Last week I posted that I had no code, just the work that makes the code possible. The PRD, the prompt spec, the architecture doc, the build brief for Kiro. I went into this week thinking I had every decision pre-made.

Then I started building.

By Block 2, real testing surfaced a phrase the model was using that no court employee would say. "Strip identifiers" sounds reasonable to a developer. To a court clerk it sounds like nothing, opaque and technical, the kind of thing you'd skip past in a training. Not a bug, exactly. Not in the PRD either. But noticeable.

By Block 3, I was flagging contrast questions, helper text microcopy, a disabled state that needed verifying. None of these were in the original scope. None of them were stop-the-build issues. All of them were real.

By Block 4, I was holding a list of polish items in my head while running an active build. Around 11pm I asked Claude how it was tracking everything. The answer was honest: it wasn't, in any reliable way.…