We need user-led research on exactly how to phrase a prompt to cause this, while still avoiding the crazy guard rails.