Index

Design agent: from request to handover-ready UI

My role
Product design lead
Team
1 product designer, front-end team
Timeline
Ongoing

Description

An AI design agent that turns a plain request into an on-system design in three fidelity stages, checks each stage automatically, and improves with every designer correction.

Read case study

The problem

Early ideas need to reach stakeholders fast, but every new screen starts from scratch. Without a shared base, visual drift creeps in, and reviews turn into debates about colour and corner radius instead of the idea itself. The design system solved this for the product team, but not for everyone who needs to explore an idea.

What the agent does

A designer describes the page in plain words. The agent works in three stages, and each one is cheaper to change than the next:

  1. Low-fi: a text wireframe that settles structure and flow. It’s the cheapest stage, so ideas get rejected early.
  2. Styled: the same structure in the product’s look, loosely applied.
  3. Strict: a high-fidelity mockup built only from confirmed components, with tokens named so engineers can hand it over.

A coordinator picks the stage and enforces the order, so nobody jumps to polished visuals before the structure is agreed.

How it stays on-system

The agent reads from one source of truth for every stage: tokens, the app shell, rules, personas, terminology, voice and templates. These were extracted from the frontend code and its docs, so the agent designs with the same values engineers ship.

Four automatic checks run alongside the stages:

  • Lint hook: blocks raw colour values and off-system pixel breakpoints before a designer sees the file.
  • System checker: flags components that duplicate an existing one, or use off-scale spacing and type.
  • Flow writer: checks that every state is covered, such as empty, loading, error and offline, with clear interface text.
  • Visual fix: renders the result and corrects misalignment, overflow and responsive breakage.

Learning from corrections

Each correction a designer makes is logged with its cause. A repeated mistake becomes a new lint rule or a regression test, so an improvement can’t be undone by a later change. The agent gets better without being retrained.

The deliberate choice

The interaction rules are intentionally loose. Locking the agent to only reassemble existing components would cap it at what the library already contains, which limits what it can propose. It keeps headroom to suggest better solutions, while staying inside the visual guardrails. As the component library matures, the constraint tightens step by step.

Status

The scaffold is built: the stages, the design system source and the checks. The improvement loop is still planned, so I treat this as an experiment under test rather than a finished product. See the design system project for the foundation it builds on.