Why domain expertise is the real prompting skill for LLMs

The common assumption is that LLMs flatten skill differences: since everyone queries the same models, a first-time user should get the same output as a seasoned ‘prompt engineer.’ Goedecke argues the opposite. The decisive skill isn’t clever prompting technique but deep knowledge of the subject you’re prompting about. He points to Terence Tao’s ChatGPT exchange on a counterexample to the Jacobian Conjecture as evidence — a conversation the author says he could never reproduce no matter how many tokens he spent, because he lacks the mathematics to do so.

What makes an expert effective isn’t a set of tricks, though the surface behaviors are visible: terse messages that respond only to the gist, gentle pushback (‘this looks more complex than I was hoping’) rather than flat contradiction, and a willingness to steer the model’s direction instead of following its suggestions. Signaling expertise even shifts the model into a more concise, peer-to-peer register. But those habits work only because the expert already knows what a good answer looks like — which idea to extract from a sprawling response, which alternative to propose, what ‘looks weird.’ The author sees the same dynamic in software: a strong mental model of a codebase lets you press the LLM far harder than generic knowledge of software systems ever could, because you can ask pointed, specific questions.

The broader implication is that human expertise won’t be made obsolete as models improve — it becomes the bottleneck. The knowledge is often already latent in the model; the hard part is communicating precisely what solution you want and recognizing it when it appears. Since most people are expert in some domains and novices in others, the practical takeaway is that we’ll all oscillate between clinging to the model for areas we don’t understand and aggressively steering it in the ones we do.