使用方法
- 复制安装命令,在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物,并对照来源检查输出。
npx skills add https://github.com/bongrealty/skillcraft --skill grill-me开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况,不等于独立质量或安全测试。
按源仓库内容呈现,保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Enter adversarial review mode. Challenge the current changes, implementation, or approach. Do not make a PR or proceed until the user passes the review.
Modes
A. Code Review (default)
"Grill me on these changes and don't make a PR until I pass your test."
- Read the current diff (
git diff mainor staged changes) - Identify risks: edge cases, missing error handling, performance issues, security holes, missing tests, unclear naming
- Ask the user 3–5 hard, specific questions about the changes
- Do not accept vague answers — probe further if needed
- Only approve (and proceed with PR) once you're satisfied
- Alternatively: "Prove to me this works" → diff behavior between main and feature branch, run tests, check logs
B. Elegant Rewrite
"Knowing everything you know now, scrap this and implement the elegant solution."
- After a mediocre or working-but-messy fix, start fresh
- Identify the root insight that makes the solution simple
- Rewrite with that insight as the foundation — no legacy cruft
- The elegant solution is usually 30–50% less code
C. Spec-Driven Development
"Write detailed specs before handing off this work."
- Ask the user clarifying questions until requirements are unambiguous
- Produce a written spec: inputs, outputs, edge cases, constraints, acceptance criteria
- Only begin implementation once the spec is approved
- The more specific the spec, the better the output
Review Checklist (Mode A)
- [ ] Does this handle the unhappy path?
- [ ] Are there race conditions or async issues?
- [ ] Is this the right abstraction, or is it over/under-engineered?
- [ ] Will this break anything that depends on it?
- [ ] Are there missing tests for the new behavior?
- [ ] Is the naming clear to someone reading this cold?
- [ ] Could this be 50% simpler?
Example
User: grill me on these changes and don't make a PR until I pass your test
→ Claude reads diff, asks 4 hard questions
→ User answers
→ Claude probes weak answers
→ User satisfies all concerns
→ Claude: "You passed. Creating PR now."

