Craft — Engineering Discipline for AI Agents
9 skills that turn AI agents into disciplined software engineers. Battle-tested on a 4-agent team shipping production software.
Skills
| Skill | What It Enforces |
|---|
| INLINECODE0 | No production code without a failing test first |
| INLINECODE1 |
No fixes without root cause investigation |
|
craftwork:verification-before-completion | No "done" claims without fresh evidence |
|
craftwork:brainstorming | No code without exploring approaches first |
|
craftwork:writing-plans | No implementation without exact file paths, code, and tests |
|
craftwork:subagent-driven-development | One subagent per task, structured review between each |
|
craftwork:dispatching-parallel-agents | Safe parallel execution for independent tasks |
|
craftwork:requesting-code-review | Structured diff review with security and spec checklist |
|
craftwork:receiving-code-review | Technical rigor over performative agreement |
Install
CODEBLOCK0
Philosophy
Agents are fast but sloppy. They skip tests, guess at fixes, claim "done" without checking, and agree with every review comment. These skills add guardrails — not to slow agents down, but to make their speed count.
Every skill follows the same pattern: stop, verify, then act.
技能名称：craftwork
详细描述：
Craft — AI代理的工程规范
9项技能，将AI代理训练成严谨的软件工程师。已在4个代理组成的团队中通过生产级软件交付实战验证。
技能列表
| 技能 | 强制要求 |
|---|
| craftwork:测试驱动开发 | 无失败测试则不得编写生产代码 |
| craftwork:系统性调试 |
未查明根本原因则不得修复问题 |
| craftwork:完成前验证 | 无最新证据则不得声称已完成 |
| craftwork:头脑风暴 | 未先探索方案则不得编写代码 |
| craftwork:编写计划 | 无精确文件路径、代码和测试则不得实施 |
| craftwork:子代理驱动开发 | 每项任务一个子代理，任务间进行结构化评审 |
| craftwork:调度并行代理 | 对独立任务执行安全的并行处理 |
| craftwork:请求代码审查 | 基于安全和规范清单的结构化差异审查 |
| craftwork:接收代码审查 | 技术严谨性优先于表面认同 |
安装
bash
clawhub install craftwork
核心理念
代理速度快但易粗心。它们会跳过测试、猜测修复方案、未经检查就声称已完成，并对每条评审意见表示认同。这些技能增加了防护栏——不是为了拖慢代理，而是让它们的速度真正发挥作用。
每项技能遵循相同模式：暂停、验证、再行动。
