I just found out this is a technique called coercive prompting, where the system applies emotional pressure on the LLM to generate more accurate responses. It seems like a clever hack but I wonder if it scales well. Have you tried something like this?