Skill：agent harness bootstrap
Specs both you and the AI understand. An agent team fitted to your repo. Guardrails that block, not advise. Two Claude Code skills - spec-builder + harness-bootstrap - for greenfield, brownfield, or audit-only.
适用场景
招聘人事
适合角色
介绍
以下内容整理自开源仓库 `nguyenhx2/agent-harness-bootstrap` 的 README / SKILL.md 摘要：
<p align="center">
<img src="docs/assets/readme-banner.png" alt="Agent Harness Bootstrap - the frame that lets AI agents operate autonomously, and safely" width="100%">
</p>
<p align="center"><b>Give an AI agent a repo it can actually understand, and a harness it cannot escape.</b></p>
<p align="center">by <a href="https://github.com/nguyenhx2">nguyenhx2</a> · <b>English</b> · <a href="README.ja.md">日本語</a></p>
[](https://github.com/nguyenhx2/agent-harness-bootstrap/actions/workflows/eval.yml) [](LICENSE) [](harness-bootstrap/assets/claude/agents/)
[](eval/guardrail_eval.py) [](https://claude.com/claude-code) [](https://github.com/nguyenhx2/agent-harness-bootstrap/releases/latest)
📊 [Slide presentation](https://nguyenhx2.github.io/agent-harness-bootstrap/presentation/) · 🎥 [Video gallery](https://nguyenhx2.github.io/agent-harness-bootstrap/video/) · 📦 [Latest release](https://github.com/nguyenhx2/agent-harness-bootstrap/releases/latest) · 📚 [Docs map](#-docs-map)
---
## 🎬 What it does
Two skills for **Claude Code** that fix four specific, recognizable failure modes of an unmanaged AI
coding agent:
| Before | After |
|---|---|
| You ask an agent for a feature. It edits 14 files across 3 modules and force-pushes to `main`. | It works inside scoped agents and path-based rules; a blocking **hook** (a script that intercepts a risky action and refuses it) stops the push before it lands. |
| The session compacts and the agent forgets the plan it was three steps into. | The task board (`docs/tasks/`) and its session log live on disk, not in context - a fresh session resumes exactly where the last one stopped. |
| The 40th agent-generated doc quietly contradicts what the spec says. | `spec-builder` gives every requirement a stable ID; the traceability graph (`docs/context/specs-graph.html`) flags the doc that drifted. |
| It reads `.env`, a private key, or `~/.ssh/` while "just fixing a bug." | Those paths are denied at the permission layer before the read happens - it cannot leak what it was never allowed to open. |
<p align="center">
<a href="https://nguyenhx2.github.io/agent-harness-bootstrap/video/">
<img src="video/gif/04-solution.gif" alt="The complete solution: the pain points, spec-builder wr
---
安装方式
1. 打开仓库：https://github.com/nguyenhx2/agent-harness-bootstrap 2. 查找 SKILL.md 或 README 中的安装章节 3. 克隆或下载所需 Skill 子目录到本地 skills 路径（如 `.cursor/skills/` 或 `.claude/skills/`） 4. 重启 / 重载 Agent 客户端 5. 用一条真实任务验证 Skill 是否被正确加载
适用平台
- Claude