菜单
Use when about to create a skill, start a feature, design a system, or plan any non-trivial work. Also use when user says grill me, question my plan, challenge this, stress-test, or validate my thinking. Prevents building the wrong thing by resolving decisions before implementation.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use after generating code, after accepting AI suggestions, or when reviewing AI-written modules. Also use when code "works" but feels brittle, when error handling seems thin, when you notice orphaned resources or missing cleanup, or when the agent says "done" but you suspect hidden debt. Catches the specific failure patterns AI agents produce that humans wouldn't.
Generate architecture diagrams, dependency maps, and system documentation directly from your codebase. Reads source files, traces dependencies, identifies boundaries, and outputs Mermaid/PlantUML/C4 diagrams. Use when asked to document architecture, visualize dependencies, generate system diagrams, create C4 models, map service boundaries, understand a codebase structure, or produce technical documentation from existing code.
Generate a CONTEXT.md shared vocabulary for any project. Reduces agent verbosity by 50-75%, improves variable naming, makes conversations more precise. Interviews the user about domain terms, then produces a structured glossary the agent references every session. Use when starting a new project, onboarding to unfamiliar code, the agent is being too verbose, or conversations have too much jargon confusion. Also use when user says "build context", "shared language", "define terms", or "the agent doesn't understand my project vocabulary".
Enforce schema discipline: queries-first design, index-every-filter, safe migrations, verify-at-scale thinking. Prevents the common failure of designing schemas for demos that break in production. Use when creating tables, planning migrations, designing data models, or when user mentions database, schema, migration, indexing, multi-tenant.
Use when making UI/frontend changes in Cursor's Design Mode, when the user selects elements visually, draws annotations, or uses voice commands to describe changes. Also use when editing components from the integrated browser, when multi-selecting elements to match styles, when making iterative visual refinements, or when the agent receives visual context (screenshots, element selections, DOM references) alongside a change request.
Build resilient systems with proper error handling, retry patterns, circuit breakers, graceful degradation, and timeout management. Use when implementing API calls, database connections, external service integrations, background jobs, or any code that can fail. Also use when the user mentions retry logic, circuit breaker, graceful degradation, error boundary, timeout handling, dead letter queue, or asks how to make code more resilient.
| name | grill |
| description | Use when about to create a skill, start a feature, design a system, or plan any non-trivial work. Also use when user says grill me, question my plan, challenge this, stress-test, or validate my thinking. Prevents building the wrong thing by resolving decisions before implementation. |
Resolve every decision branch before committing to implementation. Ask questions ONE AT A TIME until the full picture emerges.
Ask me questions ONE AT A TIME about what I'm about to build. For each question, provide your recommended answer. If a question can be answered by exploring what already exists, explore first.
Cover these branches:
1. WHO is this for? (specific user, not "developers")
2. WHAT specific pain does this eliminate? (named, felt, recurring)
3. WHY doesn't a solution already exist? (searched? what did you find?)
4. WHAT'S the simplest version that would be useful? (not the dream — the MVP)
5. HOW would someone find this? (discovery path — search terms, marketplace, word of mouth?)
6. WHAT makes someone SWITCH from their current approach? (the switching cost must be worth it)
7. WHAT could go wrong? (failure modes, edge cases, misuse)
8. HOW do you KNOW it works? (verification — not "I think" but "I proved")
ACTIVE until all branches resolved. Don't let me start building until questions are answered. If I try to skip ahead: "We haven't resolved [branch]. Answer first."
Stop grilling when:
"Let me just start and see where it goes" ← This is the PROBLEM grilling prevents. Every hour spent grilling saves 10 hours of building the wrong thing.
Agent: "WHO is this for? My recommendation: solo developers working on
side projects who hit the same deployment friction every weekend."
User: "Actually it's for teams — 3-5 devs sharing staging environments."
Agent: "Good — that changes scope significantly. WHAT specific pain does
this eliminate? My recommendation: 'We broke staging again because
two people deployed at the same time.'"
Key: ONE question, recommend an answer, let the user correct. Don't ask three things at once.
"No-one knows exactly what they want." — The Pragmatic Programmer
The fix is to DISCOVER what you want through structured interrogation before committing to implementation.