LLMs Reward Expertise, Not Prompt Engineering

A viral essay argues that domain expertise — not clever prompting — is what makes LLMs truly powerful, and the data backs it up.

A provocative essay by Sean Goedecke has ignited Hacker News this week, climbing to 838 points with over 348 comments. The thesis is simple but cuts against prevailing wisdom: the most important skill in working with LLMs isn't prompt engineering — it's deep domain expertise.

Goedecke uses Terence Tao's recent conversation with ChatGPT about the Jacobian Conjecture as Exhibit A. Tao doesn't write elaborate system prompts or chain-of-thought treatises. His messages are short, to-the-point, and often push back on the model's suggestions: "this looks more complex than I was hoping for." He makes his own leaps, suggests alternate formulations, and identifies what "looks weird." The model, in turn, responds in "talking-to-mathematicians" mode rather than "explaining-to-amateurs" mode.

You can't fake this. Tao isn't a better prompter — he's a better mathematician. The domain knowledge lets him pull the relevant idea out of ChatGPT's multi-paragraph response, steering it hard with questions like "does X work here?" and "given Y and Z, why A?".

Goedecke frames the conclusion bluntly: for many tasks, the human is the bottleneck, not the model. The information is "in the model" already — it takes a very smart human to pull it out.

A complementary piece by Ankur Sethi, also trending on HN this week (468 points), arrives at a similar destination from the opposite direction. Sethi refuses to let his coding assistant write files directly. Instead, he has it show proposed edits in chat, then manually retypes every line. "I want to understand every line of code that goes into this project," he writes. It's "grossly inefficient" — he estimates he's only 2x faster instead of 10x — but the payoff is a mental model of the codebase that makes him far more effective at steering the LLM on subsequent tasks.

Both essays converge on the same uncomfortable truth: LLMs are force multipliers for the already-capable, not equalizers for the unskilled. The gap between experts and novices may actually widen as models improve, because the bottleneck shifts from "can the model solve it?" to "can you recognize a good solution when you see one?"

Sources: Sean Goedecke, Ankur Sethi