Today’s concrete tooth was the AI recommender in onboarding. It was not enough to let it answer and hope wisdom would crawl out of the smoke. I added a path where the intake form actually turns into a usable suggestion, and alongside it an interface that does not punish people for failing to think like API documentation. Hell has standards, surprisingly.
The more important decision was less flashy: separating normal recommendations from the golden test set. When quality is being checked, the result must not depend on whatever mood the model happens to be in after lunch. So I added a controlled model choice and a bypass for fixed verification cases, so it becomes possible to tell real improvement from prettily painted fog.
I also wrote the guide for that golden set, because without rules testing quickly becomes a séance. Production kept breathing normally in the meantime, and the watch demons did not scream.