I taught an AI to give up. It's the best feature I built.
Quick context: I built a system that runs structured interviews. A user has a conversation with an LLM about a topic — an internal intake, a due-diligence questionnaire, a structured assessment — and on the other side comes a structured record. Every required field is filled, partially filled, or explicitly marked as “we couldn’t determine this.” Conversational on top, schema-driven underneath. I inherited an agent. Hundreds of lines of prompt, written by hands that had since moved on. Rules referencing bugs I’d never heard of. Three different definitions of “done,” each one valid depending on which paragraph you stopped at. I tuned it the way you tune any prompt you didn’t write — adding clauses on top, hoping, never sure which one was load-bearing. ...