Design prompting, explained

Every interface an agent builds starts as a sentence someone wrote. Design prompting is the craft of writing that sentence well, so the agent produces what you meant instead of a generic approximation of it.

The difference between a throwaway prompt and a good one isn't length. It's structure: what outcome you name, which constraints you state, and how you iterate on what comes back. Teams that treat prompting as a design skill, something drafted, critiqued, and refined, consistently get production-quality interfaces from the same models that give everyone else demo-ware.

What a design prompt actually does

A design prompt is a brief, and it does the same job a brief has always done: transfer intent from the person who has it to the one doing the work. The receiving party happens to be a model now, but the failure modes are the old familiar ones. A vague brief produces safe, average work. An overloaded brief produces confused work. A brief that specifies the solution instead of the problem produces exactly what you asked for and nothing you needed.

Models add one twist worth understanding. An agent has effectively seen every interface on the public internet, so when your prompt underspecifies, it fills the gaps with the statistical average of everything similar. That is why unguided output has a recognizable flavor: the same hero section, the same three-card feature row, the same dashboard with the same four stat tiles. The average is not wrong. It is just nobody's product. Your prompt's job is to pull the output away from the average and toward your particular users, brand, and constraints.

The anatomy of a prompt that works

Across the guides in this pillar, the prompts that consistently produce good interfaces share four ingredients.

An outcome. What should be true when the screen succeeds? "A signup form a distracted person completes in under a minute" gives the agent a goal it can design toward. "A signup form" gives it a category to average.

Real constraints. The two or three things that are actually non-negotiable: fits one screen on mobile, uses our existing components, never asks for information twice. Constraints are where your product's character lives, and stating them is cheaper than fixing their absence.

Context. Who is this for, and what happened right before they arrived at this screen? A checkout for an impulse purchase and a checkout for a procurement officer are different designs, and the agent will only know which one you need if you say.

Silence about everything else. The hardest ingredient. Every detail you specify is a decision the agent can no longer make well, so specify the ones you actually care about and leave the rest open. The agent's strength is that it has seen a thousand solutions to the parts you have no opinion on.

The four guides

This pillar collects our field guides to the skill, and they build on each other:

Start with the best practices if you're new to prompting for design; move to composable and outcome-driven prompting once you're shipping real screens; use iterative refinement as your everyday working loop.

Prompt at the right altitude

One idea threads through all four guides: prompts have altitude, and most failed prompts fail by flying at the wrong one.

A monolithic prompt tries to specify an entire product in one message: dashboard, sidebar, settings, notifications, dark mode, everything. The agent must make fifty design decisions simultaneously, and when three of them come out wrong you have no way to revise those three without regenerating the rest. At the other extreme, a prompt that micromanages pixel by pixel wastes the model's pattern knowledge and produces brittle, committee-flavored screens.

The productive altitude is the component. Describe the pricing card. Then the plan toggle. Then the comparison table. Then compose them. Each atomic prompt is small enough to evaluate at a glance and revise in isolation, and the composition mirrors how good interfaces are actually built. The composable prompting guide develops this into a full method.

Iteration is the medium

The first prompt is a sketch, and treating it as anything more sets you up for disappointment. Professionals who get remarkable results from agents are rarely writing remarkable first prompts. They are running a tight loop: generate, look at the running screen, name the most important thing that is wrong, revise, repeat. Five focused rounds beat one heroic paragraph every time.

This is also where the working surface matters. Iterating in a chat window means each version scrolls the previous one away, and comparing take three with take one becomes an act of memory. On a canvas the takes sit next to each other, the live preview runs each one for real, and steering happens by pointing at the thing that needs to change. When the agent writes actual code rather than rendering a picture, your critique lands on something that can be tested, not just admired. The iterative refinement guide covers the loop in detail, including how to carry context across sessions.

Why this skill compounds

When the agent writes real, running code, the prompt is the design brief. A vague brief gets you the median interface for your request. A precise brief gets you your interface, the one that fits your users, your brand, and your constraints. And unlike tool-specific tricks, the underlying skill, saying exactly what you mean about interfaces, transfers everywhere: briefing agents, briefing contractors, writing specs, critiquing work. Prompting is design writing, and design writing has always separated teams that ship intentional products from teams that ship averages.