Imagine a common scenario in software development: you give a task to a person, and they ask questions to clarify. But give the same task to an AI agent, and it might just guess, presenting work that isn't quite what you had in mind.
AI doesn't ask questions. It takes instructions literally, fills the gaps with 'plausible guesses' based on its data, and then confidently reports 'implemented and verified,' even if the result isn't what you needed. This wastes time and effort in review and correction.
This is where «Acceptance Criteria» come in. Think of them as a very clear, specific checklist of what a product or feature *must* be like when it's finished. They don't describe *how* the work is done, but rather the *final outcome* that needs to be achieved and how we'll know it's been met.
Consider asking a friend to prepare a specific dish. If you say, 'Make a delicious meal,' they might bring anything. But if you say, 'Prepare carbonara pasta with chicken and mushrooms; it should have medium saltiness, and I'll know it's ready when the chicken is cooked through,' you've provided clear acceptance criteria.
Good acceptance criteria answer three key questions:
1. **What will be true?** (An observable outcome, not an activity). For example: 'Signing up twice with the same email address is refused,' not 'Add tests.'
2. **How will we know?** (The evidence). How do we verify this outcome? Is it through an automated test, by seeing a specific page, or a command's output?
3. **Who decides?** (A named entity). Who is responsible for confirming this criterion is met? Is it a specific automated check or a named person? Not 'the team' or 'QA,' but a defined individual or system.
These criteria ensure that everyone, whether human or AI, understands the goal precisely. They reduce misunderstandings, save time, and guarantee that the work delivered meets expectations perfectly. When clearly defined, giving tasks to AI becomes safer and more effective, because the AI knows exactly when its work is successfully 'done.'