Most teams tune prompts the way superstitious gamblers pick numbers. Change a word, run it once, feel better, ship. Then quality drifts and nobody knows why. The fix is boring and it works: build an eval set before you touch a single prompt. Twenty to fifty real examples with a known good answer or a clear rubric. Now every prompt change is a measurement, not a mood. We caught a change that improved one flashy demo case while breaking nine ordinary ones. Without the eval set we would have shipped it and celebrated. Evals turn prompt engineering from folklore into engineering. Write the test before the fix, same as any other code.