There are two conversations about AI and interfaces happening right now. One is about AI features in the UI. The other — the one that changed my daily work more — is about AI building the UI. I want to talk about the second, because most of what's written about it is either hype or fear, and neither matches my experience shipping production React.
My current pipeline
My stack is React, TypeScript, and Tailwind, with Figma as the design source. The AI-accelerated version of my workflow looks like this:
- Design in Figma, but structured. Components named properly, text bound to a small set of type roles, colors bound to variables. This discipline predates AI, but it pays double now: an LLM translating a well-tokenized frame produces dramatically better code than one guessing at a soup of raw hex values.
- Automate the Figma side. I drive Figma programmatically — batch-editing copy, binding nodes to design tokens, auditing frames for unstyled text. Tasks that used to be an afternoon of clicking are a script. Tokenization audits that nobody ever did manually now run on every major frame.
- LLM-assisted design-to-code. With Claude Code, I hand over the frame context and the design system constraints, and review the component it writes. Emphasis on review. The model is a fast junior with encyclopedic Tailwind knowledge and no accountability. The senior judgment — is this the right abstraction, does this handle the empty state, will this break at 47-character labels — is still my job.
Where it genuinely enhances the interface
The speed isn't just convenience; it changes what ships. Three examples from recent B2B work:
- More states, actually built. Loading, empty, error, partial-data. Historically these got designed and then quietly dropped under deadline. When a state costs ten minutes instead of two hours, it ships.
- Real responsive behavior. I test three width tiers as a rule. AI-generated first drafts get me to "working at all breakpoints" fast enough that I spend my time on the judgment calls — what collapses, what truncates — instead of the plumbing.
- Faster iteration with stakeholders. In a fintech project, replacing static mockup reviews with clickable coded prototypes cut an entire round of "I didn't realize it would work like that" feedback.
The discipline that makes it safe
One rule I hold hard: whoever writes doesn't review. AI-generated code goes through a review pass by a different model or a human — ideally both — before it merges. Cross-checking catches the confident nonsense that any single model produces. I've caught accessibility regressions, phantom props, and subtle state bugs this way that I would have missed skimming my "own" output.
The other rule: nothing is done without evidence. "The component should work" isn't a status. A screenshot at three breakpoints, a passing test, a keyboard-only walkthrough — that's a status.
AI didn't lower the bar for interface quality. It removed the excuse for missing it.