phase1-prd-grill

The first phase of a project's life: ideation → a PRD that captures the essence, hardened by adversarial interviewing of the human and refined in a loop until the human declares it good enough.

The PRD is the durable artifact the whole trilogy stands on. A weak PRD here makes phase 2's backlog and phase 3's implementation confidently build the wrong thing. So this phase's job is not to write prose — it is to find the soft spots in the idea before they become code.

When to use

• A new project / feature at the idea stage; "write a PRD", "refine the spec", "help me scope this", "is this idea sound".
• Any time the next phases would otherwise proceed on unstated assumptions.

Do NOT use once a satisfactory PRD exists (go to phase2-backlog-snowball), or for a mechanical change with no design question.

Core principle

• Loop, don't draft-once. Elicit → draft/refine the PRD → adversarially attack it → grill the human on the weak points → commit → repeat. Many rounds is normal and expected.
• The human owns "done", not the AI. The exit condition is the human explicitly saying they are satisfied. The AI never declares the PRD finished.
• Grill, don't gather. Default mode is skeptical interviewer, not stenographer. Surface weak/unstated assumptions, missing failure modes, absent success criteria, scope ambiguity, "why not the simpler thing", who the user really is, what would make this fail.
• No yes-manning. Anti-sycophancy is a hard rule: name the problems, gotchas, and limitations of the idea plainly. Validating a weak idea to be agreeable is the primary failure mode of this phase.

The loop (one round)

1. Elicit / re-elicit

Ask for the idea (or the part still fuzzy). One tight cluster of questions at a time — not a 20-question wall. Prefer offering concrete options/strawmen the human can react to over open-ended prompts; reacting is higher-bandwidth than composing.

2. Draft or refine the PRD

Write/expand PRD.md (repo root or agreed path). Required sections, each kept honest:

• Essence / problem — one paragraph; what is actually being solved and why it matters. If you can't state it crisply, that is itself the finding to raise.
• Users / actors — who, concretely.
• What good looks like / what bad looks like — observable success and failure. This directly seeds phase 2's test-grounding; an absent or vague version here is a blocker, not a TODO.
• Scope & non-goals — explicit non-goals carry as much weight as goals.
• Key decisions & open questions — with the chosen option and why.
• Weak assumptions & risks — list them explicitly; this section must not be empty for a real project.
• Irreversibility map — which surfaces of the project CANNOT be cheaply walked back once shipped (on-disk/wire formats, data- integrity write paths, migrations, published APIs, anything that freezes) versus which are tentative/replaceable. This is a load- bearing deliverable, not prose color: the later phases run functionality-first — features ship at velocity on lean review, deep rigor is applied ONLY at the surfaces this map names, and hardening is deferred until surfaces stabilize. A missing map means phase 3 either deep-reviews everything (slow) or nothing (dangerous); grill the human on it like any other weak point.
• Tentative-vs-Committed — an explicit field with exactly one value: firm or experimental. Phase 2 branches on this exact field (firm → full breakdown; experimental → wave planning), so it must be a discrete stated value, not a vibe inferred from prose. Set it deliberately and revisit it each round.

3. Adversarial critique (independent)

Once the PRD has enough substance to attack (typically round 2+, not the first one-paragraph sketch), spawn a skeptical architecture reviewer subagent (mped-architect, read-only) on the current draft with: "attack this — unstated assumptions, missing failure modes, scope traps, simpler alternatives, what a hostile reviewer would say. Do not be nice." Fold its findings into the next grill instead of quietly absorbing them. Skipping this on a near-empty early draft is correct, not laziness; spawning it on a mature draft and ignoring its findings is the anti-pattern. (Read-only review never collides; same honest-failure spine the rest of the trilogy uses.)

4. Grill the human

Put the sharpest unresolved weak points to the human as pointed questions (use the question UI; make options concrete; mark a recommended option only when you genuinely have one and say why). Pursue the assumption that, if wrong, breaks the project — not cosmetic gaps.

5. Commit

Commit PRD.md every round (small, message like prd: <what sharpened> (round N)). The git history of the PRD is the record of how the idea hardened; it makes rollback and "why did we decide that" free. Incremental commits throughout phase 1 are mandatory.

Repeat until the human explicitly says it's good enough.

Exit condition & handoff

• Exit only on explicit human satisfaction. If asked "is it done?" the honest answer is "that's your call — here are the weak points I still see"; never close it yourself.
• Handoff to phase2-backlog-snowball: the PRD must carry (a) the good/bad/observable-success seeds phase 2 turns into test grounding, (b) the Tentative-vs-Committed field that decides wave-planning, and (c) the irreversibility map phase 2 turns into irreversible task labels and phase 3 turns into deep-review gates. If any is weak, phase 1 is not done regardless of how polished the prose is. Record the PRD's final path in the last commit message so a cold phase-2 agent locates it without guessing.

Anti-patterns

• Yes-manning a weak idea; softening findings to be agreeable.
• Drafting the PRD once and presenting it as done.
• Accepting vague/absent success & failure criteria ("it should work well").
• Declaring the PRD finished on the human's behalf.
• Boiling the ocean: an exhaustive PRD for an idea the human flagged as experimental — capture the essence and the unknowns, not a fantasy of certainty.
• Asking 20 shallow questions instead of pursuing the one load-bearing assumption.