The first time most designers use an AI tool seriously, they type something like "design a landing page for a SaaS product" and feel vaguely underwhelmed by what comes back. So they try a few more times, get a few more generic results, and conclude the tool isn't that useful for real work. The problem almost never is the tool. The problem is the prompt.
Prompting is a skill. It follows rules, it improves with practice, and the difference between a weak prompt and a strong one is not a matter of luck — it's a matter of understanding how these systems actually work. You are not typing a search query into Google. You are writing a brief for a very capable but entirely context-free collaborator.
Why Vague Prompts Fail
AI language and image models are trained on enormous amounts of existing content. When you give them a vague prompt, they default to averaging across everything they've seen. "Design a landing page for a SaaS product" produces something that looks like the mean of thousands of SaaS landing pages — competent, forgettable, and wrong for whatever specific problem you're actually solving.
This isn't a limitation of the technology. It's the logical outcome of under-specification. You wouldn't brief a freelance designer with "make something nice for my software company" and expect a compelling result. The same principle applies here. The quality of output is almost entirely determined by the quality of input.
Vague prompts also leave too many decisions for the model to make — and models make those decisions based on statistical likelihood, not strategic judgment. The more you leave unspecified, the more average the output will be. This is why most people's first experiences with AI tools are disappointing: they're using them like search engines rather than creative collaborators.
The Anatomy of a Good Design Prompt
A strong design prompt reliably contains four things: a role, a task, context, and constraints. Not always in that order, and not always in those exact terms — but these are the structural elements that turn a vague instruction into a useful directive.
- Role: Who the model should be. "You are a senior brand strategist working on a B2B fintech product."
- Task: What you want it to produce. "Write three hero headlines for a new invoice management tool."
- Context: The specific situation. "The target audience is freelancers and small agencies, not enterprise. The tone should be direct and slightly irreverent — we're positioning against legacy accounting software."
- Constraints: What to avoid or adhere to. "No jargon. No financial services clichés. Each headline under 8 words."
That's a prompt. Compare it to "write a headline for an invoicing app" and the quality gap becomes obvious. You're not just giving the model more information — you're eliminating entire categories of wrong answers before they appear.
Context, Constraint, Criteria
For day-to-day use, we've settled on a simpler mental model: every prompt needs context, constraints, and criteria for success.
Context is the situation. Who is this for, what stage of the project are we at, what has already been decided, what are the business goals, who is the audience? The more operational context you provide, the less the model has to invent. Think of it as filling in the brief that the model doesn't have access to.
Constraints are the fence posts. What format should the output take? What length? What tone? What should it explicitly avoid? Constraints are not limitations on creativity — they're the parameters that make creativity possible. A model with no constraints will produce the most probable output. A model with smart constraints will produce something surprising.
Criteria is how you'll know it worked. "The headline should make a freelance designer stop scrolling and feel understood" is a criterion. "It should be under 10 words" is a constraint. Both are useful. Criteria are particularly powerful because they force you to articulate what success looks like — which turns out to be clarifying for your own thinking, not just the model's.
Writing a good prompt is largely an exercise in articulating what you actually want. The discipline of doing that well is independently valuable, regardless of what the AI does with it.
Prompting for Different Design Tasks
The right prompting approach varies depending on what you're trying to produce. Layout work, copy, and imagery each have different failure modes and therefore require different prompt strategies.
For layout and structure, specificity about hierarchy and intent matters most. Don't just describe what you want to see — describe what the layout needs to accomplish. "A hero section where the primary CTA needs to be immediately obvious, with supporting social proof that doesn't distract from conversion" is far more useful than "a clean hero section with a button and some testimonials."
For copy, tone examples are everything. The single most effective technique is to give the model a sample of writing you want to match. "Here's a sample of the brand voice: [paste two or three sentences]. Write in this style." This anchors the output far more effectively than abstract descriptors like "professional but approachable."
For imagery, specificity about mood, light, and reference points beats generic style descriptors every time. "Shot on 35mm film, golden hour, candid rather than posed, editorial photography style similar to a Monocle magazine shoot" will consistently outperform "warm and professional photography."
Think in Rounds, Not Shots
One of the biggest mistakes designers make with AI tools is treating each prompt as a one-shot interaction. Get the result, decide if it's good, move on. This misses the most powerful part of working with conversational AI: the ability to refine iteratively.
A better approach is to think in rounds. Round one establishes the direction. Round two refines the most important dimension. Round three locks specific details. Each round narrows the space of acceptable outputs — you're not starting from scratch, you're steering.
When output misses the mark, diagnose what went wrong before reprompting. Was it the tone? The format? The strategic angle? Identify the specific failure and address it directly. "Make it better" is as useless in a refinement prompt as it is in a first prompt.
Give the model explicit feedback. "The second option was closest — take that direction, make it more assertive, remove the hedging language in the middle, and tighten to under 60 words." This is how you move from acceptable output to genuinely good output. The model has no memory of what you wanted to like about the last version unless you tell it.
Building a Prompt Library
Once you've found prompts that work consistently, save them. A prompt library is one of the most underrated productivity tools for a design studio. These aren't rigid templates — they're tested frameworks you adapt to each project. Over time you build reliable starting points for every common task: brief analysis, copy variants, image direction, client presentation outlines, competitive research synthesis.
The prompts worth saving have three characteristics: they reliably produce useful output, they require minimal editing to adapt to a new project, and they encode your studio's specific standards and voice. That last point is what separates a generic prompt library from a strategic asset. If your prompts reflect how your studio thinks about design, the output will consistently reflect your approach rather than an averaged version of the internet.
Some of the most valuable prompts in our library aren't generative at all — they're analytical. "Review this landing page and identify every place where it's making an assumption the target audience may not share." "List every element on this page that slows the user's path to the primary CTA." These prompts don't produce content; they produce insight. And that's often more valuable than another headline variant.
Real Examples That Work
To make this concrete, here are the kinds of prompts that consistently produce useful results in real studio work — contrasted with the weaker versions most designers start with.
- Weak: "Write copy for our about page." — Strong: "Write an about page intro for a two-person web design studio targeting early-stage tech startups. Tone: direct and confident, no corporate language. Goal: establish credibility without sounding like we're trying to seem bigger than we are. 80–100 words."
- Weak: "Give me some color palette ideas." — Strong: "Suggest three distinct color palette directions for a legal tech SaaS targeting mid-market law firms. Each should feel modern but credible — not trendy startup, not old-school lawyer. Describe each as primary, secondary, and accent, and explain the strategic rationale behind each choice."
- Weak: "What should go on the homepage?" — Strong: "We're redesigning the homepage of a B2B SaaS product for restaurant group staff scheduling. The main conversion goal is trial signups. The audience already knows they have a scheduling problem — we don't need to convince them it exists. Outline a homepage IA that moves a sceptical operations manager from arrival to trial signup in the fewest steps."
The pattern is consistent: be specific about the audience, explicit about the goal, direct about the constraints, and honest about what you already know. The model fills in the gaps. Your job is to make those gaps as small and well-defined as possible — and that discipline, it turns out, makes you a better designer with or without the AI.