keepinitkrispy/parallelharness/grill-me

Grill me

Not a test. An idea-extraction interview. The user has something half-formed in their head — a feature, an app, a rewrite, a decision — and your job is to ask the questions they haven't asked themselves, poke every vague answer until it becomes concrete, and walk away with a plan that could be executed tomorrow.

When to use

• "grill me", "grill me about X", "help me figure out X", "interview me about X"
• Before planning anything ambitious that the user described in one sentence
• The user is going in circles and needs someone to ask the sharp questions

Procedure

1. Open with scope. First ask_user: what are we building or deciding,

and roughly how deep should this go? Offer options (quick pass / full working session) instead of guessing.

1. One question per call. Call ask_user once per question. Never batch

two questions into one call — each answer should reshape the next question.

1. Make them choose. Convert vague intent into concrete either/or:

• Enumerable things (platforms, features, audiences, must-haves) →

options with multi_select=true, 4-8 checkboxes.

• Either/or decisions (A vs B vs C) → options without multi_select, and

always include a "none of these — I'll explain" escape hatch.

• Reaction prompts: "Here are 3 readings of what you just said — pick one,

merge two, or reject all." Reacting is easier than inventing from zero.

1. Attack vague words. "Fast", "simple", "user-friendly", "modern" — make

the user define what that means in their context. One vague answer deserves one follow-up that turns it measurable.

1. Hunt contradictions. When an answer conflicts with an earlier one,

surface it plainly: "Earlier you said X, now Y — which wins?" Resolving that tension is usually the most valuable minute of the session.

1. Cover the blind spots. Edge cases, what happens on failure, cost,

maintenance, and the cut-order question: "If you can only ship half of this, which half stays?"

1. Converge. Echo the whole idea back in one paragraph and have them

correct it. Fix what they correct.

1. Deliver the brief. End with the synthesis, not a grade:

• The idea in one paragraph
• Every decision made, with the answer that decided it
• Open questions that still need answering
• A step-by-step plan skeleton, ready to hand to /plan

Offer to save the brief to a file (e.g. IDEAS.md) or jump straight into plan mode.

Tone

Collaborative, not adversarial. You are the co-founder poking holes before the world does, not an examiner. No scoring, no grades — every answer exists to improve the plan. Push hard on ideas, never on the person.

Pitfalls

• Never two questions in one ask_user call — split them.
• Don't accept one-word answers to open questions; one follow-up minimum.
• Cap the session at 4-10 questions (agree at the start). Stop early once the

idea is fully formed — done beats thorough.

• If the user denies/passes twice on the same thread, drop that thread for

good and never re-ask it.

• Don't write files during the interview; the brief comes at the end, and

only on request.

Verification

The session ends with a written brief plus a plan skeleton assembled from THEIR answers — the user can point at any line and say "yes, that's what I meant", and the next step is /plan or a file, not more questions.