Grill hard

A Claude Code skill for stress-testing plans and designs. It asks one question at a time, starting with the most foundational decision, and gives you a summary of every decision at the end.

Without it, Claude asks long lists of questions, gets ahead of itself, and buries the decisions that actually change the plan.

It’s a fork of Matt Pocock’s grill-me, which is good and worth using as-is. I ran mine through four rounds of evaluation: baseline comparisons, LLM-graded assertions, human review.

What changed from the fork

• AskUserQuestion is mandatory. Questions go through the structured tool rather than as text, so one-at-a-time is enforced by the format instead of by instruction.
• Reasons instead of prohibitions. “Don’t list future questions” failed repeatedly. Explaining why, that focused attention beats a comprehensive preview, held across every test case.
• A decision tree. Foundational decisions first, dependencies resolved in order.
• “I don’t know” is handled. Uncertainty gets concrete options and a park-it escape hatch.
• It reads the codebase rather than asking questions it could answer itself.

Using it

Tell Claude to grill you: “grill me on this migration plan”, “poke holes in this architecture”, “tear this apart before I ship it”.

On GitHub. MIT.