MiniMax Music 3 — Agent Skills
This directory contains agent skills that extend AI coding agents (Claude Code, Cursor, Codex, etc.) with MiniMax Music 3 capabilities. Skills are plain-text, self-contained instruction packs — no external APIs, no runtime dependencies, no model weights required.
Available Skills
| Skill | What it does | Invocation |
|---|---|---|
music-caption-rewriter |
Turns a brief music description and optional tagged lyrics into a detailed Music 3.0 Structured Caption (Global Metadata, Vocal Details, Arrangement) for precise generation control. | $music-caption-rewriter |
Installation
Option 1: Install with the skills CLI (recommended)
From the repository root:
npx skills add MiniMax-AI/MiniMax-Music3 --skill music-caption-rewriterThis downloads the skill into your agent's skills directory and makes it available as $music-caption-rewriter.
Option 2: Manual install
Copy the skill folder into your agent's skills directory:
# Claude Code (user-level)
cp -r skills/music-caption-rewriter ~/.claude/skills/
# Cursor (project-level)
cp -r skills/music-caption-rewriter .cursor/skills/
# Codex (user-level)
cp -r skills/music-caption-rewriter ~/.codex/skills/The directory name must stay music-caption-rewriter — agents locate skills by folder name.
Basic Usage
Invoke $music-caption-rewriter in a supported coding agent with:
- Caption (required) — a natural-language music description, e.g. "A warm acoustic pop song with intimate female vocals, fingerpicked guitar, and a gradual emotional build into a wide final chorus."
- Lyrics (optional) — lyric text containing bracketed section tags such as
[Verse],[Chorus],[Bridge],[Instrumental]. - Constraints (optional) — desired length, output format, exclusions, or creative direction.
Example prompt:
Use
$music-caption-rewriterto turn my music description and optional tagged lyrics into a concise, section-aware Music 3.0 caption.
Output
The skill returns a structured caption with exactly three sections:
- Global Metadata — genre, subgenre, tempo, emotional progression, and production profile.
- Vocal Details — lead configuration, timbre, register, delivery, harmony/backing vocals, and restrained vocal effects.
- Arrangement — a section-by-section timeline with instrument lifecycles, groove development, and transitions.
JSON or JSONL output is available on request.
The rewritten caption is designed to be passed directly to MiniMax Music 3 — as the instructions field of the generation API — alongside the original lyrics in the input field. See the top-level README for the full workflow.
How It Works
The skill is built on a fully static, text-based library with progressive disclosure — the agent routes to the right style family first and only opens the templates it actually needs:
SKILL.md Skill instructions and workflow
├── agents/openai.yaml Agent metadata (display name, default prompt)
├── references/
│ ├── genre-router.md Entry point: maps genres to style families
│ └── index-*.md 18 family indexes with compact style cards
└── templates/ 1,000 full caption templatesThe genre router maps your genre/mood cues to one of 18 style families (pop & ballad, hip-hop & rap, metal & heavy rock, cinematic orchestral, East Asian modern, jazz & swing, and more). The agent selects up to three references with distinct roles — Foundation, Modifier, and Arrangement — then synthesizes a new, original caption around your brief. Lyric text is never reproduced or rewritten; only bracketed section tags act as musical directives.
Repository Layout
skills/
├── README.md This file
└── music-caption-rewriter/
├── SKILL.md Skill instructions
├── agents/openai.yaml Agent metadata
├── references/ Genre router + family indexes
└── templates/ Caption template libraryMaintenance
- The library is intentionally 100% text-based — no scripts, databases, embeddings, or external services.
- To add a template: add one complete caption file under
templates/, add one compact card to exactly one family index linked from the genre router, and update that index's family count. - Keep template IDs matching their filenames and their referenced paths existing.