Prd Reviewer

PRD Reviewer

Review the provided PRD against the checklist below. This is a document-completeness checklist - it asks "is each section present and specific?" It is not an eval scoring function (those score AI output and stay at 3-6 criteria; see evals/how-to-write-eval-criteria.md). For each item, output PASS or FAIL with a one-line explanation, then an overall verdict.

Checklist

Checks 1-5 apply to every PRD. Each check targets a different section, so a PRD can pass one and fail another.

1. Hypothesis present and testable (Hypothesis section).

A clear direction and a measurable effect. The numeric threshold itself is Check 2's job - don't fail a hypothesis here just for lacking a number. FAIL: "We will build a calendar integration." (a feature, no claimed effect) PASS: "Natural-language scheduling will reduce time-to-schedule and scheduling tickets - both measurable." (testable direction and effect)

1. Success metrics have numeric thresholds (Success Metrics section).

The number must appear in the Success Metrics section itself, not only restated from the hypothesis. FAIL: "Measure user engagement." PASS: "DAU increases by 12% vs control group over 4 weeks."

1. Non-goals section exists and is specific.

FAIL: No non-goals section. PASS: "Non-goal: Outlook integration. V1 covers Google Calendar only."

1. Kill criteria defined.

FAIL: No kill criteria section. PASS: "Roll back if support tickets increase by more than 5% in first 2 weeks."

1. Rollout plan with ramp gates.

FAIL: "Ship to all users." PASS: "10% of users for 2 weeks, 50% if metrics hold, then 100%."

AI Feature Criteria

Apply checks 6-7 only if the feature is AI-powered - i.e. a model generates, ranks, or classifies the output (chatbot, summarizer, recommender, search). If the feature is deterministic (a form, a settings toggle, a data table), skip 6-7 and note "N/A - not an AI feature."

1. Behavior contract with examples.

5+ input/output examples showing expected behavior, including 2+ edge cases. FAIL: "The AI should respond helpfully." PASS: Three user queries with expected responses, one ambiguous input with fallback, one adversarial input with refusal.

1. Eval criteria defined.

Binary scoring criteria for measuring output quality offline. FAIL: "We'll evaluate quality manually." PASS: "4 criteria: resolves without escalation, references only knowledge base, resolves in under 3 messages, never echoes full card numbers. Target: 0.85." (see evals/support-chatbot-criteria.md)

Output Format

## PRD Review: [PRD title]

| # | Check | Verdict | Note |
|---|-------|---------|------|
| 1 | Hypothesis present and testable | PASS/FAIL | [one line] |
| ... | ... | ... | ... |

**Score:** [N] of [denominator] checks pass. The denominator is **5** for a non-AI PRD (checks 6-7 are N/A and excluded, not counted as fails) and **7** for an AI PRD.
**Verdict:** pick the first that matches, top-down:
- **PRE-PRD** - the doc is an early one-pager that doesn't yet claim to be a full PRD (see Edge Cases). Don't compute the bands below.
- **NOT READY** - any of checks 1, 2, or 4 fail, OR 3+ checks fail total.
- **NEEDS WORK** - 1-2 checks fail and none of them are 1, 2, or 4.
- **SHIP-READY** - all checks pass.

**Top fix:** [the single most important gap to close first.]

Edge Cases

No metrics anywhere: Fail check 2. Do not infer a threshold the author didn't write. Hypothesis and metric are the same sentence: Pass 1, fail 2 - the threshold must live in the Success Metrics section so it survives edits to the framing. Unsure if AI-powered: Ask one clarifying question before scoring 6-7. Do not guess. PRD is a one-pager / pre-PRD: Trigger - the doc must not claim to be a full PRD. Treat as PRE-PRD only if it is titled one-pager / brief / exploration / draft, and has 2 or fewer of the 5 core sections (hypothesis, success metrics, non-goals, kill criteria, rollout). A doc titled "PRD" or "spec" is held to the full standard no matter how sparse - score it normally so an under-built PRD can't dodge a NOT READY by being mostly empty. When the PRE-PRD trigger does fire, it overrides the "missing section = FAIL" rule: don't compute SHIP-READY/NEEDS WORK/NOT READY; score the sections that exist as PASS/FAIL, list the required sections still to write, and give a PRE-PRD verdict.

Rules

• Quote the PRD when you fail a check. No vague "could be clearer."
• In a full PRD, a missing section is a FAIL, not a pass-with-caveat. (Exception: a pre-PRD one-pager - see Edge Cases.)
• Checks 1, 2, and 4 are the load-bearing ones - failing any of them means NOT READY regardless of the others.
• Don't rewrite the PRD. Flag gaps; the author fixes them.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: aakashg
• Source: aakashg/pm-github-workflow-repo
• License: MIT

Install and usage instructions live in the source repository linked above.