Designing for machines 

When PMs and agents started generating product surfaces, the design system became responsible for bugs it didn’t produce. There were more screens being produced than we could review.

Mid-year, the company pivoted hard to an AI-first operating model. Several things happened at once: 

The consumer base for design systems changed.
Everybody began building prototypes with AI tooling. Engineers and PMs began generating UI without a designer in the loop. 

The forcing function disappeared.
When designs were in Figma, things were pulled from the component library automatically. AI prototyping removed that entirely. Nothing was structurally obliged to generate output using the system.

Accountability didn’t change.
Finally, but suddenly, design systems were acknowledged as the backbone of output. It was accountable for bugs in code and features I had never laid eyes on, produced by people who had never seen the design system libraries. Everything became a design system problem.

Outputs vary wildly depending on who was prompting – and results even vary between runs of the same prompt. A design system is a carefully defined set of constraints. Without these reaching the generation layer, AI tools improvise happily and confidently. 

The original challenge I was handed was to turn the design system into a machine-readable asset that AI agents, PM-built prototypes, and engineer vibe-coded features all consume from. But the decision that actually mattered was narrower: speed was the name of the game, so do we offer guidelines and constraints at generation time, or review the output against the system afterward? 

Reviewing after the fact was considered too late for us, and it wasn’t going to scale with output volume. 

Constraining at generation means the system has to learn towards determinism, but it requires structural work on the system itself: a new component architecture designed for AI consumption (more on that here), tokens mapped directly into the config, and consolidation of documentation sprawl into a single structured source. We had our documentation scattered across Figma, Storybook, and mobile dev tools, which was blocking AI integration as there was no single source to point at.  

The solutions we are working toward are a prototyping kit and integrated docs site, agent consumable, that lets anyone in the business spin up a sandbox prototype that’s design system compliant by default. 

Metrics we’re looking at:

• Reduction in time from prompt -> production
• % of AI-generated layouts requiring no major rework
• Ratio of accurate first-time decisions
• Reduction in inconsistency across surfaces generated by AI tools 

I accept that deterministic is aspirational, and this work might be obsolete in three months — but we’re having fun experimenting!