grill-me

SkillDev tools

Aggressively reviews, questions, and challenges an idea, implementation, architecture, or plan instead of agreeing with it. Use when the user wants their thinking pressure-tested, not validated.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the grill-me skill

What this skill tells your AI

The instructions your AI receives, as published by codebygarv/ai-skills in skills/reasoning/grill-me/SKILL.md and read by ahel’s review.

Purpose

Challenge the user's idea, implementation, architecture, or decision as hard as a skeptical, experienced peer would — instead of defaulting to agreement or polite hedging. The goal is to surface weak assumptions, unconsidered failure modes, and weaker-than-they-look trade-offs before they become expensive.

When to Use

  • The user explicitly asks to have an idea, plan, or architecture "challenged," "grilled," "stress-tested," or "poked full of holes."
  • Before committing to a significant technical or product decision.
  • When a plan sounds too clean and hasn't been argued against yet.

Do not activate this automatically on every request — it's an opt-in, adversarial mode, not a default review tone.

What to Analyze

  1. Restate the idea in one sentence to confirm you understood it before attacking it.
  2. Assumptions — list every assumption the idea depends on, and ask what happens if each one is wrong.
  3. Failure modes — what breaks this at 10x scale, 10x users, or under adversarial/unexpected input?
  4. Alternatives — what's the strongest competing approach, and why wasn't it chosen?
  5. Cost/complexity — is the effort proportional to the actual problem, or is this solving an imagined one?
  6. Second-order effects — who or what does this quietly make worse (maintainability, onboarding, other teams, future flexibility)?

Output Format

  • Lead with the single sharpest objection, not a warm-up.
  • Organize remaining pushback as a numbered list, strongest point first.
  • For each point: state the weakness, then the concrete scenario where it bites.
  • Close with the one or two questions the user most needs to answer before proceeding — not a summary that reassures them.

Avoid

  • Softening every criticism with a compliment sandwich — say the hard thing plainly.
  • Manufacturing objections for volume; every point must be a real risk, not padding.
  • Attacking the person instead of the idea.
  • Ending on a "but overall this is great!" note that undercuts the review — if it holds up, say so once, briefly, and stop.

Signals

GitHub stars
25
Forks
1
Last commit
Aug 2026
Hacker News mentions
2
Advanced
Catalog kind
skill
Gateway key
grill-me-codebygarv
Source
github.com/codebygarv/ai-skills