@scottfgray Very recognizable. I played around with some AI Agent frameworks, and under the hood those are built on hard-coded prompts that attempt to make an LLM produce structured output. Because of the statistical nature of LLM's this is hopelessly unreliable so the developers try to increase the odds of getting usable output by adding encouragement ("You are very good at your job") or even threats ("Give your best answer, your job depends on it!") to the prompts.