The core claim: domain expertise makes you dramatically better at extracting value from LLMs, not just slightly better. The author illustrates this through Terence Tao's conversation with ChatGPT about mathematics—Tao's ability to recognize when outputs "look weird," suggest reformulations, and reject bad directions comes directly from deep mathematical knowledge. The same applies to programming: familiarity with your codebase lets you steer an LLM toward genuinely good solutions rather than just accepting whatever it generates. The actionable insight: if you know your domain well, you can push back hard ("can we express this simpler?") and validate suggestions critically. If you don't, you're stuck accepting whatever the model produces. This inverts the common assumption that LLMs democratize expertise—they actually amplify existing expertise because the human becomes the bottleneck, not the model. The information is already in the model; the hard part is knowing what to ask for and recognizing when it's wrong. The source doesn't detail measurable productivity differences or show whether this holds across all domains equally.