I think it does more harm than good on recent models. The LLM has to override its system prompt to role-play, wasting context and computing cycles instead of wo

I think it does more harm than good on recent models. The LLM has to override its system prompt to role-play, wasting context and computing cycles instead of working on the task.