Most designers I talk to either aren’t using AI at all, or they’re using it like a search engine. They ask vague questions, get generic answers, and walk away assuming Claude just isn’t that useful for design work.
I used to feel the same way. Then I changed how I set it up.
The setup is the work
What you get out of Claude depends almost entirely on what you give it before you ask for anything. Vague input produces vague output, and specific input produces specific output. That sounds obvious, but most people skip the setup and go straight to the prompt, then wonder why the results feel generic.
I started treating Claude the way I’d treat a junior designer: capable, fast and willing, but in need of direction. Not a genie. A collaborator who needs context.
Give it structure, not adjectives
The first thing I do in any design session is drop in a structured snapshot of the design system I’m working with:
{
"colors": {
"primary": "#2216FF",
"canvas": "#FFFEFC"
},
"spacing": "8px grid",
"radius": 12,
"font": "DM Sans",
"minBodySize": 16
}Not a PDF, not a mood board. Specific values.
“Modern and clean” means nothing to a model. radius: 12 and gap: 24px do. The more concrete the input, the less correcting you’ll do on the other end.
Codify your taste
Claude doesn’t carry memory between sessions. Without some kind of preferences file, you’ll spend the first few minutes of every conversation correcting the same mistakes.
So I keep a short markdown file of my recurring design rules: always use an 8px grid, buttons at a 12px radius, no drop shadows on cards, icons at a 1.5px stroke. I paste it in whenever I’m doing design work. It sounds tedious, but it saves me a lot of time. Think of it as writing your taste down so you don’t have to explain it again.
Learn enough to speak the language
You don’t need to be a developer. But a basic understanding of HTML and CSS changes how useful Claude is to you.
Claude produces code. If you can read it, even loosely, you can spot when something’s off and correct it precisely. Without that, you’re judging by eye and guessing at fixes.
The same goes for feedback. “Make it pop” doesn’t translate. “Increase font-weight to 600” does. CSS is a shared language between you and the model, and the better you speak it, the less gets lost in translation.
Work in layers
The other big mistake I see is one giant prompt that tries to specify everything at once. You lose control of the output, and when something’s wrong, you can’t tell which part caused it.
Working in layers works better. Start broad with “build me a card layout”. Then refine a section, then an element within that section, then a single property. Each round narrows the gap, and small, directed changes keep you in the driver’s seat. By the time you reach the details, you’re asking for something so specific that Claude can hardly get it wrong.
Save your versions
Because Claude won’t remember it for you, an iteration you’ve moved past is gone unless you saved it.
Before any significant change, I label and copy the current state. Even pasting it into a Notion block titled v2-spacing is enough. Then, when I want to go back or combine two directions, I have the pieces. “Go back to v2 but use the spacing from v3” is a powerful instruction, but only if you’ve been saving as you go.
Art direction, not magic prompts
There’s a way of using Claude that feels like wishing for the right output: you craft the perfect prompt, hope for the best and adjust from there. That’s not how I think about it anymore.
The better mental model is art direction. You know what you want. Claude is fast and capable, but it has no taste of its own. Your job is to set the direction clearly, correct course precisely and stay in control of the vision.
Once I started thinking about it that way, the outputs stopped feeling generic. They started feeling like mine.
