music-studio
Skill3 starsChanged 22 days agoApache-2.0View on GitHub
Create instrumental music and BGM (game loops, video scores, study/work background, personal tracks) locally with ACE-Step 1.5 on the user's NVIDIA GPU, with fully customisable defaults, style library, description lexicon and workflows plus a local dashboard. Use whenever the user wants to make, generate, loop, master or manage background music or a soundtrack, tweak music defaults or styles, fix bad-sounding results, or says 做音樂/配樂/BGM/循環音樂/遊戲音樂/讀書背景音. Also use for music-studio setup or troubleshooting.
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 2 are distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 17
All commands run from this skill's folder: python scripts/<tool>.py.
- Distinctive: line 21
Ask at most one question, and only when a wrong guess is expensive (exact video length, must it loop?).
- Distinctive: line 22
Do not narrate every command.
- Very specific: line 23
Never make the person run commands you can run.
- Very specific: line 27
not set up -> run scripts\setup_windows.ps1 (or tell them to double-click setup.bat).
- Very specific: line 35
Write the English caption yourself - it is the biggest quality lever: 6-12 concrete tags (genre, instrumental, lead instruments, rhythm, …
- Very specific: line 55
Do not guess. Read references/quality.md: open the take's effective / warnings (what really ran: model, steps, LM, quantization, ignored …
- Very specific: line 70
No exact cue points in time, no stems.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Music Studio
Generates instrumental music on the user's own GPU (ACE-Step 1.5), post-processes it (trim, exact length, fades, seamless loop, loudness, export) and keeps a library. Everything customisable lives in the user's workspace, never in this skill folder. Reply in the user's language; write captions in English.
Phase 1 scope: instrumental music/BGM only. Vocals, SFX, remixing existing audio are planned (references/roadmap.md) - say so plainly if asked.
All commands run from this skill's folder: python scripts/<tool>.py. Workspace: %USERPROFILE%\MusicStudio (or $MUSIC_STUDIO_HOME).
Conversation style: act, do not interview
The person usually wants music, not a questionnaire.
- Ask at most one question, and only when a wrong guess is expensive (exact video length, must it loop?). Otherwise decide, say in one line what you assumed, and go.
- Do not narrate every command. Report: the plan (one short block), a progress line per stage while generating, then the result and what to try next.
- Never make the person run commands you can run. If you cannot reach their machine, give exactly one thing to double-click (
setup.bat,start_dashboard.bat).
1. Make sure it can run (silent when fine)
python scripts/engine.py check -> read problems / notes. Fix them for the person:
- not set up -> run
scripts\setup_windows.ps1(or tell them to double-clicksetup.bat); no GPU ->-Mock. - models missing ->
python scripts/engine.py models --download(usemodelswith no flag to list what the tier needs;--download acestep-v15-sftfor the high-quality model). engine: mock-> the audio is a synthetic test signal; say so before they judge quality.
2. Show the dashboard (do this proactively)
python scripts/msclient.py url starts the server in the background if needed (no browser) and prints its URL; hand it over. The server is the one owner of the GPU: it keeps the models loaded between jobs (the first job of a session is the slow one) and queues jobs from both Claude and the dashboard, so they can watch your job live there, audition, edit the caption and finalize. If the tools run somewhere that is not their PC, tell them to double-click start_dashboard.bat instead.
3. Plan
Write the English caption yourself - it is the biggest quality lever: 6-12 concrete tags (genre, instrumental, lead instruments, rhythm, mood, space/production), see references/prompting-guide.md. Then:
python scripts/msconfig.py plan --style S --description "<their words>" --mood a,b --caption "tag, tag, ..." [--duration N --bpm N --key K --quality Q --workflow W --variants N --seed N]
--style from-descriptionwhen no preset fits; otherwise the nearest style (msconfig.py styles; feelings -> styles inreferences/use-cases.md).- Without
--caption, the caption is built from the style + the description (via the lexicon,msconfig.py lexicon) + moods; unknown Chinese words are reported inwarnings- translate them yourself. - Instruments: each style suggests fitting instruments and avoids unfitting ones (
msconfig.py instruments --style S). When the person names instruments, pass--instruments id,id(replaces the style's set) and/or--avoid-instruments id,id; instruments in the description always win. Defeat / war / sacrifice ->war-requiem(orchestral, no synths), notnature-piano. - Show caption, BPM, key, length, workflow in one block. Keep the spec warnings (clamped length, trimmed tags). Length: 30-120 s is the stable range; for long pieces generate ~2 min and loop.
4. Generate
Save the spec and run python scripts/msclient.py generate --spec spec.json (--engine mock for a dry run). It queues the job on the server and prints JSON lines: queued, phase (loading_models, encoding, diffusing, decoding, writing), warn, then result (take id, folder, variant files) or error (code, message, hint, log tail). Give the person one short line per phase, not every line.
- Relay warnings that change the outcome: no LM, model fell back, ignored settings, OOM retry,
cpu_decode(decoding fell back to the CPU and will take minutes; the next job is automatically tightened). - Other commands:
msclient.py status(queue + whether models are loaded),cancel JOB,wait JOB,unload(free the GPU, e.g. before they play a game). engine.py generate --spec ...runs a one-off job without the server; use it only if the server cannot start.
5. Audition -> refine -> finish
- Variants are
v1,v2... in the take folder / dashboard. Iterate by changing one thing and keeping the seed of the liked variant (dashboard: 以這版為基礎微調再生). Draft ondraft/standard; re-render the keeper onhighwith the same seed. - Every variant is checked automatically when it is written:
variants[].qcin the take (clipping, clicks, cut_off, silence, near_silent, low_volume, dull = muffled/no highs, each with a time). Mention any issue in one line with its time; it is a heuristic, the ear decides. - Finish:
python scripts/postprocess.py run --take ID [--variant N] [--workflow W] [--loudness video|game|background|-14] [--preview];--loudnessoverrides the workflow's target (video -14, game -16, background -18 LUFS). Explainreport.qcandreport.issuesin plain words. Loops: if the seam is flagged, offer the loop preview or regenerate with "steady, repetitive, no build". - Library:
python scripts/library.py list|show|annotate ID rating=4 favorite=true notes=...|trash ID.
When the person says it sounds bad
Do not guess. Read references/quality.md: open the take's effective / warnings (what really ran: model, steps, LM, quantization, ignored settings), fix the first cause on its list, regenerate with the same seed, and compare.
Customisation (the point of this skill)
Four layers in the workspace, merged over the shipped defaults:
- Parameters/defaults:
config.jsonviapython scripts/msconfig.py set KEY VALUE/unset KEY(show --explaintells which layer a value comes from; keys inreferences/config-reference.md). "What can I tune?" ->msconfig.py knobs [--quality Q](every generation setting: value, layer, range, effect; same table as dashboard 設定). A one-off tweak for one take:plan ... --param shift=2.5. - Styles / moods:
styles/*.json,moods.json(references/styles-moods-workflows.md). - Lexicon:
lexicon.json- teach it how your words map to English tags (msconfig.py lexiconshows the built-ins). - Workflows:
workflows/*.json- stages, gates, post-process steps. "以後預設用 X" = a change in the workspace (neverdefaults/), thenmsconfig.py validate, then say exactly what changed and how to revert. A one-off wish applies to this request only.
Engine parameter precedence: generation < hardware tier < quality preset < engine_overrides < the request. VRAM tier is auto-detected; force it with hardware.tier.
Be honest about
- First real ACE-Step runs on a given machine may need model-folder-name or setting fixes;
engine.py checkand the take's diagnostics panel show what is wrong. ACE-Step quality is bounded by the model: turbo is fast, sft is better but slower. - Loops are crossfade-based and need steady tempo; requested BPM is a hint. QC seam numbers are heuristics - the loop preview is the real test.
- No exact cue points in time, no stems. Do not name real artists/songs/games in captions; check the model licence before commercial use.
Files
scripts/ tools | defaults/ shipped config, styles, moods, lexicon, workflows | docs/ARCHITECTURE.md (how server, queue and worker fit together) | references/ (read as needed): quality.md, prompting-guide.md, use-cases.md, config-reference.md, styles-moods-workflows.md, troubleshooting.md, dashboard.md, roadmap.md | setup.bat, start_dashboard.bat | tests/ | evals/
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

