In my 3 years working with large language models (LLMs), I’ve learned this fundamental truth:
An LLM is only as smart as the human-expert frameworks provided in its prompt.
Let's look at long-form writing as an example. For a long time, many AI researchers thought generating long texts with LLMs was impossible due to short context lengths. If you tried to get an LLM to continue writing, it would start repeating itself. The AI community struggled with this limitation for years and largely viewed long-form generation as an unsolved problem.
One day, a creative writer on our team figured out a solution - prompt chaining. Just like human writers outline before drafting, she had the LLM follow an outline-draft process.
It worked remarkably well, proving that mimicking human frameworks is key. LLMs can match human expertise, but only if the prompt provides the right mental scaffolding. Without framing the task as an expert would, the LLM flounders.
This lesson guides my work - whether it’s writing, education, or any domain. Whether it's writing a novel or diagnosing a patient, an LLM can only excel when given the right expert prompt. Deeply understand the human expert’s thought process, instill their knowledge in the prompt, and an LLM’s potential is limitless.