In today's fast-paced software development world, relying on AI agents to handle tasks is becoming common. But here's the catch: how do we ensure they understand *exactly* what we want? The bottom line is this: when you hand a vague ticket to an AI agent, it doesn't ask clarifying questions like a human would. Instead, it fills the gaps with plausible guesses, finishes the job with total confidence, and reports back: 'Implemented and verified. ✅' The problem surfaces during review, when you uncover its guesses, often against a 'definition of done' that nobody ever wrote down.

So, what does this mean for you? Acceptance criteria aren't just an optional detail; they are the foundation for effectively guiding AI agents. They must be defined *before* work starts, not after. These criteria fill every gap, preventing the AI from making incorrect assumptions or unnecessarily expanding the scope of work.

To write acceptance criteria an AI agent can actually meet, focus on answering three key questions:

1. **What will be true?** Your criteria should describe an observable *outcome*, not just an activity. For instance, 'Signing up twice with the same address is refused' is a clear outcome, while 'Add tests' merely describes an effort. An agent can always do the activity; only the outcome tells it when to stop.
2. **How will we know?** There must be clear evidence that the criterion has been met. This could be a passing test, a specific check's result, a page that opens, or the output of a command.
3. **Who decides?** The decision-maker must be named specifically, whether it's an automated CI (Continuous Integration) check or a named person. Avoid vague terms like 'the team' or 'QA,' and definitely don't leave it to 'nobody.' If you can't answer this question, it's not a criterion yet – it's just a wish.

Steer clear of vague words like 'fast,' 'clean,' 'robust,' 'properly,' or 'user-friendly.' These terms allow the AI to fill the gaps with its own standards, which tend to be quite... 'generous.' Replace them with specific numbers, clear examples, or a named judge.

If a criterion contains the word 'and,' split it into separate, distinct items. A criterion cannot be 'half met,' and this ambiguity leads to confusion. AI agents only cover the cases you write down, so write down the one you'd check first in review. Finally, to prevent unwanted scope creep, add a single line stating what this job *leaves alone*. This keeps the changes reviewable and focused.