Pulse · 2026-07-18 · Analysis
How to write acceptance criteria for AI drafts
Write acceptance criteria for AI drafts as good-looks-like and bad-looks-like bullets before you generate, then judge the output against them.
AI assistance: Human-authored craft note; no generative filler.
Direct answer: Acceptance criteria for AI drafts are short bullets that define good and bad before you generate. You judge the draft against those bullets instead of vibes.
What this means at work
Without criteria, every revision becomes taste. With criteria, you can accept, reject, or request a targeted fix in one pass. This is how you keep enthusiasm from turning into endless polish.
Example for a client status email: Good looks like concrete dates and owners. Bad looks like hype, vague progress, or invented policy.
Do this next
- Before prompting, write three “good looks like” bullets and two “bad looks like” bullets.
- Put those bullets in the prompt as the quality bar.
- Generate once. Score the draft against the bullets only.
- Ask for a fix on the failing bullets, not a full rewrite.
Criteria turn enthusiasm into a finish line. Without them, you will polish forever and still not know whether the draft is done.
FAQ
How detailed should criteria be?
Enough that a colleague would score the draft the same way you would. Not a novel.
Can the model invent the criteria?
It can propose them. You own the final bar before generate.
See also
- Field Guide: Ask better
- Field Guide: Start here
- Pulse: What is a work-ready AI prompt?
Harder questions live in Challenges.