Design Feedback for AI Agents

Luke Wroblewski says a design system gets an AI agent about 90 percent of the way to a finished screen, and pointed visual feedback closes the rest.

Luke Wroblewski, a product designer and author, describes how his team finishes screens built by AI agents from a design system. No design system can cover every layout in a real product, so the team shows the agent each remaining problem on screen.

• The design system and the agent's written instructions produce the first version, which reaches about 90 percent, compared with 50 to 60 percent with earlier AI models.
• The team tried comments in Figma, marked-up screenshots and typed requests, and then built screen markup, recordings and voice notes into Intent, the AI coding tool it makes.
• Wroblewski deliberately keeps design systems small, because they otherwise grow in cost and complexity.
• He argues that AI models would add too much documentation if allowed to update the design system, so a designer decides what goes back in.

Design teams can expect to spend less effort writing a complete design system and more on fast visual corrections to what the agent builds. This is an inference and is not the author's claim. The tool described is one his team makes.

Derived by Working Surface from the article; more in the Playbook. Source line: If given free license today's models will err on the side of over documentation instead of the pattern simplification and cohesion needed in an effective design system.

5 November 2026: teams let AI agents gather the evidence and do the building, and decide in advance where a person must approve, check or judge.