grill-me
Stress-test a plan or design by interviewing the owner about it: maximize question coverage — every assumption, every edge, every dependency, every implicit decision gets a question. Use whenever plan
SKILL.mdskills/grill-me/SKILL.md
GitHubGrill Me
Home branch in Memora:
root/skill/grill-me. Reworked per owner verdict 2026-09-19 and answers 586–590 (2026-10-18): the vendor "one question at a time" discipline was wrong for this workspace. The job is maximum coverage; what survives from the original is walking the tree depth-first, exploring the codebase before asking what it can answer, and attaching a recommended answer.
Choosing a part
| Situation | Part |
|---|---|
| A plan or design exists and must be interrogated branch by branch | this page, inline (Rules below) |
| Intent is still fuzzy and the owner wants a slow, one-question-at-a-time discovery with a brief at the end | parts/interview/ |
| The request arrived already shaped as a solution and the direction itself may be wrong | parts/nodumb/ |
| A product or interface must be invented or reworked and the options look equivalent | parts/ask-nodumb/ |
Rules
- Coverage is the metric. An unasked question is a decision made silently later. One branch at a time (dependency order), but the questions inside a branch are batched.
- Explore first. Anything answerable from the codebase, the workspace indexes or prior owner verdicts — resolve it yourself and state the finding as an assumption for confirmation, not as a question.
- Extract the decision tree, then store it. Scope, data, behavior, edge
cases, failure modes, tooling, rollout, rollback, what is out of scope. Write
that tree as a Memora node under
root/skill/grill-mebefore the first question; the answers land in the same node. - Batch by branch through
ask_questions— up to 20 questions per call while the batch stays inside one branch; each question carries a one-line recommended answer with its consequence. Unresolved branches spawn the next batch; a plan of any real size takes several batches, which is success. - Record as you go. Confirmed answers go into the Memora node in the same turn; the chat gets a short summary, not the transcript. Write your version of the understanding down once the owner confirms it, not before.
- Stop when coverage is complete: every branch resolved or explicitly deferred by the owner, and no question left that has no consequence. Then summarize the locked decisions and the open deferrals.
Not a substitute for AGENTS.md: during execution, questions come at the moment
a deviation appears. Grill-me is for the plan stage, where the owner wants the
full interrogation. No CONTEXT.md / ADR ceremony, no session-tracker scripts.
Related
planning, brainstorming, the challenge part of the review router, interview.