LLMs reward expertise
Sean Goedecke argues the most important prompting skill is domain expertise, not prompt craft — using Terence Tao's ChatGPT conversation about a Jacobian Conjecture counterexample as the proof. Tao's messages are short, he signals expertise so the model drops into peer mode rather than tutor mode, he pushes back obliquely when answers look wrong, and he supplies most of the leaps himself.