How to brief an AI the way you'd brief a designer

If you've ever typed “make this better” into a chat box and wondered why you got back something worse, this essay is for you. Most people who are frustrated with AI tools aren't frustrated with the models. They're frustrated with what their own sentences produce when fed into a model. Better prompts, better results. And better prompts look a lot like the briefs you'd write for a human designer, if you ever had the budget to hire one.

The shift in mindset is the important part. When you're talking to a designer you respect, you don't hand them a two-word assignment. You sit down for a coffee, you explain the thing you're trying to build, you show them a reference or two, you talk about the mood, you name the constraints. That conversation is the brief. It's also exactly what a good prompt looks like.

The shape of a good brief

Professional design briefs, the ones agencies use, tend to have five ingredients. They're usually written as a document, but you can collapse them into a sentence or two and feed them to an AI and get a dramatically better result than the typical two-word prompt.

The thing.What are you making? A landing page, a portfolio, a shop, an invite. This is the easy part; most people get this right. “An Italian café landing page.”

The feeling.How should it feel when someone arrives? Warm, elegant, playful, serious, dramatic, cozy, minimal. Adjectives the model can act on. This is where most prompts fall apart — people skip the feeling because they think it sounds vague, but it's the most load-bearing ingredient in the whole brief.

The audience.Who is going to see it? “A reservation form for a neighborhood café” is completely different from “a reservation form for a three-Michelin-star restaurant.” Same noun, different result.

The references.Show or name something that's close to what you want. “The feel of a New Yorker article.” “Like a Muji notebook.” “Something with the warmth of a hand-lettered chalk sign in a coastal Italian town.” References anchor the model in a specific direction and stop it from defaulting to the average.

The constraints.What's off-limits? Don't use a dark background. Don't add animations. Keep it under one viewport on mobile. Constraints are gifts. The more you give, the less the model has to guess.

Before and after

Consider these two prompts, which I've watched people type in real time. They're both for the same project.

The result is the average of every café website in the training data. Serviceable. Forgettable. A hero with a stock photo of coffee. A menu section that looks like Squarespace. A contact form. A footer.

Ten times as much text, a hundred times better a result. This isn't because the model suddenly got smarter. It's because you stopped asking it to guess and started telling it what you wanted. Every sentence in that second prompt is a decision the model no longer has to make by rolling dice.

The sentences that work

After watching a lot of people use AI builders, I've noticed a handful of sentence patterns that consistently produce good results. If your brain goes blank while writing a prompt, steal these:

“The feel of X.”“The feel of a New Yorker article.” “The feel of a Muji notebook.” “The feel of a print magazine from the seventies.” These are dense references. One sentence gives the model enormous context.

“Like X, but for Y.” “Like a film poster, but for a wedding invite.” “Like a Linear landing page, but for a café.” “Like a New Yorker feature, but online.”

“Warm and X, not cold and Y.” Specifying what to avoid is as useful as specifying what to aim for. “Warm and editorial, not cold and corporate.” “Playful and specific, not generic and safe.”

“Headline should say X.”If you have an opinion about the exact words, state them. Models are usually happy to use the exact sentence you give them; they just won't invent it.

“Don't use X.”Naming cliches to avoid is one of the most effective levers in a prompt. “Don't use gradient mesh backgrounds or floating dashboard screenshots.” “Don't use stock photography.” “Don't use emoji as section icons.”

The sentences that don't work

And a few patterns to avoid, because they make your output worse in predictable ways:

“Make it better.” Better how? The model will guess, and its guess will be to add more stuff. You almost never want more stuff.

“Make it modern.”“Modern” in training data means “average AI landing page.” You're asking the model to regress to the mean. Be specific: modern like what?

“Make it pop.”This adds bright colors where bright colors don't belong. If you mean “increase visual contrast on the hero headline,” say that.

“Professional” and “clean.”These are high-entropy words. They map to a thousand aesthetics at once. Skip them. Describe the specific professional you're imagining.

One change at a time

When you're refining an existing project, batching changes is tempting and counterproductive. The prompt “make the hero bigger, change the colors, add a footer, and remove the third section” is four requests in a trench coat. The model will try all four at once, and the result will be worse than if you'd done them one at a time.

One change per turn. Short, specific, deliberate. You can see the effect of each change, decide if you like it, and undo just that one if you don't. Batched changes are all-or-nothing and make undoing harder than it needs to be.

The whole trick

The whole trick to writing good prompts is the same trick to writing good anything: be specific, be opinionated, cut the words that aren't doing work. Treat the model like a collaborator who will take you literally. Tell it what you want. Tell it what you don't want. Give it a reference. Get out of its way.

Do this for a week and the frustration will lift. You won't be wondering why the AI keeps producing generic output. You'll be getting back exactly what you asked for, which is often the first surprise on the road to making something good.