Product discovery

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-point excavation → landscape research (three vectors) → multi-direction brainstorm → vocabulary mapping → data-model synthesis → delivery. The discovery-round route fans out against these phases; the discovery-lead role owns the cadence.

Ecosystem integration

erpaval hard-deps on two role files as a public API:

• references/roles/hmw-framer.md — erpaval's CL-RIGOR fuzzy classifier spawns this role directly.
• references/roles/ears-specifier.md — erpaval's CL-RIGOR contract-unclear classifier spawns this role directly.

Renames or scope changes to either role are breaking for erpaval. The orchestrator's ecosystem section in references/orchestrator.md documents the contract. Coordinate breaking changes in the same commit.

When NOT to use this skill

• Business and product strategy (Rumelt kernel, Wardley map, PR-FAQ-as-leadership-artifact) — use product-strategy.
• Long-form narrative composition — strategy and discovery here produce the thinking; hand the artifacts to whatever narrative-writing workflow your team uses.
• Implementation-only tasks — use erpaval for autonomous coding; this skill defines the spec, not the build.

Anti-patterns

• Don't skip framing. A Phase 2 fan-out without a frozen framing.md or Intent Profile produces divergent agent outputs the synthesizer can't reconcile.
• Don't run more than 2 critic-revise rounds. After 2, surface findings inline. Critic loops are a stop-energy pattern, not a quality gate.
• Don't fabricate job stories or HMWs to meet a count. Three real job stories beat five fake ones; the jtbd-interviewer and hmw-framer roles will flag gaps rather than invent.
• Don't rewrite across roles. Each role owns a slice. If a sibling role's output is wrong, flag it in the synthesizer's log — don't silently rewrite.
• Don't bypass the write-protocol. Batched-writes-at-the-end lose work on timeout. Edit → read → edit is the discipline every role follows.
• Don't confuse upstream and downstream. Job stories are upstream; user stories are downstream. HMW is upstream; EARS/Gherkin are downstream. Using the wrong one for the current phase produces artifacts that feel right but don't land.
• Don't delete or rename hmw-framer.md / ears-specifier.md without updating erpaval. Those are public API. Alias-then-delete for any rename; coordinate breaking changes.