Magic Words in Prompting: Domain Terms That Steer Behavior

In the early days working on GPT‑3, one thing became obvious fast: language is intensely domain-specific. Certain words and phrases do far more “work” than others. I’ve called them “magic words” before—not because they’re mystical, but because of how strongly they’re connected to particular patterns in the model.

A simple example: if you’re building a customer support chatbot, telling the model to be “polite” often works better than telling it to be “helpful,” even though those words sound similar in normal conversation. In the model’s embedding space, they aren’t similar in the way you’d expect. “Polite” tends to pull in a very particular style of writing and behavior. “Helpful” can mean a lot of things, and the model may respond with something that’s technically helpful but not in the tone you were aiming for.

The same idea shows up in professional writing prompts. If you say, “Write a professional article about X,” you might get something decent—but you might also get something that meanders. “Professional” is a vague instruction. It doesn’t anchor the model to a strongly standardized format. It’s a single adjective among many, and it isn’t tied to one consistent structure.

But if you say, “Write a Wikipedia article about X,” you’ll usually get something that follows Wikipedia’s format much more closely. Why? Because the model has seen a huge amount of Wikipedia. “Wikipedia” isn’t just a style hint—it’s an extremely strong pointer to a familiar structure: the lead section, headings, an encyclopedic tone, a certain way of defining terms, and so on. You could try “write a reference article,” and it might get close, but “Wikipedia” is simply a more powerful anchor because it appears so frequently in the training distribution and is associated with a very consistent pattern.

This became even more obvious when image models arrived—first with DALL·E, then DALL·E 2. Suddenly, a lot of people who had never written prompts for language models were getting incredible results from image generation. And the reason wasn’t that they were secretly prompt engineering geniuses. It was that they were using language they already knew from photography and cinematography.

They’d ask for “shallow depth of field,” specific lenses, specific cameras—terms that are second nature if you’ve spent time around photography. If the training data linked images with that kind of text (and a lot of it did), the model could make extremely strong connections. In many open datasets, images aren’t paired with just “a photo of a dolphin.” They come with metadata: camera model, lens, focal length, time of day, location, and other descriptive context. When models are trained on image-text pairs, that extra text matters. If “Nikon 35mm” appeared next to images shot in that style often enough, then “Nikon 35mm” becomes a very strong lever. The model doesn’t need to infer it purely abstractly—it has seen the association repeatedly.

That’s also a good reminder of something people sometimes miss about these systems: they both generalize and memorize. They learn relationships between concepts, and they also store a lot of surface-level associations from the data. Those two things coexist. And when you’re not getting the output you want, it’s sometimes because you’re leaning on the part you wish it had generalized, when what you actually need is to lean into what it may have memorized—specific terms, formats, jargon, or canonical labels.

The bigger takeaway is simple: think about the domain where your target style or knowledge actually lives, and what kinds of examples the model likely saw during training.

You can train an image model on tons of images with only basic descriptions and no camera data—nothing about lens, angle, lighting, or composition. And then you can try as hard as you want to prompt “Dutch angle” or “shot on 35mm film” or “golden hour,” but it’ll be much harder for the model to deliver if it never had those associations to learn from. Same as a person. If I kept repeating “Dutch angle” at you and you’d never seen it used, never watched something like the old Batman show where it’s iconic, and never had it explained, you wouldn’t magically know what I meant. Models are the same way: they need grounding in examples.

So when you’re prompting—whether it’s for writing or images—don’t just describe what you want in plain language. Reach for the words the domain itself uses. Find the canonical format name. Use the terminology that would show up in real captions, real documentation, real metadata, real communities. Those are often the “magic words” that turn a vague request into something the model can reliably execute.