grill-melisted

# Grill Me — Adversarial alignment before planning > **Principio destilado** (Matt Pocock + Frederick P. Brooks): the goal is NOT to produce a plan fast. The goal is to reach a **shared design concept** with your human partner. The conversation IS the asset. ## Protocol Interview your human partner relentlessly to reach shared understanding. Walk down each branch of the design tree, resolving dependencies one by one. For each question, provide your **recommended answer** and let the user agree, refine, or correct. Ask **ONE question at a time**. Skip questions where Neurox or context already provides clear answers. Stop when the design tree is fully resolved or the user explicitly says "just do it". ## Output Write `design-tree.md` (or update if exists) at the repo root or `docs/specs/<feature>/design-tree.md`: ```markdown # Design Tree — <feature> ## Resolved decisions - D1: <decision> — <rationale> - D2: <decision> — <rationale> ## Open assumptions (validate before PRD) - A1: <assumption> — <impact if wrong> ## Out of scope (explicit) - <thing not addressed> ## Ready for PRD ✅ yes / ⚠️ pending: <list> ``` ## Rules 1. **One question at a time**. Never dump a list. 2. **Always provide a recommended answer**. The user can agree, modify, or override — but you take the cognitive load of suggesting first. 3. **Skip the obvious**. If Neurox already has the answer, surface it instead of asking. 4. **No premature PRD**. Do not generate a PRD inside this skill — that's th