← All episodes Podcast

# Ep #21: Cassie Groos (Unily): Design Systems for Agents, Infinite Brands, and Why the Canvas Stays

Jonny ·
Also listen onSpotifyApple PodcastsAmazon MusicDeezerYouTube

Tom and Jonny sit down with Cassie Groos, design systems lead at Unily, to learn what happens to a design system once AI starts reading it. They dig into the prose documents that brief her agents, unpack a token architecture built for an infinite number of brands, and go behind the scenes of her first big stage at Hatch.

Key takeaways

  1. AI now reads the design system as well as helping to build it. That was the thesis of Cassie’s Hatch keynote. A system built for agents ships with context, skills and, more and more often, its own MCP server, so someone can ask an agent a question about the system and get the answer back.

  2. Every component at Unily gets one prose document that its agents work from. Cassie calls it the CRD: a semi-structured document covering a component’s properties, requirements, behaviour and accessibility needs, including the details Figma can’t hold. It briefs the agent that builds the component and feeds its tests, and any agent using the component later reads it too. Send an agent to a polished documentation site like zeroheight instead and the human-friendly prose ‘blows out its context window’.

  3. Spell out what every designer already knows. Any designer knows the red button is for destructive actions. AI might decide purple suits it better, or that the brand red belongs on every primary button. Cassie reckons AI gets things right about 80% of the time, and more context plus repeated checks are how she closes the gap on the other 20%.

  4. Have one skill do the work and a second, stripped-back skill check it. Cassie’s contribution workflow, built so people outside the design system team can contribute, pairs a skill that holds the conventions with a review skill that has the same conventions and nothing else. Each check gets a little closer because the reviewer isn’t carrying the original context. She also runs a ‘Simplify prose’ skill over everything, borrowed from a friend and modelled on Simplify for code.

  5. Feature design is leaving Figma, and component design is staying on the canvas. Cassie expects product, feature and flow design to move out of Figma, with Figma Make catching some of it. Components for complex, multi-brand systems still need a visual canvas, and Figma variables remain the hub for Unily’s tokens. She is ‘not in the Figma is dead, everything lives in code now camp’, and says Figma has become far more responsive to the design system community over the last 18 months.

  6. At Unily, clients bring their own brand at runtime. Cassie calls it ‘infinite multi-brand’: any client can set up their brand through a self-serve journey, across light and dark mode (with high contrast to come), each with several colour schemes. Clients pick from presets, so spacing comes as spacious, regular or compact, and hundreds of tokens across several layers sit behind every choice. Unily’s AI assistant, Indi, then builds clients’ intranets from the system’s blocks and components with their branding already applied.

  7. Skipping the design system shows up in brand and trust. Agents can move fast without one, and a design.md file will still spin up slightly different buttons all over the place. Cassie calls ‘we don’t need design systems anymore’ a trap, and expects people to gravitate towards brands that still feel consistent and human-made. Hence the line from her keynote: design systems matter more now than ever before.

  8. With a CRD and a rough design, Claude can build five to seven components in a day. On her single-brand side project, Cassie asked Claude for a token architecture from a handful of colours, and it did ‘honestly a great job’. A workflow now helps her write each CRD, and the CRD plus a rough Figma design is enough for Claude to build the component. For a team of one or two, she’d start with buttons and fields, and for anyone who isn’t a design systems specialist, adapting an open-source system is perfectly fine.

  9. AI is good at going wide, and converging is still the designer’s job. Cassie has had Claude Design throw out ‘a million different design ideas’, but picking out three screens, spotting that there’s something there and perfecting them is work she does by hand, in a way she can never get Claude to. Perfecting that first flow and its components also stops AI taking the edge off later. She has watched Claude Design take someone’s design and make it ‘look a bit numpty’.

  10. Claude is now Cassie’s guide to a token architecture she built by hand. She says it ‘probably understands it better than I do now’, and the idea of going back in herself is ‘terrifying’. She asks it to tidy up anything clunky, such as the shadow tokens, and before she deprecates tokens it checks they aren’t used in any component, so the change won’t break anything.

  11. Apply for the talk you think you won’t get. Ben Callahan, who runs the Redwoods design systems community, tagged Cassie when Hatch’s Damian Martone asked for more European speakers, and she applied assuming she’d never be picked. She planned to learn the talk like a theatre script, then spent the five days before Hatch in hospital and walked on stage with a walking stick, too tired by then to be nervous. Her advice to anyone thinking about it: do it, because the audience is there to learn from you.

  12. Claude can draft the weekly Ocado order with a skill, a list and a few months of past orders. Cassie’s setup is a skill, a shopping list in Notion, a couple of months of previous orders so Claude can work out what she buys weekly, every two weeks or monthly, and Claude in Chrome to do the clicking. She’s holding out for Ocado to release an MCP.

Chapters

  • 0:08 — Welcome, and meeting Cassie through Hack Weeks and Hatch
  • 1:46 — What is a design system, anyway?
  • 4:04 — AI as both contributor to and consumer of the system
  • 5:52 — Jonny on agents that start from scratch every time
  • 6:47 — The CRD: one prose document per component
  • 8:56 — Telling AI what any designer knows, and its unpredictable 20%
  • 10:07 — Contribution workflows, a prose simplifier and review skills
  • 11:22 — Where the system lives: Figma, code, or both
  • 13:19 — Product design leaves Figma, component design stays on the canvas
  • 14:55 — Jonny on the design system as software
  • 17:16 — Design systems as a product, and feeling more like an architect
  • 18:28 — Unily’s infinite multi-brand problem, and presets
  • 19:46 — A token visualiser built for an audience of one
  • 20:36 — Indi, AI-checked token pairs, and nested modes
  • 22:41 — Tom on brand, marketing and the edges of the system
  • 26:58 — Several design systems, one token architecture
  • 28:38 — Jonny on consistency for its own sake
  • 30:35 — Why ‘we don’t need design systems anymore’ is a trap
  • 33:17 — A minimum viable design system for a team of two
  • 36:25 — From brand feel to code, and Tom’s Mostly Working canvas
  • 38:45 — Exploration as the human part, then one flow at a time
  • 40:55 — Tom on diverging and converging with AI
  • 43:51 — Using AI to untangle a hand-built token architecture
  • 47:18 — Getting on the Hatch stage
  • 50:46 — Why design systems are suddenly in demand
  • 52:37 — Quick-fire: getting Claude to do the food shop
  • 53:55 — Quick-fire: life after AGI, a four-tool stack, and Terminal
  • 56:26 — Where to find Cassie

Mentioned in the episode

Cassie Groos

Unily

Communities and events

People

Tools and products

About the pod

Mostly Working (formerly the Near Future Podcast) is a weekly conversation with design leaders about what they're building with AI. Send feedback to hello@mostlyworking.co, and find everything at mostlyworking.co.