The speed of AI-assisted design comes from clear ownership, not automated decision-making. The AI does not set the direction. It expands options, surfaces research, and provides initial drafts. Every choice, pivot, and override remains entirely mine.
The split is deliberate. AI is good at the parts that scale with volume. It is not good at the parts that require having sat in the room with a user, or at being answerable for a call.
Claude works inside my tools, not next to them.
Through MCP (Model Context Protocol) connections, Claude can read from and act in the same tools I use. It doesn't describe a design in chat for me to recreate. It builds the design on my canvas, where I review, comment, and take over by hand when I want to.
The design canvas. Claude creates and edits screens directly, using variables synced from the token file.
A library of real app screens and flows. Claude searches it for proven patterns, like filter sheets, booking flows, and onboarding quizzes, to shape structure. Visual style stays mine.
Works directly in the project repo: writes the brief, tokens, strategy docs, and decision log, and later builds the app from the approved designs.
A handful of shared files make it work: the brief, the tokens, the strategy docs and the decision log all live in the project repo, so every session starts from the same place.
The classic process still holds. AI changes how much ground each stage covers, not who is responsible for it.
Design rarely follows a straight line. Because generating variations with AI is fast, looping back to an earlier stage whenever something feels off carries very low overhead. If prototyping exposes a flaw in the business logic or workflow, the strategy gets reworked without sacrificing weeks of manual polish.
Understand user needs, motivations, and behaviors.
I define target personas and user archetypes.
The AI runs parallel desk research across competitive landscapes, pattern libraries, and marketplace models to compile notes on standard behaviors.
AI desk research supplements user empathy. It never replaces direct observation or interviews with real people.
Frame clear problem statements and outline core workflows.
I draft the primary user problems and prioritize core features.
The AI suggests secondary problems, maps user jobs to concrete workflows, and helps pair each problem with a measurable success metric.
Explore a wide range of solutions before committing.
I establish the strategic boundaries and aesthetic constraints.
The AI generates multiple distinct visual directions, color systems, copy options, and monetization approaches side by side for direct comparison.
Translate concepts into tangible, testable designs.
I keep a centralized token file in code as the single source of truth, mirrored directly into design tools, then review and refine each screen as it lands.
The AI builds out interface screens, component states, and complete user flows based on those tokens.
Evaluate quality, catch edge cases, and refine the experience.
I conduct continuous critique on screens as they are generated.
The AI runs automated checks against the initial brief to surface missing states, inconsistent patterns, or unaddressed personas.
AI critique catches spec gaps. Real-user feedback dictates actual product validation.
These exist because the failure modes are predictable. Left alone, AI is fast, fluent, confident, and occasionally wrong in ways that are expensive to find later.
For customer-facing decisions, I require multiple variations side by side, so I am choosing, combining, or overriding rather than rubber-stamping a single output.
AI provides velocity and breadth. Decisions about brand tone, navigation hierarchy, edge-case handling, and user tradeoffs rest entirely with the designer.
The same method applied inside a real product team, not only on personal projects.
At Meta I worked with role-based agents for PM, engineering, architecture, and QA to pressure-test PRDs, run gap analysis on my own designs, and sequence a build. It let me stress-test a direction before spending anyone else's time on it.
I build a hidden area inside the real product environment, so options get explored against real components and real data rather than in a sandbox that flatters them.
Front-end changes went into the product directly rather than being handed off as static mocks. No gap between a mockup and something the team could put in front of a real budget owner.
The core design process hasn't changed. What changed is the speed and depth of exploration before committing to a direction, and the clarity with which strategic decisions are documented.