The Polishing Layer

The State of AI Design 2026 dropped last week, and the headline is Claude overtaking ChatGPT among designers: 78% to 65%, a flip that looked improbable two years ago. The finding beneath it is the more interesting one. Figma↗, according to 900+ designers surveyed, has shifted from the primary design tool to "a canvas for quick exploration and polishing details as input for Claude Code↗."

That's a demotion, and it reads as accurate. Designers are now spending their most focused hours in Claude Design↗ and Claude Code↗, then returning to Figma↗ to clean up, align spacing, and hand off. The canvas became the finishing room. Figma↗ responded at Config 2026 by adding Code Layers, Motion, and Shaders↗, leaning into last-mile value rather than fighting to reclaim the primary slot. That's the right call.

Meanwhile, two tools this week take the conversation somewhere less charted.

GlassKit UI is a 44-component React library for Meta Ray-Ban Display apps. The display is 600x600 pixels, sits in your field of vision, and you navigate it with D-pad focus rings and Neural Band gestures. The library is MIT-licensed, you vendor the components directly with no lock-in, and it handles the full interaction model: single-task interfaces designed for a two-second glance, with no hover state and no touch input.

It's early (Meta's publishing platform isn't fully open yet), but it's the first serious design system infrastructure for wearable displays that doesn't come from Meta. What strikes me about GlassKit is how thoroughly it breaks the assumptions designers have built up over twenty years of screen work. There are no artboards, no hover states, no responsive breakpoints. The display is 600x600, it's on someone's face, and the interaction model is a D-pad. That's closer to designing for a 2003 Nokia than anything since. The constraints force a kind of thinking that most design tools actively discourage.

UISqueezy solves the unglamorous problem that grows as AI generates more UI: token drift. It syncs Figma↗ variables with your codebase two ways, automatically: push tokens to Figma↗, pull changes back, no copy-paste, no spreadsheet. As AI-generated components multiply, the one-source-of-truth problem compounds fast. UISqueezy isn't the exciting problem to work on, but it's the one that catches you in production.

That matters more than usual right now. Another number from the State of AI Design report: 50% of respondents have pushed AI-generated code to production, including brand designers rather than just engineering-adjacent roles. The handoff between AI-generated design and shipping code now runs at a pace where manual token management fails regularly. UISqueezy exists because that's become the actual bottleneck.

The stack has reorganised. Claude is doing the generative work, Figma↗ is the finishing layer, and the tools around them are getting more specific about what problem they solve. Each of them is narrower than designers would have expected two years ago, and more useful for it.