Learn Claude Code -- Harness Engineering for Real Agents
The model is the agent. The code is the harness.
learn-claude-code 项目的中文学习笔记，采用 QA 形式整理每个 session 的核心概念。
# 1. 安装 mdBook
./setup.sh
# 2. 启动开发服务器
mdbook serve --open笔记文件位于 src/sessions/ 目录：
learn-claude-code-notes/
├── src/
│ ├── README.md # 关于本书
│ ├── SUMMARY.md # 目录
│ ├── core-concepts.md # 核心概念
│ ├── phase-*.md # 阶段概述
│ ├── sessions/ # Session 笔记
│ │ ├── s01-the-agent-loop.md
│ │ ├── s02-tool-use.md
│ │ └── ...
│ └── summary.md # 总结
├── book.toml # mdBook 配置
├── DEPLOY.md # 部署指南
├── setup.sh # 安装脚本
└── .github/workflows/
└── deploy.yml # GitHub Actions
| 阶段 | Sessions | 主题 |
|---|---|---|
| Phase 1 | s01-s02 | The Loop - 基础循环与工具 |
| Phase 2 | s03-s06 | Planning & Knowledge - 规划与知识管理 |
| Phase 3 | s07-s08 | Persistence - 持久化与并行 |
| Phase 4 | s09-s12 | Teams - 多 Agent 协作 |
| Session | 一句话总结 |
|---|---|
| s01 | One loop & Bash is all you need |
| s02 | Adding a tool means adding one handler |
| s03 | An agent without a plan drifts |
| s04 | Break big tasks down; each subtask gets a clean context |
| s05 | Load knowledge when you need it, not upfront |
| s06 | Context will fill up; you need a way to make room |
| s07 | Break big goals into small tasks, order them, persist to disk |
| s08 | Run slow operations in the background; the agent keeps thinking |
| s09 | When the task is too big for one, delegate to teammates |
| s10 | Teammates need shared communication rules |
| s11 | Teammates scan the board and claim tasks themselves |
| s12 | Each works in its own directory, no interference |
Agent = Model（模型即智能体）
Harness = Tools + Knowledge + Context + Permissions
Product = Agent + Harness
详见 DEPLOY.md
# 本地构建
mdbook build
# 输出在 ./book/ 目录本项目使用 mdbook-zh-search 构建，它是 mdBook 的增强版本，提供了更好的中文搜索支持。
MIT
Bash is all you need. Real agents are all the universe needs.