Grill-Me Skill¶
A tiny Claude Code skill from matt-pocock whose job is to flip the interview direction: instead of the human prompting the AI, the AI interrogates the human until both agree on what is being built.
"It's the grill-me skill. And this grill-me skill is wonderfully small, wonderfully tiny, and it helps prevent one of I think the main issues when you're working with an AI, which is misalignments." — Pocock, 12:19
The skill body (paraphrased from repo)¶
Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the decision tree resolving dependencies one by one. For each question provide your recommended answer. Ask the questions one at a time.
That's essentially all of it — a few lines of markdown with a pull-trigger description header.
Why it works¶
- Alignment precedes output. Pocock cites Frederick P. Brooks (The Design of Design): successful collaborators converge on a shared design concept. Plan mode fails because the agent rushes to produce a plan artifact; grill-me forces a conversation instead.
- Provides recommendations. Every question ships with the model's own best guess, so the human can ratify rather than author — ~22 resolved questions in a typical session in the demo.
- Catches silent gaps. In the workshop example ("should points be retroactive on existing lesson progress?"), the grill surfaced a requirement the client (Sarah Chen) had never articulated.
- Pairs with a fresh session. The ~25K tokens burned here are "gold" and are then compressed into a PRD — see prd-as-destination — before clearing for the next phase.
Related¶
- claude-code-skills · contextual-prompt-engineering · push-vs-pull-standards — grill-me is a canonical pull skill
- prd-as-destination — the downstream artifact of a grilling session
- agent-taste · subagent-architecture (grill-me spawns an explore sub-agent before the first question)