Prd Creator

PRD — front door of the plan→execute pipeline

Draft docs/prd-<slug>.md — the product spec that defines what to build and why, grounded in the real codebase, structured so the downstream skills can consume it directly.

/prd-creator  →  /research  →  /plan-phase  →  /implement-phase
prd-<slug>.md   technical-decisions-<slug>.md   phase-NN-<slug>.md   phase-NN-<slug>.progress.md

The slug is the spine. Pick it once here; every downstream artifact reuses it. A PRD is not a technical-decisions doc and not a phase plan — it states the problem, the scope, the constraints that must not break, the open decisions to resolve, and a coarse phase outline. It hands the technical "how" to /research and the SI-level breakdown to /plan-phase.

References (load only what the step needs)

Workflow

Step 1 — Ground in reality first (before asking anything)

A PRD that hallucinates the current system is worse than none. Inspect the codebase to learn the actual starting state:

• Find the surfaces the feature touches (components, modules, services, routes). Use Grep/Glob on the real names — quote concrete paths in the PRD (e.g. Dashboard.tsx, insight-tools.factory.ts), the way existing PRDs do.
• Read related PRDs in docs/prd-*.md to reuse vocabulary and link them with [[wikilinks]].
• Read the upstream tech-discovery brief (docs/architecture/*-tech-discovery.md) and accepted ADRs (docs/adr/) — these are architectural hard constraints: reflect them in §4 (Restrições críticas) and don't reopen them. §6 (Decisões a tomar) is only for the granular forks left to research, never the architecture already decided in an ADR.
• Note what already exists vs. what's missing — this becomes the "O que já temos vs. o que falta" table and prevents over-scoping.

Capture findings before the interview so you ask sharper questions and skip what you can infer.

Step 2 — Interview (one question at a time, with a recommendation)

Cover, in order (rationale + good/bad examples in references/quality-bar.md):

1. Problem — what hurts today, for whom, grounded in the current state you just mapped.
2. Objective & opportunity — what changes if we ship this; why now.
3. Out of scope (Não-objetivos) — what this version intentionally excludes. Push for this — it's the strongest defense against scope creep.
4. Critical constraints (NÃO QUEBRAR) — existing contracts, flows, wire formats, auth/ tenant behavior, public URLs that must keep working. Enumerate as R1, R2… These are the "gotchas" that most control AI-generated code quality.
5. Open decisions — every fork with a real alternative (lib/strategy/pattern/storage/ limits). These feed /research; recommend a default but mark "a confirmar".
6. Acceptance criteria / success metrics — testable conditions (Given-When-Then for flows; security/parity/DoD buckets; product metrics with targets). Aim for 3–7 per story.
7. Coarse phase outline — a rough sequence of phases (not SIs). Detail goes to /plan-phase.

If the user says "just draft it / skip the interview", infer from Step 1 + the conversation, mark every assumption with (a confirmar), and proceed.

Step 3 — Choose the slug

A short kebab-case slug describing the feature: dynamic-charts, llm-gateway-module, auth-rate-limiting. This becomes docs/prd-<slug>.md and propagates to docs/decisions/technical-decisions-<slug>.md and docs/phases/phase-NN-<slug>.md. Confirm it doesn't collide with an existing docs/prd-*.md.

Step 4 — Write docs/prd-<slug>.md

Use assets/prd-template.md and the section spec in references/prd-structure.md. Rules:

• Portuguese (PT-BR) to match existing docs, unless the user writes in another language.
• Ground every claim in real files/paths; no invented frameworks or endpoints.
• Make Não-objetivos and Restrições críticas (NÃO QUEBRAR) explicit and specific.
• Put genuine forks under Decisões a tomar (preencher via /research) — don't pre-decide what deserves research; recommend, don't dictate.
• Keep the phase plan coarse — name phases, not step-implementations.
• End with Próximos passos as a checklist that names the next pipeline steps (/research, /plan-phase).

Step 5 — Validate against the quality bar

Before declaring done, self-check with references/quality-bar.md:

• Is the problem grounded and specific (not generic)?
• Is out-of-scope non-empty and concrete?
• Does every critical constraint name a real thing that could break?
• Is each acceptance criterion testable?
• Are open decisions actual forks (not obvious choices), ready for /research?
• Does the phase outline map cleanly to /plan-phase?

Fix gaps, then report the path, the chosen slug, the open decisions count, and the suggested next command (/research <slug> or, if no research needed, /plan-phase).

Output format

Return:

1. Summary — feature, slug, one-line problem & objective.
2. File — docs/prd-<slug>.md written.
3. Open decisions — list of forks to resolve in /research (or "none → go straight to /plan-phase").
4. Next step — exact command to run next.