AI Experiment • UI & Frontend

Agentic Design Systems

Turning wireframes into working screens, with AI.

Where my design time actually goes

I like solving a problem completely: every state, every edge case, every error. But on data-heavy products, much of my time went into applying design system rules, screen after screen. Necessary, but not problem-solving. So I wondered: could the rules do the assembly, freeing me for users and testing?

A wireframe of a product screen covered in sticky notes that capture design-system rules

The Idea

The design system was well documented. Every rule, every component, how to use it, its states and variations. All written from a perspective of a designer to design screen in future.

Which made me wonder. What if I rewrote those same rules for an AI agent, in a form it can understand and act on?

The design system documentation: grid and type scale, colour palette, status chips and form-input states

Setup

I connected the agent to IBM Carbon’s own Storybook, where every component, state and variation is documented and open on the web. Then I wrote the rulebook it reads before building anything.

  • Live components

    All 137 Carbon components, composed into my Storybook from IBM’s own.

  • Tokens and rules

    Every colour, spacing and type token, plus a list of what never to do.

  • Wireframe playbook

    Seven steps from wireframe to components, grid, and states.

Experiment

The wireframe was well structured. It carried every front-end rule: accepted data types, limits, cross-field checks. I handed it to the agent with one prompt: convert this into a front-end ready screen, using only Carbon’s guidelines. The prompt works like a brief to a senior engineer: analyse the wireframe, map it to components, define interactions and validation, build, then check its own work against a hygiene checklist.

The rulebook file the agent reads before building: a code editor showing the front-end rules

Result

The screen came back working. Loan amounts format with commas, percentages stop at their caps, Max has to exceed Min, and dates can’t be backdated. The wireframe’s validation rules were all applied.
It also broke. Inputs lost focus after every keystroke. Tables grew two scrollbars. Date pickers jumped the scroll position. Each failure became a rule in the file, so it wouldn’t happen twice.

The screen the agent built: a Product - Create form with tabs and an editable table

Learning

The agent didn’t replace the thinking. It took the assembly.

1.

Treat it like a junior designer

It handles execution. I handle decisions, users and testing.

2.

Structure makes AI trustworthy

Without rules, agents make things up. With a design system written for them, they stay on brand.

3.

The effort moves upstream

Good wireframe and rules in, good screen out. Teams get consistent, on-brand screens, and the thinking goes into the spec.