Using AI keeps showing me the same thing: what I want from the result changes less often than what I can give it.
"We want a short message." "We want references." "We want an evaluation."
These requirements are pretty stable. What changes is the input. What do we have? How much can we provide? Is it complete, current, useful? Sometimes we have proper documentation. Sometimes we have three notes and a screenshot.
There is a small distinction here. The message or evaluation is the output. What it helps us understand or decide is the outcome. A short answer can meet the length requirement and still be useless. References need to support the claims they accompany.
For example, imagine asking for a short comparison of two software options, with references, to decide which one to try. Today we can supply documentation for both. Tomorrow we have detailed information about one and an old description of the other. The desired outcome stays the same. The evidence available to reach it does not.
The second comparison should make that gap visible. Maybe it can compare a few things. Maybe it needs one missing document before recommending anything. Producing the same confident answer in both cases would hide the problem.
That is why I would design for uncertain inputs from the start. Keep the goal and output requirements explicit. Decide which information is essential and what should happen when it is missing. Test with sparse, conflicting, and excessive inputs, as well as the clean example.
Goals can change too. But a workflow built around today's perfect input gives us another thing to rebuild when tomorrow's input arrives.