Run a /grilling session, using the /domain-modeling skill, with three changes:

Split the frontier. Every decision in the design tree is either a functional requirement — observable behaviour the system must deliver, from the user's perspective — or a technical detail — architecture, data model, API shape, error handling, numeric defaults, anything the system can settle. Ask the user only the functional-requirement questions. Settle every technical detail yourself: pick your recommendation and record it for the recap, don't ask.
Write the questions you do ask in STE-style plain English. Short sentences, one idea per sentence, common words, no jargon. Keep the standard question shape — ❓ numbered with a title, your recommendation alone on a ➡️ line — and give a recommendation for every functional-requirement question too, so the user can answer by number.
Gate all writing on a recap. Nothing lands in CONTEXT.md or docs/adr/ until the frontier is empty and the user has approved the recap. Present, in one screen: the functional requirements you confirmed, and every technical detail you assumed — each with its recommendation and a one-line why. The user rejects any; reopen those branches and recompute the downstream tree. Only then write: terms to CONTEXT.md, decisions that pass the three gates to docs/adr/.

grill-novice