원클릭으로
grill
Pre-implementation alignment grill. Forces the agent to ask sharp questions before writing code on a non-trivial change.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
Pre-implementation alignment grill. Forces the agent to ask sharp questions before writing code on a non-trivial change.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Core working-language policy for this harness. Think in English, deliver in Chinese.
Disciplined bug-diagnosis loop. Reproduce → minimise → hypothesise → instrument → fix → regression-test.
Recipe for filling state/reference/ENV.md when the project is an Android app. Use when the project has a build.gradle / settings.gradle and targets Android. Covers emulator boot, gradle test/connectedAndroidTest, adb logcat capture.
Recipe for filling state/reference/ENV.md when the project is a backend service (API, worker, daemon). Use when the project ships a server but no UI; covers compose-based DB/cache, pytest/jest test runners, and container log capture. Adjust language ecosystem (Python/Node/Go/etc.) accordingly.
Recipe for filling state/reference/ENV.md when the project is a web app (frontend or full-stack). Use when the project serves HTTP, has a dev server, or runs in a browser. Adjust package manager (pnpm/npm/yarn/bun) and test runner to match the project.
Zoom out before zooming in. Read the surrounding system before changing an unfamiliar piece of code.
| name | grill |
| description | Pre-implementation alignment grill. Forces the agent to ask sharp questions before writing code on a non-trivial change. |
借鉴 mattpocock/skills 的 /grill-me. 修复最常见的失败模式:Agent 以为自己懂了,其实没懂。
CURRENT_TASK.md → Intent 不到 2 行 / 没说清边界跳过 grill:trivial 字面修复(typo / 改一个常量 / 改一行配置)。
逼着自己回答下面这组问题。不许跳。把答案写到 CURRENT_TASK.md 里 ## Grill 段(或者口头答完再开工,但要在 JOURNAL 留痕)。
pnpm test foo、curl /…、UI 步骤。DECISIONS.md.CURRENT_TASK.md 里追加:
## Grill (前置审讯 - <date>)
1. 变了什么:…
2. 没变什么:…
3. 谁会坏:…
4. 验证:`<runnable command>`
5. 回滚:…
6. 抽象:…
7. **最不确定的假设:…** ← 风险根源
如果 #7 的假设无法当回合验证 → 把它拆出来当一个 research 任务先做。