Domain Knowledge Beats Prompt Tricks: An Uncomfortable Experiment

I ran an uncomfortable experiment last quarter: a friend who collects prompt techniques and a colleague who simply knows our domain deeply were given the same AI tools and the same problem. The prompt collector produced more elegant prompts. The domain expert produced the useful answer. The gap between those two facts rearranged how I think about AI skill.

The mechanism was visible in the outputs. The elegant prompts were generically well-structured: role, context, constraints, format. What they could not contain was the knowledge of which constraints actually matter in our domain, which failure modes are common, and which part of the 'correct' answer is actually wrong in practice. The domain expert's messy prompts carried all of that.

This explains a pattern I kept seeing in the wild: experts with casual AI habits outperforming enthusiasts with elaborate systems. It was never about the prompting. The expert's four-sentence prompt contains twenty years of implicit specification, and specification was always the hard part. Prompt technique optimizes the easy 20 percent.

The uncomfortable implication for the prompt-engineering industry: most of it is teaching people to optimize the part that was never the bottleneck. There are real techniques (structured outputs, few-shot examples, iteration discipline), but they are the last 10 percent, and they are worthless without the domain knowledge that tells you what to ask for.

My own practice changed accordingly: I stopped hunting prompt tricks and started writing domain notes, the 'what matters and what usually goes wrong' documents that feed my prompts. The prompts got shorter and the outputs got better, which is the experiment's result in one sentence: the knowledge was the prompt all along.