You've written PRDs before. Requirements, acceptance criteria, user stories, edge cases. Standard stuff.

For AI features, that standard spec is missing the most important section: How do we know if it's good enough?

Traditional spec:

Feature: Auto-categorize support tickets
Requirement: Tickets should be categorized into the correct bucket
Acceptance criteria: Categories are assigned correctly

This spec is useless. "Correctly" is doing all the heavy lifting and it means nothing. Correctly how often? 70%? 95%? 99.9%? What happens when it's wrong? Who decides what "correct" even means when reasonable people would disagree?