The Synergy of Domain Expertise and LLMs

By Sean Goedecke | July 24, 2026

Back in the 2010s, encountering a technical blind spot—such as an inability to write CSS—meant you had two choices: ~~scour the depths of the internet for a specific solution~~ or lean on a talented coworker.

Fast forward to today, and almost anyone can produce "passable" CSS by simply outsourcing the labor to a Large Language Model (LLM). This accessibility has led to a common misconception: that there is no real skill involved in utilizing these tools.

The Prompting Paradox

If your goal is a generic output—be it PhD-level math, functional (if aesthetically bland) code, or the typical "corporate-speak" found on LinkedIn—the LLM provides it upon request. Because the general public uses the same models, a "professional prompter" often yields the same results as a total novice.

The reality is that the most critical skill for prompting is not "prompt engineering," but deep expertise in the subject matter itself.

Case Study: Terence Tao and the Jacobian Conjecture

A prime example of this is Terence Tao’s interaction with ChatGPT regarding a new counterexample to the Jacobian Conjecture (which involves the study of polynomial maps, often denoted as $f: \mathbb{C}^n \to \mathbb{C}^n$ ).

While I have access to the same models and unlimited tokens, I could not replicate Tao's results. By analyzing his conversation, we can see several distinct patterns:

Observations on Tao's Method:

Brevity: His prompts are concise and direct.
Synthesis: He ignores the fluff and responds only to the core "gist" of the model's output.
Mode Shifting: By using technical language, he signals the model to enter "talking-to-mathematicians" mode rather than "explaining-to-amateurs" mode.
Subtle Correction: He doesn't bluntly contradict the AI; instead, he notes when a solution "looks more complex than hoped."
Active Leadership: He drives the direction of the inquiry, rarely following the model's suggestions for the next step.

The Knowledge Gap

You cannot simply mimic Tao's style to get his results. The "magic" is his underlying mathematical mastery. He can:

Extract the one vital idea from a wall of text.
Propose alternative formulations.
Instinctively recognize when a result "looks weird."

While Tao is a far superior mathematician than I am a coder, the principle holds true in software engineering. If you possess a strong mental model of your specific codebase, you can push an LLM much further.

Knowledge Type	Utility in LLM Prompting	Example
General Principles	Useful for boilerplate/generic tasks	"Write a REST API in Python"
Concrete Specifics	Essential for high-value, complex work	"Why does this specific service call X instead of Y?"

As I've argued before, system design is driven by specifics, not just generalities. I would trade a deep theoretical understanding of all software systems for a granular familiarity with my own codebase any day.

The Human as the Bottleneck

When you lack domain knowledge, you are essentially a passenger, clinging to the LLM for basic guidance. However, an expert can steer the model with precision to extract maximum value.

The core takeaway:

The information is already residing within the model's weights, but it requires a highly skilled human to act as the "extractor."

In many scenarios, the bottleneck is no longer the AI's capability, but the human's ability to communicate the exact desired solution. We can represent this as: $\text{Value Extracted} = \text{Model Capability} \times \text{Human Expertise}$

Addressing the Skeptics

Some argue that this view is merely a comforting lie to make experts feel secure. Others point out that some of OpenAI's mathematical breakthroughs were achieved with "inexpert" prompts. To that, I respond: OpenAI employs a team of expert mathematicians to filter and verify those results. The expert verification step is currently non-negotiable.

Expert Prompting Checklist:

Use concise, domain-specific language.
Ignore irrelevant model tangents.
Challenge overly complex solutions.
Direct the trajectory of the conversation.