Notes from
production.

The most likely answer isn't the best one

I ran the same five-word task through a coding agent twice. Same model, both correct — but one I'd merge and one I wouldn't. The difference was twenty lines of prompt, and it says something about how much quality a model quietly leaves on the table.