AI-powered Design System workflow
Notes from an initiative I started two month ago.
At WorkJam, our design system spans three libraries: Figma for design, Storybook for web, and separate native libraries for iOS and Android. Each holds part of the truth. None of them hold all of it.
The artifacts are there: components, variants, tokens. What's not there is the reasoning: when to use this component versus that one, why a pattern was chosen, what edge cases broke previous implementations, what rules a new variant has to respect. That knowledge lives in designers' heads. Mostly in mine. This creates predictable friction:
Drift between the three libraries. A component evolves in Figma; the Storybook version lags; the native versions lag further. Without a shared source of truth for intent, each library drifts on its own timeline.
Devs blocked on Slack. Every implementation question routes back to a designer. Some questions repeat across teams. The bottleneck is human availability.
Slow onboarding. New designers and devs spend weeks absorbing tribal knowledge that should be written down once.
We have a separate design tokens pipeline that already works well, so tokens are out of scope for now (even though I'm thinking about ways to automate it with AI…). The component layer is where the pain is.
Figma holds visuals. Storybook holds web code. The native libraries hold mobile code. None of them are built to hold decisions, rules, and rationale, and just as importantly, none of them are easily consumable by AI tools.
That second point matters more than it used to. Devs increasingly work with AI assistants in their IDEs. Designers use AI to draft, iterate, and document. If your design system can't be read by an AI agent, you're forcing every AI-assisted workflow on the team to either reinvent the rules or skip them. Both outcomes produce drift.
Attending the Into Design Systems AI conference in March 2026 sharpened the direction. The initiative I proposed: a centralized, AI-consumable knowledge layer that sits alongside the three libraries and holds what they don't.
We're building it as a Bitbucket repo "workjam-design-system" containing MD and JSON files for every component, pattern, and guideline. Plain text. Version-controlled. Queryable. Accessible to any designer or developer, and more importantly accessible to AI agents.
Three concrete roles, not one:
Accelerated documentation. We use the Figma MCP server with Claude to extract component metadata and draft documentation faster than writing it by hand. This is what makes populating the repo at scale realistic for a small working group.
Self-serve guidance for developers. Because the repo lives in Bitbucket alongside our codebases, devs can query it directly through their AI tools while implementing. The expected outcome: fewer Slack interruptions, faster implementation, less guesswork.
AI-generated component scaffolds. With structured documentation as input, AI can generate scaffold code for new component implementations across Storybook and native, reducing the manual translation step that's a major source of drift today.
None of these is magic. Each is a specific lever against a specific cost.
I proposed this initiative and run a bi-weekly working group: the design team plus three developers from across web and mobile. We're intentionally small while we validate the process. Broader rollout follows once the core workflow holds up under real use.
My role in practice: defining the architecture of the repo, setting documentation standards, running the working sessions, building the first wave of component docs myself to pressure-test the format, and coordinating with the three devs to validate that what we're producing actually answers their questions when they're implementing.
Too early for results. The metrics we're set up to track:
Drift incidents between Figma, Storybook, and native libraries.
Time to ship a new component end-to-end across the three platforms.
Designer and developer onboarding time to productive contribution.
Volume of design system questions flowing through Slack.
Near-term: finish the first wave of component documentation, expand the working group, validate that the AI scaffolding workflow holds up across all three platforms.
Longer-term: connect this back to the design tokens pipeline so the full system; tokens, components, patterns is consumable as one coherent layer by both humans and AI agents. That's a project for later this year.
Two months in. The thesis holds. The work is real. More to come….