LLMs reward expertise, writes engineer Sean Goedecke

In a post published on 24 July 2026, software engineer and blogger Sean Goedecke argues that the most important skill in working with large language models is not prompting tricks, but expertise in the domain you are prompting for.

Why it looks as if no skill is needed

Goedecke recalls the 2010s: if you had a technical gap — say, you could not write CSS — you had to rely on a skilled colleague or hope that the answer to your exact problem was already somewhere on the internet. Today, he writes, anyone can produce sort-of-okay CSS by delegating the task to an LLM. Models “make everybody into a generalist”, which leads many people to conclude that no skill is involved at all: everyone talks to the same models, so “skilled prompters” seem to get the same results as first-timers.

Tao’s conversation as a case study

Goedecke’s counter-example is mathematician Terence Tao’s conversation with ChatGPT about the recently discovered counterexample to the Jacobian Conjecture. He notes that Tao’s messages are short and to the point, that the model’s replies are far more concise than in his own attempts (signalling expertise shunts the model into “talking-to-mathematicians” mode), that Tao pushes back on answers that look wrong without directly contradicting (“this looks more complex than I was hoping for”), and that he almost never follows the model’s advice about where to go next.

Those habits, Goedecke stresses, cannot simply be copied: Tao’s real advantage is understanding the mathematics — pulling the relevant idea out of a multi-paragraph response, suggesting alternate formulations, noticing what “looks weird”, and asking specific questions such as “does X work here?” or “given Y and Z, why A?”.

Expertise as the bottleneck

The author reports the same effect in his own work: a good mental model of a codebase lets an engineer push the model much harder — “could it be simpler here?”, “don’t we already do X?”. Without domain knowledge you can still get something out of an LLM; with it, you can wring far more value out of the same model.

His conclusion is that human expertise will stay useful as models get stronger, and that for many tasks “the human is the bottleneck, not the model”, because the hard part is communicating exactly what kind of solution is wanted. In an edit to the post he adds that it drew many Hacker News comments, including sceptics who find the argument conveniently reassuring; to those who say OpenAI’s mathematical discoveries required no expertise, he replies that a team of expert mathematicians checked and filtered the model’s suggested results — a step he does not think can be skipped today.