As models improve, it's often more effective to communicate the specific task directly rather than using metaphors or role-based prompts.