Build the evaluation harness before you tune the prompt

The fastest way to stall an AI project is to iterate on prompts without a way to tell whether a change helped. Teams end up defending changes with anecdotes, and nobody can approve a release because nobody can describe what good looks like.

A workable minimum

• A representative case set drawn from real questions, agreed with the business
• Expected characteristics per case — not necessarily an exact string, but what a correct answer must contain
• Automated scoring that runs on every change
• A threshold that a release must clear, decided before the pressure to ship arrives

This does not need to be elaborate. A few dozen well-chosen cases will surface most regressions, and having them changes the conversation from preference to evidence. It also gives the business a legible way to sign off — which is usually the real blocker on the path from pilot to production.