prd-design

prd-design — PRD Interview (Human Gate 1)

Produce docs/forge/prd/NNN-name.md: the contract consumed by plan-design. The acceptance criteria you write here become the loop's exit conditions, so they must be verifiable.

Two habits that run through the whole interview

These apply to every question you ask below, including the divergent phase.

Recommend, don't just enumerate. Any time you offer a choice — an approach, an AC phrasing, a constraint, the next step — present it as a lettered list and mark exactly one option Recommended with a one-line, concrete reason. Ground the reason in something real: matches an existing pattern in the codebase, lowest risk, smallest scope, fits a constraint the user already stated. The list lets the user answer with a single letter; you carry the weighing so they don't have to evaluate every option cold. Two options is the default — add a third only when it is a genuinely distinct, worthwhile path, never padding; cap at three. Ground every why in something you actually checked — a file you read, a lesson, a constraint the user stated — not an assumption. If the user's preference contradicts that evidence, say so plainly before you defer ("you can, but lessons.md shows this pattern escalated"); caving to please produces a worse PRD, not a happier user. Format:

<question?> a) — Recommended, <concrete why, e.g. "matches current pattern in auth.ts"> b) — c) —

This applies especially to next-step handoffs between skills/commands (e.g. PRD → spec → plan): always lay the choices out as a lettered list, never bury them in prose.

Close every question — assume nothing. The interview covers a fixed set of topics (the six in Step 2). Treat them as an open-questions ledger: it is done only when the user has answered every one. A default you proposed is not an answer until the user confirms it — nothing enters the PRD on your authority alone. Users will interrupt, jump ahead, or try to short-circuit ("looks good, just write it"); when they do, absorb the new information, then name what is still open and return to it. A silently skipped question is a hole in the contract that resurfaces downstream as drift or a missing acceptance criterion — closing all of them is exactly what this gate exists for. Resuming sounds like: "Got it — that covers US-2's criteria. Still open: Non-Goals and Constraints. Next: ."