What does an AI-ready design system look like in practice? Kaelig Deloumeau-Prigent walks Brad Frost through his updated State of AI in Design Systems report (first covered in #258) and singles out Meta’s Astryx for running nightly evals to check that its skills produce good outcomes. Then they run the audit on Eddie. Brad’s takeaway: the tooling surfaces the gaps, and people still have to decide what matters.

Blair Metcalf, principal UX engineer at Reddit, demos Reddit Proto on How I AI. It’s an in-house prototyping harness built on the design system, used by 100% of Reddit’s UX and product org, with about 700 prototypes live. The clever part is real data. Prototypes can render a signed-in user’s actual feed, because leadership dismisses mocked screens with ​“This isn’t relevant to me.”

The CSS Working Group resolved to give buttons, selects, checkboxes, and radios cursor: pointer in the new appearance: base form controls. The decision is partly based on Nicole Sullivan’s analysis of design systems: ​“27 of the 33 systems that specify a cursor use pointer.” The old ​“pointer is for links” debate is back in the comments, and both sides make good points.

A book by Andrew Marcus about interface design from the Xerox Alto to the Apple Vision Pro, with one chapter per component: buttons, checkboxes, text fields, menus, dropdowns, down to their anatomy and states. It has about 500 illustrations drawn from 180 well-known apps. It’s also a practical guide to building a design system.

High contrast mode in Notion came out of an accessibility audit and focuses on foreground content: borders, text, and icons, applied systematically so every page still feels considered. Tamara Chu, who worked on it, wanted it to be beautiful too, and feedback during the work compared it to the no-fluff UI of the 1984 Macintosh.

Mentha is Mintlify’s design system for an agent-heavy team. Fay Dakrouri’s takeaway: the more instructions, skills, and component constraints you feed an agent, the better its first direction, though someone still has to judge whether the result solves the problem tastefully.

Guidelines let a design systems team upload Markdown rules to a published library, and the Figma agent reads them every time someone prompts with that library. Include dos and don’ts, anti-patterns, and unwritten taste. When the agent keeps inventing navigation, add a rule.

At Config India, Figma also put code sources in closed beta and previewed one place to manage a whole system.

Josh Puckett launched Graphical, ​“a simple, powerful, and fun tool to design visual languages, style components, and work with coding agents to create interfaces that look memorable and feel unique.” You tune colors, type, spacing, shape, and motion springs visually, then hand your agent a GUI​.md, theme references, and three skills for building and reviewing screens.

DS Manager adds an MCP server so agents can edit a shadcn/​ui design system without breaking the rules you locked. Handy if your team builds on shadcn and the shadcndesign Figma kit.

Evil Martians’ eight-part framework for keeping AI-built interfaces consistent. Component contracts say when to use and not use each component, an inventory lets agents shortlist components before reading any code, and a design linter fails the build on a raw hex color. The agent skill is published too, split into craft and use workflows.

Anton Lovchikov, Head of Design at Evil Martians, makes the case that ​“a designer’s main job isn’t drawing screens. It’s pulling rules out of chaos and making sure they’re followed.” His proof is a 700-line localization refactor, after which Claude rewrote half the strings. The fix was a three-layer system of product principles, copy rules traced back to them, and skills built on those rules. Now when he dislikes a string, he asks Claude which rule produced it and fixes the rule.

“Benchmark and read the transcripts” – the most useful lesson in Atlassian’s write-up on building a CLI for its design system. The dashboards told them whether the agent succeeded. The transcripts showed agents ignoring carefully written instructions and reading source files instead. Charlie Sutton, Atlassian’s Chief Design Officer, adds that the public ADS MCP and CLI together cover 90%+ of atlassian.design.

Anthony DiSpezio, Designer Advocate at Figma, builds a Make kit from an npm package. Make auto-generates a starting version of the guidelines, and Anthony offers a few rules for making them more effective – assume the agent knows nothing, split guidelines into categories so it doesn’t drown in context, include real usage examples, and spell out the gotchas. Before publishing your kit, calibrate against the model’s pre-training biases and test how it actually reads your UI.

“So a good rule of thumb is that semantic tokens do not include a scale.” Donnie D’Amato gives three tests for whether a token is truly semantic: it doesn’t suggest its value, it describes a generic family of components rather than a single one, and it doesn’t depend on its neighbors. That last test is the ​“smedium” problem, and it quietly disqualifies ​‘space-sm‘ and ​‘space-md‘, which is to say most of the space scales shipping today.

His recommended token construction: ‘[purpose]-[priority]-[property]‘. Priority is optional and capped at three levels, because ​“having more than 3 levels is useless for user interpretation within the experience.” The test underneath is the useful part. If you can’t point at the element in a low-fidelity wireframe, it isn’t semantic. That’s why ​‘surface-primary-shadow‘ earns its name and ​‘surface-highest-shadow‘ doesn’t.

“Names that hold up describe the decision a designer is making about the interface. Elements get renamed, colors get rebranded. The decision persists.” Murphy Trueman, Director of Design Systems at Baseline Design, calls Donnie’s piece above one of their favorite things written on the middle tier, then turns it into a working method. Walk a wireframe and ask what a designer is deciding at each element. A ​“card” is really ​“this is an elevated surface for grouped content”, so the token becomes ​‘Surface/​Elevated‘ and it survives the rename to ​“panel” next quarter.

“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.

“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.

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.