Claude music

claude-music — AI Music Production for Claude Code

Quick Reference

Command	What it does
/music	Interactive mode — describe what you want
/music generate	Text/lyrics to full song (text2music)
/music cover	Style transfer from reference audio
/music repaint	Edit a specific section of a song
/music extract	Separate tracks/stems (base model only)
/music lego	Add instrument layer (base model only)
/music complete	Continue/extend audio (base model only)
/music compose	Songwriting: craft caption + lyrics + params
/music analyze	BPM, key, loudness, duration analysis
/music export	Platform-optimized export (Spotify, YouTube, etc.)
/music enhance	Post-processing: normalize, denoise, stem separate
/music random	Quick random generation with smart defaults
/music library	Browse and manage generated music
/music web	Local browser dashboard: generate, play, rate
/music lora	LoRA/LoKr fine-tuning management
/music setup	Verify installation and dependencies

Orchestration Logic

Command Routing

When the user provides a specific command, load the matching sub-skill:

• /music generate or intent is create song/make music/text-to-music/lyrics-to-music → Read skills/claude-music-generate/SKILL.md
• /music cover or intent is cover/style transfer/remake/version of → Read skills/claude-music-cover/SKILL.md
• /music repaint or intent is edit section/fix chorus/change part/modify section → Read skills/claude-music-repaint/SKILL.md
• /music compose or intent is write lyrics/craft caption/plan song/songwriting → Read skills/claude-music-compose/SKILL.md
• /music analyze or intent is BPM/key detection/loudness/audio info → Read skills/claude-music-analyze/SKILL.md
• /music export or intent is export for Spotify/YouTube/platform/format conversion → Read skills/claude-music-export/SKILL.md
• /music enhance or intent is normalize/denoise/stem separate/master → Read skills/claude-music-enhance/SKILL.md
• /music random or intent is quick generation/surprise me/random song → Read skills/claude-music-random/SKILL.md
• /music library or intent is list songs/browse output/manage music → Read skills/claude-music-library/SKILL.md
• /music web or intent is dashboard/browser app/web UI/visual player → Read skills/claude-music-web/SKILL.md
• /music lora or intent is train/fine-tune/LoRA/custom style → Read skills/claude-music-lora/SKILL.md
• /music setup → Run bash ~/.claude/skills/claude-music/scripts/setup.sh

Interactive Mode

When user says /music without arguments or describes a task in natural language:

1. Run bash ~/.claude/skills/claude-music/scripts/check_deps.sh to verify tools
2. Run bash ~/.claude/skills/claude-music/scripts/detect_gpu.sh for GPU info
3. Identify intent from the user's description
4. Route to the appropriate sub-skill
5. If ambiguous, ask the user to clarify

Multi-Step Pipelines

For complex requests spanning multiple sub-skills (e.g., "compose lyrics, generate a song, then export for Spotify"):

1. Compose lyrics/caption with /music compose
2. Generate with /music generate using composed output
3. Export with /music export
4. Clean up temp files

Generate-Listen-Iterate Loop

After any generation:

1. Present output file paths and metadata (seed, duration, format)
2. Suggest playback: ffplay -nodisp -autoexit "<path>"
3. Ask if user wants to:
   • Re-generate with different seed (same params)
   • Refine params (adjust caption, BPM, quality)
   • Repaint a specific section
   • Cover to change style while keeping structure
   • Export for a platform

Safety Rules — MANDATORY

1. Run preflight before writes: bash ~/.claude/skills/claude-music/scripts/preflight.sh "$INPUT" "$OUTPUT"
2. Never overwrite source files — all operations produce new files
3. Check VRAM before GPU operations: bash ~/.claude/skills/claude-music/scripts/detect_gpu.sh
4. Confirm before: batch >4 generations, operations with --quality max (3-5 min)
5. Auto-execute without confirmation: single generation (draft/standard), analysis, format conversion, setup
6. Temp files: /tmp/claude-music/ with cleanup trap
7. Output directory: ~/Music/claude-music-output/ (auto-created)

ACE-Step Configuration

• Installation: Set ace_step_dir in config.json (default: see config.json)
• Invocation: bash ~/.claude/skills/claude-music/scripts/music_engine.sh <command> [args]
• Config: ~/.claude/skills/claude-music/config.json
• Output: ~/Music/claude-music-output/

Quality Presets

Preset	Model	LM	Steps	Speed	Use for
draft	turbo	none	8	~15s	Quick exploration, batch 4 variants
standard	turbo	none	8	~15s	Default, batch 2 variants
high	turbo	1.7B LM	8	~25s	Better lyrics/structure, thinking mode
max	base	1.7B LM	65	~3-5min	Highest quality, single output

VRAM Management (RTX 5070 Ti — 16GB)

Configuration	VRAM	Offload	Notes
Turbo (no LM)	~8GB	CPU offload	Default, fast generation
Turbo + 0.6B LM	~10GB	CPU + DiT offload	Thinking mode, lightweight
Turbo + 1.7B LM	~14GB	CPU + DiT offload	Full thinking, tight on VRAM
XL Turbo	~14-16GB	Full offload	Maximum quality DiT, no LM room

Rule: Never run two heavy models simultaneously. The music_engine.py handles VRAM automatically.

Script Invocation

All ACE-Step operations go through the bash wrapper:

bash ~/.claude/skills/claude-music/scripts/music_engine.sh <subcommand> [args]

The wrapper handles: path setup, environment variables, VRAM pre-check, uv run invocation.

Output is always JSON to stdout. Parse with jq for specific fields.

Reference Files (Load On-Demand)

Reference	When to load
references/prompt-guide.md	When crafting captions or lyrics
references/parameters.md	When user asks about specific params or tuning
references/genre-recipes.md	When targeting a specific genre
references/music-theory.md	When discussing keys, scales, BPM, song structure
references/post-processing.md	When exporting, mastering, or enhancing
references/song-structures.md	When planning song layout
references/lora-training.md	When training custom LoRA models

Sub-Skills

Skill	Type	Description
claude-music-generate	Generation	Core text2music via ACE-Step Python API
claude-music-cover	Generation	Style transfer from reference audio
claude-music-repaint	Editing	Selective section regeneration
claude-music-compose	Reference	Songwriting guide (caption, lyrics, params)
claude-music-analyze	Analysis	BPM, key, loudness, metadata via ffprobe/librosa
claude-music-export	Processing	Platform-specific audio export via FFmpeg
claude-music-enhance	Processing	Loudness normalization, denoise, stems (reuses video skills)
claude-music-random	Generation	Quick random generation with genre presets
claude-music-library	Management	Browse/search generated music output
claude-music-web	UI	Local browser dashboard (generate, play, rate)
claude-music-lora	Training	LoRA/LoKr fine-tuning wrapper

Scripts

Script	Purpose
scripts/music_engine.py	Core ACE-Step Python API wrapper (all 6 task types)
scripts/music_engine.sh	Bash wrapper (env, VRAM, uv run, show-once star nudge)
scripts/music_export.sh	FFmpeg platform export commands
scripts/rank.py	Batch-rank outputs vs caption (stub — Theme 3 of research plan)
scripts/detect_gpu.sh	GPU detection + tier recommendation → JSON
scripts/preflight.sh	Safety checks for audio files → JSON
scripts/check_deps.sh	Dependency verification → JSON
scripts/setup.sh	Installation verification (invoked by /music setup)

Audio analysis (BPM, key, loudness) is handled by claude-music-analyze via ffprobe + FFmpeg's loudnorm filter — no separate Python script; see that sub-skill.