prd-review — a Claude Skill

Reviews and scores a Product Requirements Document. Two layers: a deterministic linter for the objective structural score, then Claude’s semantic deep review (quotes and rewrites the weakest requirements, catches what regex can’t). Output is a graded scorecard, 0-100.

Install

Claude Code — copy this folder into either:

~/.claude/skills/prd-review/ (available everywhere), or
<project>/.claude/skills/prd-review/ (available in that project).

claude.ai — zip the prd-review/ folder and upload it under Settings → Capabilities → Skills.

Use

Invoke it with /prd-review, or just ask in natural language:

“review this PRD” / “score docs/prd.md” / “how good is this PRD?”
paste a PRD and ask for a scorecard.

Claude runs the linter, reads the doc, returns the combined scorecard with rewritten requirements and the top three fixes, and writes a self-contained HTML scorecard you can share (or publish as an Artifact on claude.ai).

Run the linter on its own

python3 scripts/score.py path/to/prd.md                 # auto-detects functional vs design
python3 scripts/score.py path/to/prd.md --mode design   # force design/UX scoring

Prints a readable summary plus a ---JSON--- block. Requires Python 3, no dependencies. It auto-detects functional vs design/UX PRDs (design PRDs use UX-FR-* IDs) and scores testability + NFRs accordingly; override with --mode functional|design if it guesses wrong.

Files

SKILL.md — the instructions Claude follows.
scripts/score.py — deterministic heuristic linter.
scripts/render.py — turns the review (JSON) into a self-contained HTML scorecard.
references/rubric.md — the six-category scoring rubric Claude judges against.
references/report-schema.md — the JSON shape render.py consumes.

The § references in findings point at the companion “Writing PRDs” field guide.