How might we rebuild a fragmented legacy design system so that people and AI can both build from it reliably?

Enabling AI-assisted development while battling AI-accelerated inconsistency
The Chief Product Officer wanted three things: ship faster, ship more consistently, and let product managers do exploratory design without needing Figma. AI-assisted development was the route.
The starting point made it difficult. eClinical solutions has a legacy design system that had grown organically until it was unwieldy. Its documentation was out of date, its code was full of hard-coded values, and how it worked lived as tribal knowledge with a few people, which bottlenecked collaboration. An AI tool given that system would reproduce its inconsistency at speed.
So the real problem came before the AI. The product needed a design language that was explicit and structured enough for a machine to follow, and clear enough for a person to trust.
At a glance
Company
eClinical Solutions
Industry
Healthcare technology
Timeline
Dec 2025 to Nov 2026
Role
Principal UX Designer. Primary designer on the work, partnering directly with the Chief Product Officer and product and engineering leads.
Team
Product management and Engineering in Platform and Risk Management teams
Established the foundations
I started from DevExtreme’s token structure, since it’s the component library elluminate is built on, then added color, typography and spacing ramps with semantic tokens on top, so every value had a name and a purpose. I also standardized product and design nomenclature, such as what separates a product, a module, a feature and a workflow, so one name meant one thing everywhere. That consistency is what lets AI read the system reliably.
Designed for flexibility, not uniformity.
I framed out governance and a subsystem architecture so the system could scale and individual products could adapt where they needed to. Flexibility was a core principle because rigid alignment for the sake of consistency can produce a poor user experience, so the system was built from the start to accommodate differences between products.
Made it usable by AI and humans.
A GitHub repository became the shared source. Figma variables and components became AI-drafted Component Reference Specifications: plain, structured descriptions of usage, variables, accessibility, when to choose one component over another, and open questions. An engineering colleague turned the specs into code tokens so design and code shared names, and the team built from the repository with Claude and Cursor.
Gave colleagues a way in
Early AI output was inconsistent, so we built a Claude skill that helped PMs and designers create on-system HTML prototypes without Figma. It connected to the GitHub repository, then asked the person questions to draw out the context it needed, such as the persona, their challenges, linked Jira tickets, and competitive material. Building a prototype became a collaboration instead of one giant upfront prompt, and the goal was exploration through the specs, not cookie-cutter screens.
Outcomes
1
More than 500 foundational tokens, 30 components, and several UI patterns codified, replacing raw values and stale documentation with one source.
2
One major greenfield feature, Portfolio Risk Oversight & AI Analysis, was built through the pipeline.
3
Several PMs and engineers working with the new system for their daily work since mid-Q2 early release, along with contractors brought on to produce concept designs for the company’s annual conference in Q3.
Lux provided reusable tokens, detailed component guidance, and instructions supporting both manual and AI-assisted development. This dramatically reduced the time spent scaffolding components, aligning styles, and correcting UI inconsistencies. The result was faster development, fewer UI-related bugs, and time savings ranging from hours to days on individual tickets.
Lead Developer, eClinical Solutions
What I learned
1
AI does not fix an inconsistent system, it amplifies it, so the unglamorous work of naming and structuring decisions is what makes AI-assisted design possible.
2
Writing a great spec with AI is also a fine balance. Too much context up front works against you, and the best specs came from a focused start followed by rounds of refinement with the AI.




