Your expertise is the real prompt

Everyone is hunting for the right prompt. The thing that actually changes what a chatbot gives back is sitting in their own head.

For two years the conversation about large language models — the AI systems behind ChatGPT, Claude, and Gemini — has orbited prompt engineering. Buy the course, copy the template, chain the thought. Sean Goedecke, a working software engineer, argues in a widely-discussed blog post that the whole industry is aimed at the wrong target. The skill that actually moves the output is domain expertise, not prompt cleverness (LLMs reward expertise).

The mechanism is uncomfortable. Large language models turn everyone into a generalist: a non-designer can ship passable CSS, a non-writer can draft a memo, a non-programmer can scaffold an app. The same model serves all of them, and it has no reliable way to tell which user actually knows their field. So it defaults to the same register for everyone: "explaining-to-amateurs." That is the level most people are getting back, and it is the level the prompt-craft industry has been trying to dress up.

Goedecke's exhibit is not a developer. It is Terence Tao, a Fields-medal mathematician and one of the most cited living researchers, in a conversation that Goedecke describes with ChatGPT about a recent surprise: a counterexample to the Jacobian Conjecture, a long-standing open problem in algebra that most mathematicians expected to stay unsolved. Goedecke's claim is not that he could reach Tao's output even with unlimited tries. It is that watching Tao work shows what he describes as the four moves most users never make.

Goedecke describes Tao's approach as having four visible hallmarks. First, his messages are short and respond to the gist, not the literal text. Second, his outputs are concise because the register signals the model into "talking-to-mathematicians" mode rather than the default pedagogy mode. Third, he pushes back on answers that look wrong without contradicting them directly — "this looks more complex than I was hoping for," the kind of soft refusal that gives the model room to correct itself. Fourth, he proposes his own next steps instead of asking the model what to do next. Each move is a way of telling the model what kind of conversation it is in.

The caveat cuts both ways. Those four techniques are not portable. Pulling the right idea out of a multi-paragraph response, suggesting an alternate formulation, and noticing when something "looks weird" — that is mathematics, not prompting. A reader cannot adopt the moves without doing the underlying work, which is exactly Goedecke's point. The prompt industry is selling the wrapper, not the thing inside it.

The same shape shows up in software. A working theory of a codebase — what the modules are for, what the failure modes look like, where the seams sit — lets a developer push a large language model much harder than a novice can, because the developer already carries a sense of what a good answer should look like. The model fills in the typing; the human supplies the judgment. Strip the judgment and the same prompts return the same "explaining-to-amateurs" sludge.

Not everyone is buying it. In the Hacker News thread under the post, several commenters push back that the "be an expert" advice only addresses the cohort who learned to code by copying Stack Overflow answers and never grew up, and that LLM delegation is a natural extension of that pattern (HN discussion). Others note that a system trained on a million papers by Tao would partly be pattern-matching the name, not just responding to expertise — a real possibility the post does not rule out. Both are worth taking seriously, and they narrow the claim. The thesis is not that expertise is magic. It is that, in the median case, the user's own knowledge is the largest unused input the model has to work with, and almost no advice column is telling people to invest in it.

For a worker, student, or creator who already knows their field, the practical version of the argument is plain. The next time an AI assistant hands back a draft, a diagnosis, or a proof, the most productive move is rarely a cleverer prompt. It is the same move a domain expert would make with a junior colleague: read it, push back on the part that looks wrong, and propose the next step yourself. The model will meet the register the user sets, and the user sets it by what they already know.