grill — converge on intent before work

Intent: an agent that implements a half-understood ask wastes the user's time twice. The deliverable of this skill is shared understanding, not code.

Question zero — is this the right question? Verdicts:

1. Wrong question — the ask, even done perfectly, wouldn't serve the actual goal. Stop, name the underlying problem, propose the question that would serve it.
2. Right problem, misaimed ask — the goal is sound but the requirement overshoots, undershoots, or aims at a symptom. Adjust it and confirm the adjusted version.
3. Right question — proceed silently.

An honest "this smells wrong because X, though I don't have a better formulation" is a legitimate report.

Guardrails:

• One question at a time. Challenge, don't validate — your value is pressure-testing the user's judgment, not agreeing with it.
• Check the wiki first (llm-wiki skill) so settled decisions aren't re-litigated.
• No code yet. Scale depth to ambiguity: a nearly-clear ask needs one question or none.

Exit: a compact "here's what I understood — confirm" summary. For most tasks that confirmed summary IS the spec — work from it directly. If the task looks large or multi-session, offer the spec skill; create the file without asking when obviously needed.