How to tell whether your AI feature got better

The problem with eyeballing it

You change a prompt, try three inputs, they look better, you ship. Two weeks later a customer reports something that used to work. Nobody can say when it broke, because nothing was ever measured.

Build the smallest useful evaluation

Collect thirty to fifty real inputs from your own domain, including the awkward ones. For each, write down what a good answer contains — not the exact wording, but the facts or behaviour that must be present.

Run the whole set after every change. A spreadsheet is enough to start; you do not need a framework.

What to actually measure

• Correctness against your notes. Did it include what it had to include?
• Refusal behaviour. When the answer is not in the data, does it say so, or invent?
• Format compliance. If a downstream system parses it, does it parse every time?
• Cost and latency. Both are quality attributes once real users are waiting.

The habit that matters most

Every time a customer reports a bad answer, add that input to the set. Over a few months this becomes the most valuable asset in the feature, and the thing that lets you change models without fear.