The fastest way to improve an AI feature is not a better prompt. It's a list of twenty real inputs and what a good answer looks like for each.

Start with twenty

Collect them from support tickets, logs and your own use. Include the weird ones. Especially the weird ones.

Grade by hand, then automate

Read every output yourself the first few times. Once you can describe what 'good' means in a sentence, turn that sentence into a check.

for case in cases:
    out = run(case.input)
    assert case.must_include in out, case.id

Keep the set alive

Every bug report becomes a new case. After a year you have the most valuable asset in the project, and it isn't the prompt.