Every way an LLM got my week wrong, and the six rules that fixed it

Week one it booked four calls back to back, because nobody had told it that a call has a cost after it ends. Week two it protected a focus block by moving a doctor's appointment. Week three it learned to ask one question first, and asking turned out to be the entire product.

I read the whole log on screen: thirty days, forty-one interventions, every one of them with the correction I typed at the time. Six of those corrections became rules in Steward. Every single one is a sentence I had been enforcing in my head for years and had never written down.

That is the actual finding, and it is not about models. The agent was not bad at scheduling. It was working from a specification that existed nowhere outside my skull, and the month I spent correcting it was the first time that specification got written down.