0%
0%
GitHub
Cursor
Claude Code
Claude
FigmaContext
Verification, scheduling, claims, the agents could handle it all. But every hour sales spent walking users through setup was an hour they couldn’t sell. I had four months to change that.
AI agents required a human to configure
UI built for engineers, not healthcare teams
No orchestration without a sales call
“Incredibly useful, impossibly confusing to use without a demo.”, User during discovery interviews
Research
Two main user groups emerged from research: nurses managing patient workflows and admin staffhandling scheduling, claims, and verifications. The recurring blocker wasn’t trust. It was abstraction. Both groups needed examples before setup.
What I learned
Users couldn’t describe an agent setup until they saw one.
Shared blocker
Different roles, same friction: configuration was too abstract.
Why it mattered
Every workflow still depended on a sales-led demo.
Iteration
I explored forms, a guided wizard, and chat. Chat won because it let users discover possibilities while they configured.
Discarded
Forms-first felt familiar, but it still assumed users knew what to ask for.
Also tested
A wizard improved onboarding, but it was too rigid for ongoing use.
Chosen
Chat let users explore, clarify, and self-serve in the same flow.
Design
XY had never had a dedicated designer, just engineers making UI decisions. Before I could ship, I needed the foundations every screen would depend on:
What this is
The complete color token system, primitives, semantic tokens, accessibility audit, and proposed system
Why it matters
Every component, every agent status, every interaction maps to this semantic system, no more guessing hex values
The result
A three-tier token system that lets engineers theme any new component without design input
Every design system decision was made with the same question: “Does this make the next screen faster to ship?” If the answer wasn’t yes, it didn’t go in., Design principle
Solution
Agent configuration cards, workflow status indicators, data extraction previews, built for fullscreen, sidebar, or embedded. Alongside the company’s first design system: 8px grid, semantic tokens, language & tone, motion docs.
What this shows
The 60+ component library, documented and ready for engineering
Why it matters
Components designed for AI-agent interaction patterns, not generic UI
The result
Engineers pulled components directly from Storybook without designer intervention
Figma Design → Figma Make / Magic Patterns → GitHub → Engineering. Prototype-to-production in hours when a customer demo needed it. Storybook gave engineers direct access, no design handoff wait.
AI-native design stack
Figma Design → Figma Make / Magic Patterns → GitHub → Engineering · Tracked in Linear
Validation
Three rounds of live demos with enterprise prospects, each building on the previous. I designed the experience but the CEO carried every conversation, I never talked to clients directly.
Every demo confirmed the same thing: once people saw the chat interface, the agent model clicked. The forms-based approach needed a guided walkthrough every time. The CEO didn’t need me in the room, the design spoke for itself., What the CEO reported back
Impact
Users orchestrated agents themselves. That self-serve UX became a core sales asset.
“This is so cool! It makes perfect sense to make these complex flows chat-friendly. I would have trouble knowing where to start.”, User during testing sessions
“Your design instinct is really strong, and that's hard to teach. The visual design combined with the UX… you did some really good work here.”
Reflection
I was the only designer, reporting directly to the CPO and CEO. The workflow was mine to own.
Design-to-engineering pipeline ownership
Linear tracked every move from Figma to production. Clear tickets, estimates, review cycles.
Rigorous design standards
The company's first design practice: 8px grid, semantic tokens, typography, spacing, language & tone, motion documentation.
AI-powered workflow tracking via Claude MCP
Claude MCP connected Figma to production in hours, not sprints.