The Vocabulary Bottleneck

Why experts get better work from the same AI

I had a conversation with a non-engineer friend recently about what makes me, an engineer, better at getting AI to build software than him. In theory, we’re both using the same model, and so we should both be able to get the same outputs out of it. But for some reason, the software my AI built was consistently more robust and just plain worked better.

It doesn’t make any sense.

So I asked him to use his agent while I watched. He booted up his Codex, switched to a new task, and started writing his first prompt:

“Build me a website for my pickup soccer group. People should be able to create an account, schedule games, join games, and see who else is playing. Make it clean, modern, and easy to use on a phone. Use whatever technologies you think are best.”

Almost immediately, it clicked. That was nothing like how I would have prompted it. His prompt was entirely about the outcome of the task, what the final product would look like. It was entirely unopinionated on how it should get there.

Now, if I were to ask my Codex to build the same product, I would’ve used way more engineering jargon. I would’ve specified how auth should work, what platforms people should be able to create accounts with, where the website should be hosted, what frameworks to use, and way more.

For something this simple, it didn’t end up making a difference. Both our websites worked pretty well. But then again, this is a product that any second-year computer science student should be able to build, so that’s not surprising.

But what about when you’re building at the frontier? Then, the gap becomes way wider. To understand why that’s the case, we first need to understand how these new models are being trained.

What goes into frontier models?

The first models that were released earlier this decade all had (more or less) the same training data. Model companies scraped everything into the internet, pumped it into a transformer, and trained it to predict the next word. Those models were slightly better than mediocre at just about everything, which funnily enough is how I would’ve described most things on the internet.

At some point, that way of sourcing data hit a wall. This was amplified with the release of reasoning models, as there’s only so many examples of high quality text showing how experts reason through difficult problems. Model companies needed a way to get a lot of transcripts of the best in the world’s reasoning while they work.

So, they started creating that data. They, with the help of data startups like Mercor, started hiring skilled professionals to narrate their thoughts while they do their job. Those narrations become the data behind model reasoning.

The problem is that experts use jargon. An engineer narrating their thought process is never going to talk about how they’re “making accounts work”. They talk about setting up auth, figuring out email recovery, making sure passwords are hashed, and so on. Similarly, a designer narrating their thought process doesn’t talk about “making things look clean”. They talk about hierarchy, information architecture, spacing systems, interaction states, accessibility, and responsive behavior.

This strategy worked. The models got better and more capable of solving hard problems. The hard part became getting them to know to use those capabilities.

Generic prompts build generic software

So we’ve established that these models understand high level expert jargon. It’s somewhere in their training data, and so they should be able to apply it. So why do most AI-designed websites look the same, and why do vibe-coded websites barely ever work out-of-the-box?

It’s because every person is asking for the same thing. When you ask a model to make a website look good, it falls back to the most generic form of making something “look good”. In contrast, the people who are paid the most to make things look good don’t ever think of the problem as that black and white.

You make something look good by having opinions on how it should look. You make something work well by having opinions on how it should work. Those were core primitives before agents, and they remain core primitives. We don’t even need to implement every concept ourselves anymore. We still need to understand when it applies and recognize whether the model used it well.

Let’s say I want my website’s design to be Art Deco-inspired. Just being able to name Art Deco as a style unlocks a whole subset of the agent’s knowledge that it would never have fallen back to otherwise. And the more design jargon you know, the more of those “hidden rooms” within the model you can unlock.

This generalizes to everything. Engineers know the right engineering jargon to unlock the expert engineer rooms in the model. Doctors know doctor jargon, mathematicians know math jargon, and so on.

Given this, it’s not surprising that the recent counterexample to the Jacobian conjecture came from an IMO level mathematician prompting a model. Knowing what to ask for is a prerequisite to unlocking the full capabilities of these models.

Will this always be the case?

I personally see this going one of two ways.

The first is that jargon becomes unimportant, and the model companies make it so that everything “just works” regardless of how you ask for it. I find this possibility extremely unlikely, though, because there is a wide array of tasks where “just works” is entirely subjective.

What would that look like with design, for example? There is no universal marker of good design. Sure, there are frameworks and heuristics that designers are taught in school, but that’s still converging into genericness. I guess they could make the model resolve a generic query to some random configuration of design primitives, but even then that’s just randomized genericness.

A good argument here would be “What about plan mode? Why can’t we just have the agent ask us questions about how it should look?'“. Well, you still need to understand what it’s asking about to make the right decisions.

The second possibility is that jargon becomes the most important thing, and I do think that is way more likely of an outcome. The people who know the most about design remain the best designers, the people who know the most about engineering remain the best engineers, and so on. Your opinions on how things should work become taste that is encoded in every line of output from the model.

So what does that mean for the average person?

If you’re a skilled professional, this is a good thing. AI probably won’t be taking your job on its own. Who could take your job, however, is someone who understands more of the field than you do and can express that understanding to the model.

My advice (and this is what I’ve been doing myself) is to just consume, consume, consume. Subscribe to blogs, read Substack posts, watch YouTube videos, read books, consume anything that can expose you to jargon you’re not aware of.

And don’t just memorize words. Understand these concepts well. Why does one design style work well in one circumstance, but fail in another? Why do certain models perform well at certain tasks but not at others? The more of these questions you can answer, the more you’ll be able to get out of your agent, and the more your skills and outputs will stand out in the sea of generic software.

And take risks in your prompting. Make things that look different, that work different. We’re all tired of the same old generic slop software. Become informed, and use that edge to stand out. I’d love to see what you build.