Why LLMs Reward Expertise More Than "Good Prompting"

A widely-discussed essay by engineer Sean Goedecke makes a claim that cuts against a year of "anyone can prompt" marketing: the single biggest multiplier on LLM output quality is domain expertise, not prompting technique. The proof he offers is Terence Tao's publicly shared ChatGPT conversation about the recently discovered counterexample to the Jacobian Conjecture — a chat that reads nothing like a typical AI conversation about advanced mathematics, because Tao already knows exactly what he's looking for.

The idea has struck a nerve, generating over 550 points and 246 comments on Hacker News, and the debate is a useful stress test for anyone assuming LLMs are now a flat playing field. We covered Tao's shared transcript itself in detail when it first circulated; this piece is about the broader argument Goedecke builds on top of it.

TL;DR — what people are asking

The Terence Tao example, explained

Tao's exchange with ChatGPT about the Jacobian Conjecture counterexample — a discovery Fable 5 was independently involved in surfacing — has become a reference case for what expert-level LLM use actually looks like. Goedecke pulls out four specific patterns from the transcript:

1. Tao's messages are short and to-the-point. He responds to the gist of a multi-paragraph model output, not point-by-point.
2. The model's replies get denser, not more explanatory. By signaling expertise, Tao shifts the model from "explaining-to-amateurs" mode into "talking-to-mathematicians" mode — responses assume shared vocabulary instead of re-deriving basics.
3. He pushes back without contradicting directly. Phrases like "this looks more complex than I was hoping for" redirect the model without a flat "you're wrong," which tends to produce a more useful correction than a blunt override.
4. He steers the exploration himself. Tao rarely accepts the model's suggested next step; he proposes his own leaps and formulations, using the model to test and extend ideas he's already forming.

Goedecke's point isn't that Tao used some secret prompt template. It's that none of those four behaviors are reproducible by someone who doesn't already have Tao's mathematical intuition. You can't ask "does X work here?" about the Jacobian Conjecture unless you already understand what X and "here" refer to at a level most people never will.

Why this generalizes past mathematics

Goedecke draws the same conclusion from his own work as a programmer: familiarity with a specific codebase beats a generic understanding of software architecture. If you know a system well enough to say "but don't we already handle that with the existing retry logic?" or "can we express this in terms of the queue we already have?", you can push an AI coding assistant far harder than someone encountering the code for the first time.

This tracks with what Ethan Mollick's Wharton research on prompting has found: precise specifications, not prompt-hacking tricks, are what reliably move output quality. Domain expertise is effectively the deepest form of specification a person can provide — it's not a technique layered on top of a prompt, it's the substance the prompt is made of.

The pushback worth taking seriously

The Hacker News discussion surfaced a real counter-example. A separate AI-assisted math result — a counterexample to the Dinitz–Garg–Goemans conjecture — reportedly came from a chat that was mostly just "keep going" and "now finish with a complete unconditional counterexample," with little visible domain steering. Commenters also pointed out that OpenAI's own mathematicians reviewed and filtered the model's proposed results before publishing them — meaning expert judgment was still load-bearing, just moved to a verification step instead of the prompting step.

That's a meaningful qualifier, not a rebuttal: it suggests expertise remains necessary somewhere in the loop — either steering the generation or auditing the output — even when brute-force iteration produces a usable first draft. It's consistent with the broader shift Goedecke describes: as models get better at generating plausible candidates, the scarce human skill moves toward evaluating and directing them, not typing the first message.

How to apply this without a PhD

Most people aren't Terence Tao in any domain. But the mechanism Goedecke describes is available at every skill level, not just the top of it:

• Use your field's actual terminology. It signals to the model what register to respond in, the same way it changes what expertise level a human conversation partner assumes.
• State your specific constraints, not the general problem. "This needs to run inside our existing rate limiter" beats "make this faster" for the same reason Tao's targeted pushback beats a vague "try again."
• Push back with specifics when something looks wrong, rather than accepting the first answer — this is the actual skill transfer, more than any prompt template.
• Bring your own next step, even a tentative one, instead of asking the model "what should I do next?" — this is what separates steering a model from being steered by it.

None of this replaces years of built-up domain knowledge. But it explains why two people with the same access to the same frontier model — whether that's Claude, GPT-5.6, or Gemini — can walk away with wildly different results from the same session.

The bigger implication: expertise doesn't get obsoleted, it gets leveraged

Goedecke's closing argument is the one worth sitting with: if domain knowledge is what determines how much value someone extracts from a frontier model, then human expertise doesn't become less valuable as models get stronger — it becomes the bottleneck. The information is "in the model" already; getting it out in a useful form still takes a human who knows what a good answer looks like.

That reframes a lot of current anxiety about AI and jobs. The risk isn't that expertise stops mattering — it's that the value of expertise increasingly shows up as prompting leverage rather than as manual execution, which is a different skill to build than the one most people spent their career developing.

A concrete comparison: novice vs. expert prompting the same task

One Hacker News commenter ran an informal experiment that illustrates the mechanism cleanly. A non-technical friend asked an LLM to build a simple single-page web app. Because she lacked the vocabulary to describe what she wanted in software-engineering terms, the conversation spent over ninety minutes and dozens of messages circling through feature brainstorming without producing working code — the model kept treating vague requests as an invitation to propose more features rather than execute. A developer watching the exchange estimated that a single precisely worded prompt ("write an HTML page that does X, Y, Z") would have produced the same result the model eventually reached, in one exchange instead of ninety minutes.

That gap isn't about the model being smarter for one user than another — it's the same model both times. What changed was whether the human could translate their goal into the specific terms the model needed to act decisively instead of continuing to explore. This is the practical, non-mathematical version of what Tao demonstrates at the frontier of research: expertise isn't a nice-to-have layered on top of using an LLM well, it is what using an LLM well consists of, for any task harder than the model's default assumptions can carry on their own.

Why "just ask better questions" undersells the skill

A common response to this argument is that anyone can learn to "ask better questions" without needing years of domain training. That's half true. Some transferable habits genuinely help regardless of domain — using precise terminology, stating constraints instead of vague goals, pushing back with specifics rather than generic "try again" — and these are teachable in an afternoon.

But the deeper layer Goedecke and Tao's example both point at is harder to shortcut: knowing which follow-up question actually moves a problem forward requires having already built a mental model of the domain well enough to recognize a wrong turn. Tao doesn't ask better questions because he learned a prompting framework; he asks better questions because he can tell, from experience, that a proposed factorization "looks more complex than I was hoping for" before he's finished reading it. That kind of judgment is exactly what takes years to build in any field — mathematics, software architecture, medicine, law — and no prompting technique substitutes for it. The honest conclusion is that LLMs have made the last-mile translation of expertise into output faster and cheaper, without making the expertise itself faster or cheaper to acquire.

Related on explainx.ai

• Is AI out-thinking mathematicians, or just out-remembering them? — a memory-based account of why a well-organized context beats a clever prompt
• LLM simulation games for learning: the ChipTycoon HN debate (Aug 10) — eterm's HN comment applies this same expertise-scaling idea to learning simulations
• What is the Jacobian Conjecture counterexample, explained
• Fable 5 and the Jacobian Conjecture counterexample (Alpoge)
• Ethan Mollick: prompting is a science of specs, not tricks
• OpenAI Astra's 10 math proofs: what was actually proved?
• Will AI replace mathematicians? The Big Mathematics debate
• Claude Opus 5: top 10 use cases
• Prompt engineering vs fine-tuning vs RAG

Primary source: Sean Goedecke, "LLMs reward expertise," seangoedecke.com · Hacker News discussion thread

This article reflects the essay and discussion as published through August 4, 2026.