Receber por email
← Voltar para o catálogo

product-discovery

Optional path to raw content (brief, notes, tickets) or a one-line topic.

25estrelas
Atualizado há 5 meses

Ver no GitHub ↗ Licença: MIT

Como adicionar

/plugin marketplace add theagenticguy/erpaval

O comando exato pode variar conforme o repositório. Confira o README no GitHub.

Para o autor da skill

Cole no README do seu repo

Mostra que sua skill está catalogada na Skillteca, gera backlink e tráfego rastreável.

[![Skillteca](https://www.skillteca.com.br/api/badge/product-discovery-mrt6bq9b/svg)](https://www.skillteca.com.br/skills/product-discovery-mrt6bq9b)

Skills relacionadas

Ver todas de Pesquisa e Web →

Alerta por categoria

Receba novas skills de Pesquisa e Web toda segunda

1 email curto, só com as skills novas de Pesquisa e Web. 4 minutos de leitura, sem spam, cancela com 1 clique.

Você confirma o email no primeiro envio. Sem spam. Cancela com 1 clique.

Ler descrição completa

Contents

| Reference | When to load | | -------------------------------------------- | -------------------------------------------------------------------------------------------------------- | | references/orchestrator.md | Running the pipeline — 5 routes, parallel Phase 2, critic review, deliver | | references/write-protocol.md | Canonical write-protocol block copied into every subagent prompt | | references/frameworks/INDEX.md | Routing guide — discovery + spec decision tables, discovery → spec composition | | references/frameworks/*.md | Per-framework files — double-diamond, how-might-we, jtbd-job-stories, user-stories-invest, ears, gherkin | | references/methodology/discovery-rounds.md | Six-phase discovery methodology | | references/inference-heuristics.md | Signal-to-inference tables for PRD-route Intent Profile capture | | references/quality/prd.md | Per-section PRD quality bar, cross-section consistency checks | | references/roles/product-analyst.md | PRD route — vision, personas, user stories, IA, milestones (sections 1-6, 12) | | references/roles/system-architect.md | PRD route — NFRs, data models, API contracts, edge cases (sections 7-10) | | references/roles/research-scout.md | PRD route — competitive landscape, open questions (sections 11, 13) | | references/roles/discovery-lead.md | Discovery-round route — pain-point excavation, Double Diamond cadence | | references/roles/hmw-framer.md | HMW-only route (also hard-dep entry point for erpaval CL-RIGOR fuzzy) | | references/roles/jtbd-interviewer.md | JTBD-only route — job stories from PM-provided source material | | references/roles/ears-specifier.md | EARS-only route (also hard-dep entry point for erpaval CL-RIGOR contract-unclear) | | references/roles/prd-synthesizer.md | PRD route Phase 3 — composes final PRD from three parallel work logs | | references/roles/discovery-critic.md | Phase 3.5 critic — rubric-graded review of any synthesized artifact | | templates/prd-template.md | 15-section PRD skeleton with [FILL] markers | | templates/worklog-skeleton.md | Per-role work-log skeleton with embedded write-protocol | | templates/discovery-memo.md | Discovery-round output shape | | templates/double-diamond-workbook.md | Double Diamond four-stage workbook | | templates/hmw-skeleton.md | HMW brainstorms file — brainstorms/NNN-<slug>-requirements.md | | templates/jtbd-skeleton.md | JTBD job-stories file | | templates/ears-spec-skeleton.md | EARS spec file — specs/NNN-<slug>/spec.md |

Product discovery

One skill, five routes. Takes a problem at any maturity (one-sentence idea, research notes, fuzzy framing, clear feature) and routes to the right framework pipeline. File-first write protocol throughout. Every role is a Markdown prompt copied verbatim into a general-purpose Agent call — zero custom agent files.

Pipeline at a glance

flowchart TD
  A[User brief] --> B[Phase 0: Route]
  B -->|PRD ask| P1[Phase 1: Intent Profile]
  B -->|Discovery round| D1[Phase 1: discovery-lead framing]
  B -->|HMW only| H1[Phase 1: framing]
  B -->|EARS only| E1[Phase 1: framing]
  B -->|JTBD only| J1[Phase 1: framing]

  P1 --> P2[Phase 2: product-analyst + system-architect + research-scout in parallel]
  D1 --> D2[Phase 2: hmw-framer / jtbd-interviewer / research-scout as routed]
  H1 --> H2[Phase 2: hmw-framer]
  E1 --> E2[Phase 2: ears-specifier]
  J1 --> J2[Phase 2: jtbd-interviewer]

  P2 --> P3[Phase 3: prd-synthesizer]
  D2 --> D3[Phase 3: discovery-lead composes memo]
  H2 --> P35
  E2 --> P35
  J2 --> P35
  P3 --> P35
  D3 --> P35

  P35[Phase 3.5: discovery-critic] --> P4[Phase 4: Deliver]

Full route-by-route runbook, role prompts, check-in cadence, and stuck-detection recovery live in references/orchestrator.md.

When to run which route

| User signal | Route | Primary output | | ------------------------------------------------------------- | --------------- | -------------------------------------------- | | "write a PRD", "draft requirements", one-sentence idea | PRD | {product-slug}-prd.md (15 sections) | | "brainstorm a product", "run discovery", "think through this" | Discovery round | discovery-memo.md | | "turn these notes into HMW", "reframe this" | HMW-only | brainstorms/NNN-{{ slug }}-requirements.md | | "spec this for Kiro / Spec Kit", "write EARS ACs" | EARS-only | specs/NNN-{{ slug }}/spec.md | | "write job stories for [segment] from these tickets" | JTBD-only | jtbd-job-stories.md |

Default-in-doubt: PRD route. Absorbs the most source-material shapes and the 3-role parallel research pattern is the skill's strongest lane.

When the brief is genuinely ambiguous (audience unknown, source unclear), use one AskUserQuestion call with up to 4 questions. Then proceed. Do not loop.

Write-protocol discipline

Every role follows the same rhythm: edit the output file → read → edit. Partial work on disk survives subagent termination; plans held in working memory do not. The canonical block lives in references/write-protocol.md and is copied verbatim into every role prompt and every task skeleton.

{product-slug}-prd.md, discovery-memo.md, brainstorms/NNN-.md, specs/NNN-/spec.md, jtbd-job-stories.md — one of these is the source of truth per run. Nothing else becomes load-bearing.

Framework family

The skill ships one routing guide plus per-framework files. The orchestrator reads references/frameworks/INDEX.md first — it covers BOTH discovery frameworks (upstream, problem-space) and spec frameworks (downstream, contract-shaped) in one decision table because they compose as a single pipeline. Each framework's canonical structure, templates, and validation checks live in its own file in the same frameworks/ directory.

  • Discovery: ${CLAUDE_PLUGIN_ROOT}/skills/product-design-shared/references/double-diamond.md, frameworks/how-might-we.md, frameworks/jtbd-job-stories.md.
  • Spec: frameworks/user-stories-invest.md, frameworks/ears.md, frameworks/gherkin.md.

Common heuristic across both: discovery (HMW/JTBD) → PRD (user stories) → spec (EARS for invariants, Gherkin for scenarios). Routes compose — one run can fire multiple frameworks via the orchestrator's Phase 2 fan-out.

Discovery methodology

The six-phase methodology lives at references/methodology/discovery-rounds.md. Pain-poi

Compartilhar X LinkedIn