“Instead of individual assets, you design systems.” Justin Jay Wang, designer at Cursor and before that the person who grew design at OpenAI from 0 to 1, on code as a creative medium for brand work: graphics, charts, posters, logos, animations. Define an idea as a set of rules and you can generate endless variations, dial in parameters, and converge on the strongest one.

Kaelig Deloumeau-Prigent, co-chair of the W3C Design Tokens Community Group, audited 20 open-source design systems and 6 platforms, logging 187 AI affordances and 157 techniques. 17 ship an official MCP server, 17 ship official agent skills, 14 publish llms.txt. Two ship Figma Code Connect: Carbon and Primer. Design systems are being rebuilt as machine interfaces right now, and the design side of that bridge is the part almost nobody has wired up. Code Connect is still Organization and Enterprise only.

“The failure is not in the components; it is in the composition.” Matt Rothenberg, Principal Engineer at Cloudflare, handed three isolated agents the same shadcn/​ui library, the same tokens, the same app shell, and three ordinary settings tickets. Back came tabs, cards, and a long form, all three of which a token linter would happily pass. He then gave the next three runs a named vocabulary and a short page of rules, and they finally read as one product. The agents didn’t cause the drift, they just made it impossible to keep pretending it wasn’t there.

Luis Ouriach makes the provocative case that when LLMs make code cheap to write, edit, and migrate, the contribution model and the component-as-source-of-truth lose their reason to exist. His point is that consistency was always carried by tokens, not components, so a well-defined token set could govern ten different LLM-generated inputs at once. He stops short of full obituary, asking whether components live on as the unit of constraint.

Luis Ouriach lays out a clear mental model for design systems at scale: before adding a component, ask whether you actually need a structural change or just a value override. The piece covers theming hierarchy across modes, Extended Collections, and direct token updates; when platform separation is worth the headcount cost (often it isn’t); and why you shouldn’t restructure your system for AI tooling until AI tooling can handle multi-file setups.

Colm Tuite, co-creator of Radix UI and its widely-used color system, pushes back on one of the most persistent design systems myths: that a ​“good” color scale means uniform chroma and lightness steps across hues. In perceptual color spaces like OKLCH, this is physically impossible — yellow simply cannot reach the chroma depth that blue can. But the deeper point is that even if it were achievable, it wouldn’t serve UI design. A palette that works needs purposeful contrast variation, not mathematical regularity.

Matt, designer at Lovable, writes up how they rebuilt their entire color system on OKLCh, derived from just three inputs: a neutral seed, a contrast value, and an accent. The technical choices are solid throughout — perceptually uniform ramps, WCAG contrast by construction, box-shadow borders that composite without layout shift, the state layer — but the most interesting design decision is the intentional reduction of token surface area specifically to make AI-generated UI more consistent. The fewer valid choices a model has, the less it can quietly go wrong.

Check Designs scans your file against your design system and flags four categories of drift: hard-coded color, text, radius, and spacing values that should be tokens; color contrast violations below WCAG 2.0 AA or AAA; components sourced from unsubscribed libraries; and fully detached components. Each issue comes with a one-click fix. Unfortunately, it’s only for Organization and Enterprise plans now — a big miss for many teams on Professional.

Luis Ouriach makes the case against single-number design system adoption metrics. His argument: one number collapses three things that should stay separate, across artifacts (a brand token and a complex data table component need different definitions of ​“used well”), surfaces (a logged-in dashboard component has no business on a sign-up screen), and people (a marketer, a senior product designer, and a front-end engineer all want different things from the same system). The throughline is that compliance with a benchmark is not the same as value, and most design system dashboards are quietly measuring the wrong one.

TJ Pitre’s Figma Console MCP is the most ambitious MCP server for Figma I’ve seen — 94+ tools spanning design-system extraction, file and FigJam and Slides creation, plugin debugging, and full-spectrum WCAG accessibility scanning across both design and code, including 13 design-side lint rules and code-side scanning via axe-core. Pair this with TJ’s Claude Design piece in the section above — it’s the same author making the case for what design-systems-aware AI tooling actually looks like.

“Claude Design can read a design system carefully when the prompt is about the system. When the prompt is about a composition that uses the system, it stops respecting the components and just generates lookalikes.” TJ Pitre spent a few hours testing Claude Design against two real design systems and concludes that the tool references your system, but doesn’t consume it. Claude would happily inline HTML tags with style props instead of importing them from your component library.

“The pitch for Claude Design’s workflow is roughly: I have a design system, I want to generate new product surfaces from it, and I want AI to do most of the lift. That workflow exists today. You can pair Figma with an MCP server like our Figma Console MCP, or with Figma’s own MCP server, or with Code Connect, and then point an AI app generator at it. Lovable, v0, Replit, Figma Make, Claude Code working inside your repo. Your Figma file stays the canonical source. Your codebase stays the production surface. AI does the generation in between.That flow is more linear, more honest about where source of truth lives, and it produces output that actually uses your component library, because the AI is operating inside the repo where the components live.”

FigmaLint is an AI-powered plugin that audits Figma components for design system compliance, accessibility, and developer readiness. It surfaces hard-coded values, naming issues, and missing interaction states, then helps you fix them directly — binding to design tokens, renaming generic layers, staging component properties.

Carina B. Velasquez, a Design Systems Lead locked out of AI tools at work, spent a Friday afternoon on her personal laptop with Claude Code and the Figma MCP server. By 6pm she had built Prisme: a complete three-tier token system, four brand themes, and a bidirectional plugin (Prisme Bridge) translating between Token Studio and native Figma variable formats. Her conclusion: ​“AI is amazing, Claude Code built my design system for me, the future is here. That is not what happened. What happened is that I showed up with years of accumulated knowledge, a specific problem I understood deeply, and a clear enough mental model to describe what I wanted with precision. Claude Code was fast; impressively fast, and smart in ways that genuinely surprised me. But it was executing a vision I already had. The intelligence in that Friday afternoon was not artificial.”

A thorough, practical guide from Christine Vallaure on what design systems need to work well with agents. Five things most systems are missing: semantic tokens, full state coverage, slot-based composition, exact Figma-to-code property matching, and token-based auto layout. Pairs well with the Brad Frost video — both cover the same design system moment from different angles.

Brad Frost and Chromatic cofounder Dominic Nguyen walk through how Storybook’s new MCP exposes design systems for agents to use reliably. The key insight: AI doesn’t invent states your components don’t have — it amplifies what’s documented. Coverage, completeness, and constraints are now load-bearing.

Luis Ouriach argues that modern design systems have become the centerpiece of software development in an AI-first world, shifting from a siloed support function to the primary mechanism that tells LLM-based tools what to build, and therefore directly shaping code quality, speed, and brand differentiation.

Luis on what the ​“shadcn-ification” debate is actually about — not the visual uniformity, but the organizational misread: ​“The mistake isn’t in the ingredient. It’s in thinking that having access to good ingredients is the same as knowing how to cook.”

Stakeholders are concluding that the design system infrastructure is done because a great foundation exists. The teams that have spent years practicing design systems understand exactly why that conclusion is dangerous.

Five practical strategies for integrating Slots into a design system, sourced from teams who’ve been testing the feature since Schema 2025. The most actionable tip is to start with the components people detach the most — that’s where Slots will deliver the fastest payoff. I also liked the tip of using pre-filled slots where content is predictable, and empty slots where designers need to add something on their own.

Code Connect keeps getting more ​“real-world design system” friendly: you can now attach multiple code files to a single Figma component, preview code across frameworks (with GitHub connected), and set framework-specific MCP instructions so agents implement the right thing. This is one of those unsexy updates that makes Code Connect actually usable across multi-framework orgs (or even just ​“web + native + marketing site”).

Once your design system is in Figma Make, you can really reap the benefits of working with design and code side by side and start actually using your system. This article walks through the specific technical problems of pulling a design system out of a monorepo to make it accessible in Figma Make.