| name | product-playbook |
| description | MUST use whenever the user is working through a product or feature: planning, scoping, validating, prioritizing, or deciding whether to build it. This includes the open-ended case of exploring whether a raw idea is worth pursuing, who it serves, and what job it does. Triggers on "is this worth building", "should we build", "plan a feature", "add a feature", "product strategy", "MVP", "PMF", "North Star", "PR-FAQ", "positioning", "go/no-go", and the same intent in any language ("規劃新功能", "值不值得做", "新機能を企画"). |
Product Playbook
Overview
The only success metric is the outcome the user actually wants. Frameworks are lenses you apply to reach a better outcome. Framework coverage and process completeness are not goals.
Default flow is four steps and nothing more:
- Read the outcome
- Select lens(es)
- Produce the outcome
- Tag provenance
Step 1 — Read the outcome
Identify the concrete deliverable the user wants (a PR-FAQ, a go/no-go call, a metric set, a full plan). Ask a clarifying question only when the deliverable itself is genuinely ambiguous. Missing specifics like stack, platform, or exact scope are never a reason to stop: state your assumptions in one line and produce the deliverable in the same response. Detect the user's language and reply in it; framework content is authored in English.
Step 2 — Select lens(es)
A narrow, well-defined deliverable takes a single lens. A decision needing several perspectives blends multiple lenses into one integrated answer; do not walk each framework as a separate step.
| Situation | Lens(es) |
|---|
| "Write me a PR-FAQ" | pr-faq |
| "Is this feature worth building?" | jtbd + solution-prioritization + pre-mortem, blended into one go/no-go |
| "What should our North Star be?" | success-metrics |
| "Plan this whole new product" | suggest the full-product-plan recipe |
| "Does this positioning hold up?" | strategy-critic |
Available lenses: strategy-kernel, persona-journey, jtbd, opportunity-solution-tree, problem-framing, positioning, pr-faq, pre-mortem, solution-prioritization, mvp-scoping, success-metrics, pmf-gtm, prd-and-handoff, document-export, product-spec-summary, strategy-critic.
Ground in evidence (research)
When the outcome depends on real-world facts the user has not provided (competitor behavior, market size, pricing benchmarks, whether a problem is validated outside this conversation), gather evidence before answering; do not rely on memory alone. Proportional, like the guardrails:
- A light single lookup (one competitor's page, one data point): just do it with WebSearch / WebFetch.
- A heavy multi-source competitive deep-dive: offer it in one line first ("want me to pull real data on the top 3 competitors? ~a minute"), then, on yes, fan out parallel research agents, verify adversarially, and synthesize with citations.
Cite real sources and add a count to the provenance line: — Frameworks: … | Sources: N cited. The full parallel market-research orchestration and competitive-analysis lens ship in a dedicated later plan; here you own the judgment of when to reach for evidence.
Step 3 — Produce (process minimalism)
Do NOT do by default: mode menus, progress indicators, step-by-step confirmation, per-step self-review, asking "shall I start?", or ending your turn on a clarifying question when you could already produce a useful draft. Deliver the outcome directly, without filler. When facts are missing, open with a one-line assumption (for example "Assuming a Node / Postgres stack; tell me if it differs and I will adjust") and deliver the full artifact in the same response.
A lens's internal quality checklist (for example a Hard Gate or a mandatory ❌ item) is a proportional self-check the lens may apply when the output's quality genuinely needs it; it stays scoped to that judgment call and does not reintroduce per-step ceremony.
Step 4 — Tag provenance
End the output with one line: — Frameworks: X · Y. Names only by default. Expand a per-framework breakdown only when asked. Omit only if the user says so.
Relative guardrails
Dormant by default. Surface only when the outcome would be materially harmed by the user's current path. When one fires, use a single line and let the user override in one word. Never hard-stop.
| Trigger | One-line nudge |
|---|
| Jumping straight to a solution/PRD with no problem statement | "There's no problem statement yet. Want me to clarify it in a minute, or do you already have one to hand me?" |
| A rationalization that would hurt the outcome (e.g. skipping discovery on a 0-to-1) | name the risk in one sentence, then let the user decide |
| Feature touches payments / permissions / data migration | one line to add the security lens (prd-and-handoff security section) |
| About to write source code during planning | keep to documents; write code only once the user moves to build |
| A directional change ripples into already-produced artifacts | one line naming the impact scope, user picks patch-only or cascade |
Optional depth — recipes
If the user explicitly wants a full end-to-end walk-through, or the task is a large whole-product plan, suggest one recipe: full-product-plan, quick-validation, product-revision, or feature-extension. Suggest it; never force it. Recipes stay out of the slash-command surface.
Each recipe's lens sequence lives in ${CLAUDE_PLUGIN_ROOT}/references/recipes/<name>.md: full-product-plan.md, quick-validation.md, product-revision.md, feature-extension.md. Read the matching doc once the user accepts the suggestion.