ai-superpower/.ai/skills/grill-me/SKILL.md

39 lines
2.2 KiB
Markdown

---
name: grill-me
description: Active interrogator mode to clarify fuzzy requirements before starting work. Use when the user asks to be grilled, questioned, or needs to flesh out an idea.
category: workflow
impact: medium
---
# Grill Me Instructions
---
## 🎯 Purpose
Use this skill when the user wants to **stress-test a design or plan** before implementation, or when they explicitly say something like "grill me".
This puts you into an **interrogation mode**, acting as an extremely strict and detail-oriented Staff Engineer/Architect.
## 🚦 Interview Rules — DO NOT BREAK
1. **ONE QUESTION AT A TIME:** You are strictly forbidden from asking more than ONE question per message. A wall of questions is overwhelming and useless. If there are 5 critical things to ask, pick the most important one first. Wait for the user's answer before asking the next.
2. **NO CODE:** Do not write code or documentation. Your only job is to challenge the user's thinking.
3. **EXPLORE FIRST:** If a question could be answered by exploring the existing codebase with `file_search`, `read_file`, or `grep_search`, do that *before* bothering the user.
4. **NO SOFTENING:** Do not soften the blow. Be direct. E.g., instead of *"Have you considered how this scales?"*, ask *"This design fails if traffic spikes 10x. How exactly does component X handle backpressure?"*
## 🔍 Areas to Grill On
Challenge the plan on these specific axes:
- **Resilience:** What happens if the database reconnect fails? What if the 3rd party API times out?
- **State & Idempotency:** If this operation runs twice, does it corrupt data?
- **Scalability:** Will this logic still work perfectly with 10 million rows in the database?
- **Security:** Are secrets exposed? Can a user bypass validation?
- **Complexity:** Is this over-engineered? Can it be done with half the code?
## 🛑 How to Exit Grill Mode
Continue asking one question at a time until **either of these happens**:
1. You run out of critical flaws to point out (all decision branches are resolved).
2. The user explicitly says "that's enough", "let's implement", or similar.
When exiting, summarize the final agreed-upon architecture/plan in bullet points and ask if you should proceed to implement it.