AI has no design taste of its own. What you can do is supply taste from the outside — and some ways of doing that work far better than others.
What 'taste' means for an AI
Taste, in design, is a consistent set of intentional choices — a point of view applied the same way across every screen. An AI has none of this by default. Left alone it returns the average of its training data, which reads as competent but anonymous.
So 'giving your AI taste' is not about unlocking something inside the model. It is about feeding it a specific point of view from outside, in a form it can apply consistently. The question is only which form works best.
Lever 1 — better prompts (weakest)
The first lever is the prompt itself: be specific, name constraints, show one example. This is real but limited. The instruction shapes a single output and then evaporates; across many screens the model drifts back to its default, so consistency is hard to hold.
Prompts are the right tool for a one-off and the wrong tool for a system. If you need ten slides or twenty components to feel like one product, the prompt is too thin a thread to carry the taste.
Lever 2 — reference images and examples
The second lever is showing, not telling: paste a screenshot, link a site you like, give a few examples to imitate. This lands better than adjectives because it points at something concrete.
But references are lossy. The model infers what it thinks matters from a picture, and different runs infer differently — it might copy the color but miss the spacing logic, or borrow the type but not the grid. You get closer to the look without ever pinning it down.
Lever 3 — a machine-readable design spec (strongest)
The strongest lever is to hand the AI an actual design system in a form it can read: the real colors, type scale, spacing rhythm, shape language, and component rules, written down. Now the AI is not guessing at taste from a picture or an adjective — it is applying a defined system, value by value, the same way every time.
This is what closes the consistency gap. A spec does not evaporate between prompts and it does not have to be re-inferred from an image; it is the same input the model can apply to slide one and slide forty. The clearest form of this is a DESIGN.md — a single machine-readable file that carries a complete style, ready for an AI to consume.
Curio exists to supply exactly this lever: each design style packaged as a spec your AI can apply directly, so you are not relying on prompt-luck for a consistent, intentional look.
Putting it together
Use the levers in order of strength. Start from a real design spec as the backbone, lean on references where the spec is silent, and reserve prompts for the small, local adjustments. The spec carries the taste; the prompt carries the task.
The practical move: pick one documented style, give your AI its spec, and keep every screen on that same system. Taste, for an AI, is just specificity supplied consistently from outside.
FAQ
Can't I just tell the AI to 'use good design'?
No — 'good design' is an adjective, and to a model an adjective resolves to its average. Every style in the training data claims to be good, clean, and modern, so the instruction points nowhere specific and you get the mean. Taste has to be supplied as concrete, consistent rules.
Reference images or a design spec — which is better?
A spec, for anything beyond a one-off. References are lossy and re-inferred differently each run, so consistency suffers. A spec is the same deterministic input every time, complete and reusable across many screens — which is exactly what a consistent look requires.
What format does the AI actually consume?
A machine-readable spec it can parse and apply — for Curio styles, a DESIGN.md: one file holding the full palette, type, spacing, and component logic. You hand it over by download, share link, or an MCP connection, and the AI applies the values directly.
Originally published on designbycurio.com. Curio is a design style library for AI agents — 1,000+ real design styles as machine-readable specs your AI can actually follow. Related: Why AI-generated design all looks the same and What is DESIGN.md?
Top comments (0)