bitwize-music-studio/claude-ai-music-skills

I love music but never learned an instrument. AI became the creative outlet that was always out of reach. This project started as a way to go deep on Claude Code plugin architecture, agentic workflows, multi-model orchestration, and MCP tooling. Music was the domain because it was personal.

What it actually does: a Claude Code plugin that turns a conversation into a full album production pipeline. You describe what you want to make, and it handles concept development, lyrics, Suno prompts (an AI music generation platform), audio mastering, and release prep — with quality gates and source verification at every stage.

Questions? See the FAQ.

You:    "Let's make an album about the 2016 Bangladesh Bank heist"
Claude: Creates album structure, runs 7-phase concept planning

You:    "Start the research"
Claude: Dispatches legal, financial, and security researchers in parallel
        Gathers DOJ filings, SWIFT documentation, malware analysis
        Cross-verifies sources, flags claims that need human review

You:    "Sources look good. Let's write track 1"
Claude: Drafts lyrics, checks prosody and rhyme schemes
        Scans for pronunciation risks, suggests phonetic fixes
        Builds the Suno style prompt and generation settings (model, Variety, Max Mode)

You:    "Track sounds great, here are the stems"
Claude: Imports stems from Suno, polishes per-stem
        Masters to -14 LUFS for streaming
        Generates promo video and social media copy

Concept to released album. You generate on Suno, everything else happens in the terminal.

/plugin marketplace add bitwize-music-studio/claude-ai-music-skills
/plugin install bitwize-music@bitwize-music

Then run /bitwize-music:setup to detect your environment and install dependencies. Run /bitwize-music:configure to set your artist name and workspace paths.

Platform: macOS, Linux, WSL2, and native Windows are all fully supported. The full test suite runs on windows-latest in CI (plus dedicated Windows legs for the MCP boot check and MuseScore PDF export) — the MCP server, state cache, non-audio workflow, and the ffmpeg audio pipeline all run natively there. Promo video works on Windows too, though it's verified by hand rather than continuously guarded (its tests mock ffmpeg). Sheet music works natively too: MuseScore PDF export is CI-verified on windows-latest, and AnthemScore transcription runs against a licensed install (its free trial exposes no CLI on any OS, so that caveat isn't Windows-specific). Python 3.11+ for the MCP server and audio tools. See the compatibility matrix for the per-feature breakdown.

This is where the engineering lives. The plugin is a case study in how far you can push Claude Code's plugin system.

Each skill is a self-contained markdown file with a YAML frontmatter that declares its model, description, and when it should activate. Skills range from simple clipboard operations to multi-step creative workflows. Claude routes to skills automatically based on context, or you invoke them directly with /bitwize-music:<name>.

The lyric-writer knows prosody rules, rhyme scheme analysis, and Suno's pronunciation quirks. The mastering-engineer knows loudness targets per platform and genre-specific EQ curves. The researcher coordinates parallel sub-agents across 10 domain specializations.

See docs/skills.md for the full reference.

Skills declare which Claude model they need. Creative work that directly impacts music quality runs on Opus. Coordination and reasoning tasks use Sonnet. Mechanical operations (imports, validation, clipboard) run on Haiku.

Skills declare the tier alias, not a pinned version, so each one tracks the current frontier model in its tier automatically — no per-skill edit when a new Claude generation ships.

This project pushes Claude Code hard — multi-agent research, real-time audio analysis, sub-agent orchestration across model tiers. It works best on the Max subscription. The standard Pro subscription will hit rate limits during multi-track sessions.

See reference/model-strategy.md for per-skill rationale.

A Python MCP server exposes 80+ tools for instant state queries, audio analysis, lyrics processing, and database operations. The server is the plugin's nervous system — skills call MCP tools instead of reading files directly, which keeps responses fast and state consistent.

Key tool categories:

• State management — album/track lookups, session context, cache rebuild
• Lyrics analysis — syllable counting, readability scoring, rhyme detection, section validation, cross-track repetition
• Audio processing — mastering, stem analysis, QC checks, promo video generation
• Database — tweet/promo content management via PostgreSQL

For documentary and true-story albums, the research system coordinates parallel investigation across 10 domain-specific sub-agents. A lead researcher dispatches to specialists (legal, financial, security, government, journalism, etc.), each trained on where to find primary sources in their domain. A verification agent cross-checks all claims before human review.

The full pipeline: gather sources, verify citations, require human sign-off, then — and only then — allow lyrics generation. Every claim in the music traces back to a captured, verified source.

Nothing ships without passing gates:

• Lyrics: 13-point checklist (rhyme, prosody, pronunciation, POV consistency, factual accuracy)
• Pre-generation: Sources verified, explicit flags set, style prompt complete, artist names cleared
• Audio: 7-point QC (loudness, clipping, silence, phase, stereo width, frequency balance, dynamic range)
• Structure: Album directory validation, file location checks, content integrity

72 genre directories with production guides, mastering presets, artist deep-dives, and Suno-specific tips. From afrobeats to vaporwave, each genre includes subgenre breakdowns, lyric conventions, and reference artists.

6 GitHub Actions workflows: tests (4,412 across ubuntu/macOS/Windows, plus lint, security scanning with bandit + pip-audit, and static validation), real-service integration (Postgres, SeaweedFS/S3, MuseScore), nightly deep tests, auto-release from changelog, PR target enforcement, and version sync. Dependabot watches pip and Actions versions weekly.

Coverage is measured on all three OSes and gated on the combined total, not one platform's view. That matters because every sys.platform == "win32" branch is unreachable on Linux — measuring only there made the platform-specific code invisible to the gate, which is exactly where this project's real bugs have lived. The merge is asserted rather than assumed: a mis-specified path mapping makes coverage combine report the Linux-only number while looking like success, so the job fails unless a known win32-only line is genuinely covered.

skills/              53 skill definitions (markdown + YAML frontmatter)
servers/             MCP server (Python, 80+ tools)
tools/               Audio mastering, promo videos, sheet music, cloud uploads
reference/           46+ docs — Suno guides, mastering workflows, genre references
genres/              72 genre directories with production guides
templates/           Album, track, artist, research templates
tests/               4,412 tests across 14 categories
config/              Example config and setup docs

If you make something with this, I'd genuinely love to hear it — @bitwizemusic on X, join the Discord, or open a discussion.

CC0 — Public Domain. Do whatever you want with it.

Artist and song references in the genre documentation are for educational and reference purposes only. This plugin does not encourage creating infringing content. Users are responsible for ensuring their generated content complies with applicable laws and platform terms of service.