Berating an LLM's response as a way to improve output has worked for at least some subset of LLMs and outputs. I have a coworker who told me that "hack" anyways.