Multi-Agent Collaboration System Skills

Universal multi-agent collaboration methodology for Claude Code. Model-tiered cowork + document-driven sync + self-evolution.

Agent Sync Skill | 智能体协作技能

For Claude Code | 适用于 Claude Code

Note: This methodology works with any multi-agent system (Cursor, LangChain, OpenAI Assistants, OpenClaw, etc.). This file is the Claude Code skill wrapper. See README.md for platform-agnostic usage.

说明：本方法论适用于任何多智能体系统（Cursor、LangChain、OpenAI Assistants、OpenClaw 等）。本文件是 Claude Code 技能封装。通用用法见 README.md。

---

Core Mechanisms | 核心机制

1. Model-Tiered Cowork | 模型分层协作: Expensive thinks, cheap executes, cheapest archives 贵的想，便宜的干，更便宜的归档
2. Document-Driven Sync | 文档驱动同步: TASK / CHANGELOG / CONTEXT = collaboration protocol TASK / CHANGELOG / CONTEXT = 协作协议
3. On-Demand Retrieval | 按需检索: QMD indexing, no full injection QMD 索引，不全量注入
4. Self-Evolution | 自演化: Repeated patterns → candidate skills 重复模式 → 候选技能

Quick Start | 快速开始

bash

# Initialize project | 初始化项目
./scripts/init.sh <project-name>

# Index documents (requires qmd) | 索引文档（需要 qmd）
qmd index .

Agent Workflow | 智能体工作流

Before Work | 开始前

text

1. qmd query "project X current tasks" → get relevant fragments
   qmd query "项目 X 当前任务" → 获取相关片段
2. Or read TASK.md directly (enough for small projects)
   或直接读 TASK.md（小项目够用）

After Work | 工作后

text

1. Update TASK.md (done → move to recent completed)
   更新 TASK.md（完成 → 移到最近完成）
2. Append CHANGELOG entry (one line, with #tag + identity)
   追加 CHANGELOG 条目（一行，带 #标签 + 身份）
3. Major decisions → update CONTEXT.md
   重大决策 → 更新 CONTEXT.md

Weekly Report | 周报时

text

1. Aggregate CHANGELOG by #tags
   按 #标签 聚合 CHANGELOG
2. Fill "Pattern Discovery" section
   填充"模式发现"板块
3. Operations 3+ times → mark as candidate skill
   出现 3+ 次的操作 → 标记为候选技能
4. Archive old data to archive/
   归档旧数据到 archive/

Model Roles | 模型分工

Role	Model Examples	Responsibilities
Lead	Opus / GPT-4 / High-capability	Architecture, decisions, task breakdown 架构、决策、拆任务
Engineer	Sonnet / GPT-4o-mini / Balanced	Execution, coding, review 执行、写代码、review
Maintainer	Flash / GPT-3.5 / Cost-effective	Archive, cleanup, weekly aggregation 归档、清理、周报聚合

Note: Model names are examples. Use equivalent models from your platform. 说明：模型名称仅为示例。请使用你平台上的同等能力模型。