LLMs reward domain expertise, not just prompting skill
Original: LLMs reward expertise
Why This Matters
Reframes the AI productivity debate: expert knowledge, not prompt tricks, determines how much value users extract from LLMs.
Software engineer Sean Goedecke argues that the most critical skill when using LLMs is domain expertise, not general prompting technique. Using mathematician Terence Tao's ChatGPT session on the Jacobian Conjecture as evidence, he contends that experts extract far more value from the same model than novices.
In a July 24, 2026 post, GitHub engineer Sean Goedecke challenges the notion that LLMs have made all users equally capable. While LLMs enable anyone to produce 'sort-of-okay' output in unfamiliar domains, Goedecke argues that domain experts consistently extract far more value from the same models. His primary evidence is a published conversation between Fields Medal-winning mathematician Terence Tao and ChatGPT about a newly discovered counterexample to the Jacobian Conjecture. Goedecke observes that Tao's messages are short and targeted, that the model's responses were notably more concise and technical than typical exchanges, and that Tao steered the conversation himself rather than following the model's suggestions. Crucially, Goedecke emphasizes this is not replicable simply by mimicking Tao's style: the underlying mechanism is deep mathematical knowledge, which allows Tao to identify errors, propose alternate formulations, and extract the relevant signal from verbose model output. Goedecke extends this to software engineering: developers with strong familiarity with a codebase can push LLMs harder, suggesting simplifications or reframings the model would not surface unprompted. His conclusion is that the human, not the model, is often the bottleneck — because communicating precisely what kind of solution is wanted requires expertise. This, he argues, means human domain knowledge remains valuable even as models continue to improve.