LLMs reward expertise more than prompt tricks

Sean Goedecke’s essay argues that the strongest LLM users are not merely better at phrasing prompts. They know enough about the domain to notice when an answer is wrong, overcomplicated, or pointed in the wrong direction. The example is Terence Tao’s short, precise exchange with ChatGPT about a mathematical problem, where the human contribution is not verbose instruction but expert steering.

Tao does not treat the model as an oracle. He pushes back when a path looks too complex, proposes alternatives, and extracts the useful part from long responses. Those moves are only possible because he understands the mathematics. Copying the surface style of the prompts would not reproduce the result for someone without that background.

Goedecke connects the same pattern to software work. Engineers who know a codebase can ask an LLM to reuse a local pattern, simplify a proposed design, or explain why a familiar abstraction is not enough. Without that context, the user can still get something usable, but the model is harder to evaluate and harder to steer.

The Hacker News thread turned the essay into a broader discussion about verification. Commenters shared examples where expertise made AI assistance feel dramatically stronger, while unfamiliar areas made plausible output risky. The interesting claim is not that LLMs are weak, but that stronger models raise the value of human judgment.

The practical lesson is blunt: LLMs make generalist work easier, but they do not erase the advantage of knowing the field. The bottleneck often becomes communication and evaluation. The information may be latent in the model, but a skilled human is still needed to pull out the right shape of answer.