The cheapest way I’ve found to catch wrong-shape architecture is a single question, asked at design time: what does the iteration loop look like on a typo fix?
I learned this the expensive way. We landed a content-deploy approach that, on paper, was clean — image build, registry pull, rollout. Three minutes per change. I approved it. We built it.
Then I made a typo. Three minutes later it shipped. I made another. Three minutes. I felt the cost in my body.
If I’d asked the question first, the agent and I would have caught it before a single line was written. We didn’t, so we paid for the lesson with a cleanup that retired what had just landed.
The question is cheap. The lesson, if you skip the question, is not.