzinstruments project. CC0 1.0.
• jSteelDrum by John Learman. Unlicense.
• The University of Iowa Musical Instrument Samples, University of Iowa Electronic Music Studios: "Since 1997, these recordings have been freely available on this website and may be downloaded and used for any projects, without restrictions."