Design system for generative UI in a clinic operations tool, from tokens to deployed screens.

OVERVIEW

A design system built for AI orchestration.

This system serves as the foundational architecture for a clinical operations concept where pages dynamically assemble around real-time user intent. I designed the tokens, components, and layout rules in Figma, then mirrored the system in code to ensure the AI assembles layouts using real, deterministic components.

ROLE

Solo UI/UX designer + developer

TOOLS

Figma, Vercel, Claude (via Figma MCP server & GitHub)

YEAR

2026

CHALLENGE

Building a deterministic framework for non-deterministic UI.

How might we design a strict, semantic system of tokens and components that allows an AI engine maximum flexibility to assemble screens, while guaranteeing clinical clarity and data accessibility?

PROCESS

An automated pipeline from canvas to code.

Designed in Figma

Established visual foundations, semantic tokens, and rigid state properties within a centralized component library.

Engineered via Claude

Leveraged the Figma MCP server and GitHub to translate component properties directly into clean, executable code.

Deployed to Vercel

Refined and stress-tested the generated code against the Figma library before deploying it as a live prototype.

COLOR & TYPOGRAPHY

Semantic tokens and tabular lining to protect clinical legibility.

Semantic Tokens

Primitives stay hidden. Components reference functional roles (e.g., brand-primary, clinical-alert), ensuring a single change to a primitive safely updates the entire ecosystem.

Tabular Alignment

Utilizing high-legibility typefaces optimized with tabular lining figures so clinical dates, times, and patient IDs remain scannable in dense views.

SHAPE & ELEVATION

Proportional geometry and semantic shadows to define layout depth.

Nested Geometry

Corner radii scale proportionally to maintain visual consistency when components are nested in layouts.

Functional Depth

Shadows are semantic. A subtle lift indicates clickability, while distinct elevation reserves focus for floating modal layers.

INTERACTION

Root-level states to prevent layout fragmentation.

States (hover, focus, disabled, loading) are standardized at the root component level, eliminating visual fragmentation when the AI dynamically assembles layouts.

AI REVIEW COMPONENTS

Clear visual indicators that communicate system confidence and source.

Specialized modules designed explicitly for AI-to-human communication. Dedicated tokens display data provenance (Source), machine certainty (Confidence), and explicit human validation (Status).

CORE PRODUCT COMPONENTS

Fixed structural components that support safe, dynamic screen arrangement.

Larger components combine the basics: a command bar, view header, patient card, timeline, document tracker, menus, toasts and pinned views with live counts.

CODE / IMPLEMENTATION

Syncing Figma architecture with code for system integrity.

I carried the same token names into the implementation using shadcn/ui, so the design system wasn't just a Figma library. The prototype uses the same primitives, semantic tokens and component structure.

REFLECTION

Live browser testing surfaced architectural logic flaws.

Transitioning from a static Figma library to a live browser taught me that visual perfection in Figma doesn't guarantee structural usability. Moving quickly to code exposed unexpected UX collisions. Static frames couldn't simulate a sidebar reordering unpredictably or multiple inputs overcomplicating a single task. For v2, I want to refine the visual hierarchy, simplify overlapping interactions, and test the system as the interface becomes more dynamic.