英文 |한국어

音乐作曲 技艺代理人

一套模块化的知识和工作流程技能，专为AI音乐作曲助手设计。

这个库旨在帮助经纪人将模糊的创意请求——“让副歌更有冲击力”，“给我一个更黑暗的K-pop桥段”，“重新和声为爵士但不过于密集”，“为我的播放列表找参考曲目”——转化为具体的音乐决策：和弦、旋律形状、律动、形式、编曲动作、诊断反馈和研究流程。

当前发行： v1.0 发行日期： 2026-04-27

---

安装

最简单的方法是通过 skills CLI（需要Node.js）：

该程序安装到您当前项目的代理目录（例如 .claude/skills/）。添加 -g 以实现用户级（全局）安装，或添加 --all -y 以跳过提示。

手动安装时，克隆仓库并指向你的代理（或者复制SKILL.md及相关子目录到代理的技能文件夹，比如项目中的~/.claude/skills/或.claude/skills/）：

兼容性

该技能遵循开放的代理技能标准——Claude（Claude Code、Claude Agent SDK、Claude Solutions）使用的 SKILL.md + YAML前备式格式。

English | 한국어

Music Composition Agent Skill

A modular knowledge and workflow skill for an AI music-composition assistant.

This repository is designed to help an agent turn vague creative requests — “make the chorus hit harder,” “give me a darker K-pop bridge,” “reharmonize this like jazz but not too dense,” “find reference tracks for my playlist” — into concrete musical decisions: chords, melody shapes, grooves, forms, arrangement moves, diagnostic feedback, and research workflows.

Current release: v1.0
Release date: 2026-04-27

---

Installation

The simplest way is via the skills CLI (requires Node.js):

npx skills add SJY051/music-composition

This installs to your current project’s agent directory (e.g., .claude/skills/). Add -g for user-level (global) install, or --all -y to skip prompts.

For a manual install, clone the repository and point your agent at it (or copy SKILL.md and the relevant subdirectories into your agent’s skill folder, such as ~/.claude/skills/ or .claude/skills/ in a project):

git clone https://github.com/SJY051/music-composition.git

Compatibility

The skill follows the open Agent Skills standard — the SKILL.md + YAML frontmatter format used by both Claude (Claude Code, Claude Agent SDK, Claude.ai skills) and OpenAI (Responses API tools[].environment.skills). It works on any harness that reads this format.

---

Does it work?

Two benchmarks ship with v1.0:

Approach 1 (system-prompt injection, GLM-5.1, blind eval). With SKILL.md prepended to the system prompt, the skill version wins 6 of 8 representative composition prompts in an author-bias-controlled blind comparison. The strongest signal is the boundary test (P8) — the skill version explicitly declines an out-of-scope mixing question while the baseline answers freely.

Approach 2 (native lazy loading, Claude Opus 4.7). With the skill installed via Claude Desktop’s Skills feature, the model reads references on demand. Most striking: pansori OST advice cites yoseong, toeseong, chuseong with cents-and-milliseconds implementation values that exist only inside references/genres/korean-traditional.md — direct evidence the model used the skill’s full content, not just SKILL.md.