In practice
Good criteria are testable. “Loads quickly” is vague. “Renders in under 2 seconds on a 3G connection” can be checked.
They don't describe the design or every possible edge case. They describe the result that must be true.
Why it matters for product builders
Without clear criteria, “done” is an opinion. You discover the disagreement at the end, when changing the work costs more.
What changes in AI-first products
Acceptance criteria now guide the build as well as the review. An agent can test its own work against a precise condition. It can't test “make it fast.”
See how Annsa does it: Done When →