I built Prompt Coach, an AI tool that scores a rough prompt, diagnoses the gaps, and walks the user through building a stronger version by hand. Then I tested it myself, the same way a real user would, and found a bug that most people would only catch by accident.
The obvious fix would have been to just cut the output short. That treats the symptom, not the cause, so I traced the actual instruction the tool was sending the AI instead. One ambiguous line was the real problem. Fixed that, then verified it held across four separate test runs on two unrelated topics before calling it done.
That loop, build it, use it, diagnose what broke, fix the actual cause, verify more than once, used to be split across three different people at three different desks. Now it happens in one sitting.
This video walks through that process. Built as a deck, animated, then scored and finished in Canva.