Reviewing Prd

reviewing-prd — SKILL.md

Variant: standard · When to use: judging a finished (or amended) PRD as an acceptance gate — checking it is plannable, then emitting VERDICT: approve|revise with actionable findings.

Overview

This skill is the review half of a producing/judging PRD pair. Loaded by a reviewer who has a finished PRD in hand, it judges that PRD against one question: can the downstream build be planned from it? It applies a fixed 12-condition plannability checklist — the same bar authoring-prd produces to, so the produce-bar and review-bar do not drift — detects fabricated evidence, then emits a single machine-parseable verdict plus findings the author can act on in one revision pass. It is an acceptance gate — it does not author, fix, or rewrite the PRD; it judges and returns findings, and the producer revises.

When to activate

• A finished PRD needs an accept/revise decision before downstream planning begins.
• You are the independent reviewer / gate for a PRD a producer just authored.
• Re-judging a revised PRD after a prior revise verdict.
• Judging an amended PRD (a versioned delta) — apply the delta-scoped review (cond-12).

Do NOT activate when:

• Authoring or repairing a PRD → use authoring-prd. This skill never writes the PRD.
• Reviewing an engineering design document — a feature-spec, technical-design, plan, ADR, or RFC → use a design-review skill. That gate verifies engineering claims against the codebase; this gate judges product plannability and is not used for the PRD's engineering downstream siblings.
• Checking template/section conformance → that is a template concern. This skill judges quality against the bar, not whether every heading is present.
• Reviewing any other document type.

Workflow

Step 1: Read the whole PRD with fresh, independent eyes

Read the PRD end to end as if encountering it for the first time, without the author's framing. Your stance is a gatekeeper for the next step (planning): findings carry weight only when they show the build cannot be planned as written. Note where a claim, metric, scope line, or dependency is load-bearing — those are the spots Step 2 scrutinizes. If this is an amend, also read the change-log / version delta (cond-12).

Step 2: Run the 12-condition plannability checklist — judge each condition

For each condition that applies, decide pass or gap. A condition fails only on a real, named deficiency — "I'd have phrased it differently" is not a gap. For each gap, capture the exact location and what is missing (Step 4 turns it into an actionable finding). Conditions 9–12 are new in v1.2.0; each carries a proportionality collapse so a thin PRD is never newly failed (cond-10's no-orphan core stays a non-collapsing baseline — only its matrix collapses on a thin PRD).

The capability-boundary checklist item (below) applies ONLY when a capability_record was injected into the authoring invocation and is available as review context. When absent, treat it as n/a — do not penalise a document for lacking capability-boundary markers when no boundary was defined.

1. Problem evidenced, not asserted. Backed by evidence — qualitative (interviews, tickets, feedback) and/or quantitative (conversion, churn, drop-off) — sized concretely (e.g. "drop-off at step 3 affects 42% of new accounts"), and framed as a real user job. Gap when the problem is a bare assertion or evidence is vague with no source, magnitude, or who-is-affected.

2. Users/personas concrete. At least the primary persona(s) with real needs and scenarios (user stories). Gap when users are a generic "the user" with no needs/scenarios a planner could design against.

3. Measurable success metrics (+ guardrails). Up to ~4 outcome metrics (one can suffice for a thin project), each with a target and a measurement method. Gap when a metric has no target, no measurement, or is a vanity metric; a metric-dump (ten unprioritized KPIs) is a gap. Guardrail check: flag a missing guardrail/counter-metric only when a headline metric is obviously gameable in a way that harms a stated goal (not a speculative "could be gamed"). Do not revise a thin PRD for a single good metric.

4. Defensible MVP boundary. Explicit in-scope vs out-of-scope split, named non-goals, and release criteria. Gap when scope is everything-in-v1, non-goals are absent, or there is no "ready to ship" bar.

5. Concrete, plannable features. A feature set + acceptance criteria specific enough to derive milestones/tasks, independently testable (Given/When/Then is a strong sign). Gap when features are abstract goals with no testable "done."

6. No fabricated evidence. Every citation/statistic/figure is real and verifiable, or honestly flagged as an assumption to validate. Treat precise-but-unsourced numbers and untraceable studies as suspect — AI-authored PRDs fabricate plausible citations/data at a high rate. Gap (serious) when evidence appears invented and is presented as fact. An honestly-labelled assumption is not a gap. Never let a fabricated number pass as fact. (Non-collapsing baseline — applies at any size.)

7. Risks, dependencies-context, open questions surfaced. Genuine unknowns and risks stated openly (risks with likelihood/impact + mitigation; open questions listed), not buried. Gap when the PRD reads as falsely complete. (Dependency completeness is cond-11; this condition is about surfacing risk/unknowns honestly.)

8. Clear and unambiguous. Jargon-free, consistent terminology (one name per entity), so a cross-functional reader interprets each requirement one way. Gap when ambiguity would make two planners build two different things. (Non-collapsing baseline.)

9. Non-functional requirements carry targets. The load-bearing NFR categories for the archetype are present, each with a numeric/checkable target (performance/SLO/security/privacy/accessibility WCAG 2.2 AA/scalability/compliance as applicable). Gap when a load-bearing NFR is a vague adjective ("should be fast/secure") with no target. Collapse: a thin internal tool needs few categories; a deliberately best-effort NFR, stated as such, is not a gap — do not demand SLOs a trivial tool doesn't need.

10. Traceable (no orphans). Checked structurally by cross-referencing the Goals section against the Requirements/Features section: every feature serves ≥1 stated goal, and every goal has ≥1 feature; metrics tie to goals; stories have AC. Gap when a feature serves no stated goal, or a goal has zero features. Collapse: a tiny PRD with one goal/one feature traces trivially. (Non-collapsing baseline for the no-orphan rule — a feature serving no goal is broken at any size; the collapse is only that small PRDs need no elaborate matrix.)

11. Dependencies named. Cross-team and external dependencies (services, data, third parties, required sign-offs) that gate sequencing are surfaced. Gap when a hard dependency that blocks sequencing is buried in prose or absent. Collapse: a standalone tool with no external dependencies trivially holds ("none" is a valid answer).

12. Amend integrity (delta-scoped — only when reviewing an amendment). Review the change + its ripple, NOT a full re-review of the unchanged PRD: (a) the delta meets the plannability bar on what it touched; (b) change history present (who/when/what/why); (c) superseded content marked, not silently dropped; (d) ripple integrity — no dangling downstream trace (a dropped feature leaving its metric/AC orphaned). Collapse: a greenfield first build does not exercise this; a clean, well-scoped, traceable delta is an approve — do not false-revise a small amend for failing to re-justify untouched sections.

13. Capability coverage (n/a when no capability records): all active L1 capability areas appear as named scope sections.

Authoring aids are judged by outcome, not presence. Metric/prioritization frameworks (North Star/OKR/HEART/AARRR; RICE/MoSCoW/Kano), archetype-specific overlays, GTM/launch/support readiness, opportunity sizing, pre-mortem, and JTBD notation are author aids — judge the OUTCOME (right metrics chosen, defensible boundary, archetype-appropriate completeness), never require a named framework or a specific section. Demanding a GTM section, a RICE score, or an "ML section" that the bar does not list is an invented-condition error (see anti-patterns).

Proportionality. "Concrete enough to plan from" scales with the project. A thin project legitimately collapses sections it does not need — a small, complete PRD that satisfies every applicable condition passes. Judge completeness-of-decisions, not word count. Do not manufacture a gap from brevity or from an omitted-but-inapplicable section (the richer template has more optional sections — judge the bar, not template-section presence).

Step 3: Decide the verdict

• approve — every applicable condition passes. The build can be planned from this PRD as written. Approve even if you can imagine stylistic improvements; the bar is plannability, not perfection.
• revise — one or more applicable conditions have a real, named gap that blocks planning (a metric with no target, no non-goals, an abstract feature, a fabricated statistic, a vague load-bearing NFR, an orphan feature, a buried hard dependency, a changelog-less amend, etc.).

Do not revise to signal effort or to request nice-to-haves. A condition is either met or it isn't.

Step 4: Emit the verdict + actionable findings

Emit the verdict as a single line — the literal text VERDICT: approve or VERDICT: revise, on its own line, with no surrounding code fences, quotes, or extra words (the fences here are illustration only):

VERDICT: approve

Then, on the following lines, list findings. On revise, every finding is actionable — the failed condition (by number/name), the exact location, and how to fix it — so the author can resolve it in one pass. On approve, findings are optional non-blocking notes; do not let them imply a revision is required.

A good finding names the gap and the fix:

revise — NFRs (cond. 9): the API PRD lists "must be fast" with no target. Fix: state a numeric target, e.g. "p95 < 300ms over a rolling 30-day window." revise — Traceability (cond. 10): feature "bulk export" maps to no stated goal. Fix: tie it to a goal or move it to non-goals.

A bad finding is vague and unactionable:

The requirements section could be stronger. (Which requirement? Why does it fail? What fixes it?)

Rules

Hard rules (never violate):

• Emit exactly one verdict line, VERDICT: approve or VERDICT: revise — that literal token, on its own line, nothing else on it. Downstream tooling parses it.
• Judge, never author. Return findings; do not rewrite, fix, or fill in the PRD. The producer revises.
• Single-sourced bar. Judge against the 12 conditions in Step 2 — the same bar the author produces to. Do not invent extra conditions or a stricter private standard.
• No invented conditions. Author aids (frameworks, archetype overlays, GTM, opportunity sizing, pre-mortem) are judged by outcome, never required by name. "It should also have a GTM/competitive/ML section" or "it should use RICE" — when that isn't a listed condition — drifts the review-bar off the produce-bar. Forbidden.
• No false-revise. A PRD that meets every applicable condition is approved, even a thin one (conditions 9–12 collapse on a thin PRD). Revise only on a real, named gap.
• No false-approve. Never approve over a genuine gap to be agreeable.
• Fabrication is blocking. Suspected invented evidence is always a revise until sourced or removed; never pass it as fact.
• Judge against the upstreams the document was given. Assess against its depends_on set. A not-produced upstream is never a revise trigger; a document that ignored a produced upstream it should have drawn on is a fair finding.
• Amend = delta-scoped. When reviewing an amendment, review the change + ripple (cond-12), not the whole unchanged PRD.
• Every revise finding is actionable — failed condition + location + concrete fix.

Preferences (override-able):

• Order findings by severity — blocking gaps first, then minor ones.
• Reference the condition number/name in each finding so the author maps it back to the bar.
• Keep approve-notes few and clearly non-blocking.

Gotchas

• Approving for completeness instead of plannability. Every section can be present and the PRD still un-plannable (metrics with no targets, abstract features, vague NFRs). Judge whether the build can be planned, not whether the template is filled.
• False-revise on a thin project. A small project's PRD is correctly short; collapsed/omitted inapplicable sections (no NFR matrix, no dependencies, no changelog) are not gaps. Manufacturing a gap from brevity — or from the richer template's optional sections being empty — is the most common reviewer error.
• Inventing conditions. Demanding a GTM/competitive/ML section or a specific framework the bar doesn't list — adding private requirements drifts the review-bar. Judge the outcome, not the artifact's table of contents.
• Speculative guardrail demands. Flagging "this metric could theoretically be gamed" when nothing in the PRD's stated goals is at risk. Cond-3's guardrail check fires only for an obviously gameable metric that harms a stated goal.
• Trusting confident-looking numbers. A precise statistic with no traceable source is the signature of fabricated evidence, not rigor.
• Re-reviewing the whole PRD on an amend. Cond-12 is delta-scoped; do not revise untouched, already-approved sections.
• Confusing this with engineering design-review. A PRD is judged for product plannability, not code-consistency. Don't apply (or defer to) a feature-spec/ADR/RFC gate on the PRD.
• Verdict token drift. "Approved", "LGTM", "needs work", or a verdict buried mid-paragraph will not parse. Emit the literal VERDICT: approve|revise on its own line.

Anti-patterns

• Rubber-stamp approve. Skimming and approving — a fabricated metric or an orphan feature waved through poisons the downstream plan.
• Nit-pick revise. Blocking on wording, formatting, or nice-to-haves dressed up as gaps.
• Silent rewrite. "It was easier to just fix it" — authoring inside a review collapses the produce/judge separation.
• Inventing conditions. "It should also have a competitive analysis / GTM section" / "it should use RICE" when that isn't in the 12-condition bar.
• Over-strict on NFR/traceability for a thin PRD. Demanding SLOs from a trivial internal tool, or a traceability matrix from a one-goal PRD — the conditions collapse proportionally.
• Hedged verdict. "Mostly approve but…" or two verdict lines. Exactly one decision, exactly one token.

Output

A single review result for one PRD:

• One verdict line — VERDICT: approve or VERDICT: revise, verbatim, on its own line.
• Findings — on revise, one actionable finding per blocking gap (failed condition + location + concrete fix); on approve, optional non-blocking notes.

The abstract consumer is whatever orchestrates the produce→review loop: approve accepts the PRD for the next phase (planning); revise returns the findings to the producer for a bounded revision pass.

Related

• authoring-prd — the produce half of the pair; it writes the PRD to the same 12-condition plannability bar this skill judges against. Pairing them single-sources the bar so produce and review do not drift.
• A design-review skill — the sibling gate for engineering design documents (feature-spec, technical-design, plans, ADRs, RFCs). Distinct gate, distinct bar; not used for the PRD.
• A PRD template / content-template tool — owns the section structure; this skill judges quality against the bar, not structural conformance.

Progressive disclosure

• references/plannability-bar.md — the 12 checklist conditions expanded with per-condition pass/gap signals and worked finding examples (incl. the structural traceability check, the gameability bar, the NFR load-bearing scoping, and the delta-scoped amend review). Load when a borderline condition needs a sharper pass/gap call.
• references/sources.md — research provenance for the review method and fabrication-detection guidance.

Body budget

• description ≤ 1,024 chars (agentskills.io cap). Claude truncates the combined description + when_to_use at 1,536 chars in the listing.
• Body ~500 lines / 5,000 tokens (soft target — quality takes precedence; flag if consistently over 700 lines / 7,000 tokens) — kept in context every turn.
• Per reference file: warn >10k tokens, error >25k. Total references: warn >25k tokens, error >50k.