This article republishes thoughts by Sean Goedecke from his blog with the author’s permission. It examines how large language models (LLMs) have broadened access to technical outputs, but argues that domain expertise remains essential to extract the best results.
How things have changed since the 2010s
In the 2010s, when you lacked a technical skill (for example, writing CSS), you typically had to rely on a skilled colleague or hope that an exact solution existed online. Today, LLMs let many people generate passable CSS and other technical artifacts; in that sense LLMs make people into generalists.
That shift can create the impression that there is little skill in working with LLMs: anyone can ask for PhD‑level mathematics, working code of variable taste, or formulaic professional writing and receive something plausible. Because many people interact with the same models, the outputs from experienced prompters can appear similar to those from first‑time users.
The key: domain knowledge
Goedecke argues this impression is mistaken. The single most important prompting skill is expertise in the domain you are querying. He points to Terence Tao’s conversation with ChatGPT about a recently discovered counterexample to the Jacobian conjecture as an instructive case. Goedecke notes he could not reach Tao’s level in that exchange even with extensive token usage: Tao’s technique rests on real mathematical understanding, not only prompt craft.
From Tao’s interaction with ChatGPT Goedecke extracts several lessons:
- Tao’s messages are short and to the point; he responds to the gist rather than addressing the model line by line.
- The model’s outputs are more concise than Goedecke gets when discussing mathematics with GPT‑5.6 Sol; by signaling expertise, Tao channels the model into a ‘talking to mathematicians’ mode rather than an ‘explaining to amateurs’ mode.
- Tao pushes back when answers seem wrong, but often in measured language (for example, noting a reply “looks more complex than I was hoping for”).
- Tao frequently proposes his own leaps and suggestions and seldom follows the model’s guidance about where to go next.
However, Goedecke emphasizes that mimicking these conversational patterns without real knowledge will not suffice. The critical difference is the expert’s ability to extract the relevant idea from a multi‑paragraph model response, suggest alternate approaches, and recognize what seems off.
Personal engineering experience
Goedecke reports similar phenomena from software engineering. If you possess a coherent theory of your codebase, you can press an LLM harder than if you lack familiarity. Knowing what a good solution should look like allows you to say, for instance, “No, we could simplify here,” or “Don’t we already do X?” or “Can we phrase this in these familiar terms?”
This ties into an earlier point he has made: system design problems are dominated by concrete specifics rather than only broad principles. Familiarity with the concrete system often matters more than a deep abstract understanding of software systems.
How indispensable is expert knowledge?
If you lack domain knowledge, you can still use an LLM to get something useful, which is not a bad outcome. But if you have domain expertise, you can extract far greater value by steering the model more precisely. Most people will rely on a mix of both approaches depending on the area.
This suggests human expertise will remain valuable even as models improve. For many tasks the bottleneck is the human, not the model, because the challenging part is communicating exactly what kind of solution is wanted. The information is often already present in the model, but it takes a very capable human to retrieve it in the desired form.
Community responses and caveats
Goedecke’s post prompted many comments on Hacker News. Some commenters shared anecdotes confirming that expertise helps or that lack of expertise hinders. Others found the claim plausible but cautioned it might be a comforting rationalization about human indispensability; Goedecke concurs that the landscape could change as models evolve.
Some pointed out that OpenAI’s mathematical prompts were inexpert and suggested expertise isn’t required. Goedecke answers that OpenAI relied on a team of expert mathematicians to check and filter model‑suggested discoveries, and that step cannot currently be skipped.
Conclusion
Goedecke’s position is that while LLMs lower barriers and make many tasks more accessible, domain expertise remains critical to obtain the best results. Experts can better steer models, detect and correct errors, and propose alternative approaches that produce higher‑quality outcomes. The community debate over the long‑term role of expertise is active, and responses indicate the relationship between human skill and model capability may continue to shift as models improve.
