Most of us learned to write specs from templates and examples, but I've never seen anyone taught how to red-team one: to hunt for the missing edge case, the metric with no baseline, the acceptance criteria that aren't testable. Those gaps usually surface in eng review, a sprint too late.
The AI that reviews your PRD, not just writes it. Take a spec you already shipped and thought was finished. Thinkr reads it like a senior reviewer: 11 named passes, the full rubric published so you can audit every check. Findings come back classified blocker, major, or minor and land as comments on the exact line they concern, each with a suggested rewrite you can accept inline. It ends on a deterministic score, per dimension, computed from the findings rather than a model's gut feel.
