Before an AI feature meets real users.

Write down the job the feature should do in a sentence a user would recognize. Then collect examples of a useful result, an incomplete result and a request the feature should decline or hand back. Those examples create a shared standard for evaluating changes.

Give failure paths the same attention as successful responses. Consider missing information, unavailable services and a task that takes longer than expected. Decide what the user should see, when to stop and how they can recover or continue manually.

Review what the feature is allowed to read or change, and where a person should confirm an action. Keep enough diagnostic information to understand problems while respecting the data involved. Revisit the examples when prompts, models or connected services change.