My practice / AI
How I work with AI.
I use AI to design and build UI. I also design the experiences people have with AI.
Coding agents expand my ability to implement and test designs myself. Designing AI products draws on a connected set of skills: shaping agent behavior, defining human control, and evaluating whether the system helps people accomplish their goals.
01 / UI design & implementation
Take an idea into working UI.
I use ChatGPT, Claude Design, Claude Code, and Codex to explore product ideas, build interactive UI, and refine the details in code. That lets me work directly with navigation, application state, responsive layouts, and behavior across screens.
The scope varies from exploring a new product to extending a flow or refining an existing component. I direct the design and review hierarchy, spacing, interaction states, and consistency. A working experience gives PMs and engineers something concrete to evaluate with me.
02 / Design systems & playgrounds
Help AI work within a design system.
I start with the existing codebase: its components, tokens, patterns, and product behavior. I give agents guidance on how those pieces should be used, including which patterns fit a task and how components work together across a flow.
I’ve built a design-system playground where I can prompt, assemble UI using existing components, and iterate on the experience. It gives me a working environment to explore component choices, interaction states, and complete flows before carrying the design into the product.
Correct component usage takes more than a valid API. An agent can choose an available component and still miss an important state or assemble the wrong interaction. I review the resulting UI for component choice, composition, and behavior.
03 / Coding-agent harnesses
Build the system around the work.
I build coding-agent harnesses that combine project context, design-system guidance, decision records, automated checks, and review workflows. These systems guide how agents work within an existing codebase, check their changes, and preserve decisions across tasks.
I draw on established practices and adapt them to the project, adding guidance and checks where I find gaps. I give agents room to solve implementation problems. Requirements that can be checked reliably become automatic checks; decisions about product intent and the quality of the experience receive human review.
What I’ve learned: The workflow needs evaluation too. An earlier approach to recording project state became too burdensome. I simplified it into shorter context documents and targeted reviews, then tested the setup across different kinds of changes.
04 / AI product design
Design AI features and agent behavior.
When AI is part of the product, I design the interface and the behavior behind it: when conversation is useful, when an existing control is clearer, how generated results appear, and where people need to make a decision.
I design for uncertainty and failure alongside the successful path. Missing information, interrupted tasks, partial results, and recovery all affect whether someone can understand and trust the experience.
My human–agent interaction principles guide that work: keep people oriented, connect results to evidence, make consequential choices explicit, and preserve useful work when something fails.
Explore Incident Owner Copilot →05 / Testing & evaluation
Test the implementation. Evaluate the AI behavior.
For UI development, I use unit tests, smoke tests, and end-to-end regression suites, with automated checks running through GitHub Actions and CI workflows. I also test the interactions myself, reviewing the details and complete flows in the browser.
For AI features, I run evaluation suites with representative scenarios, edge cases, and examples of incorrect outputs. I use the findings to adjust the system and run the evaluations again, assessing task performance and UX quality alongside latency and cost.
Together, these help me assess whether the software works as intended and whether the AI behavior meets the project’s quality standards.
Define intended behavior
User task · Expected outcomes · Quality criteria
Run evaluation cases
Representative scenarios, edge cases, and examples of incorrect outputs.
Assess the results
Task performance · UX quality · Latency · Cost
Adjust the system
Interaction · Instructions · Constraints · Implementation
What this brings to a team
Design, implementation, and evaluation in one practice.
I move between product definition, UI design, working implementation, and evaluation.
I can help a team explore an AI feature, build an interactive version, examine where it succeeds or fails, and refine both the experience and the system behind it. I bring design judgment into implementation and technical understanding into design decisions.