Full-cycle product studio for the AI-native era. Design systems, UX, AI, and full-stack engineering for growth-stage and enterprise teams.
USA
Joined January 2010
- Tweets3.2K
- Following534
- Followers1.1K
- Likes5.2K
Prototypes were never the point. Learning whether the idea works was.
AI tools like Figma Make and Claude Design are shrinking the gap between a concept and real user feedback.
@sweetdesignman shares his thoughts on where the design process is heading within product teams.
The most useful thing in a design system isn’t the components. It’s the rules for when not to use them.
Anyone can build a button library. What makes a system work is the judgment on top: which piece fits which situation, and which one never should.
One example: most apps have a small pop-up that slides in and disappears after a few seconds. It’s great for “Saved.” It’s the wrong choice for an error, because people miss it.
Designers carry dozens of rules like that in their heads. AI agents don’t. They pick whatever looks close enough.
So write the judgment down.
For every component:
- When to use it.
- When not to use it.
- What to use instead.
Then give your AI agents those rules, so they make the same call a designer would.
@sweetdesignman, our Head of Design, breaks it down in this clip.
Your sales team has Claude Code now. So does finance.
AI tools are landing in the hands of people who’ve never opened a design system. Sales reps are building their own dashboards. Analysts are spinning up internal tools. That’s a good thing.
The catch: the same prompt can give you two different answers. One follows your design system. The other just looks close.
A designer spots the difference instantly. Everyone else sees two screens that look fine. So the close-enough version ships, then the next one does, and your product slowly drifts off brand.
A design system isn’t just for designers and developers anymore. It’s for anyone building with AI, and it has to live where they’re working: inside the agent.
@sweetdesignman, our Head of Design, explains in this clip from our webinar on design system MCPs.
Your agent's design system context comes from three places: tokens, components, and the mapping between Figma and code.
@timgremore on why we bake all three into the MCP at build time instead of digging for them at runtime: headway.io/events/design-sys…
No developers? No repo access? You can still give AI your design standards.
@sweetdesignman on when to start with a Design MD file and when to move up to a design system MCP:
headway.io/events/design-sys…
The foundation of every design system MCP we build today. (As we know it today)
4 principles, 6 tools.
Principles
1. Fewer tools, richer params
2. Descriptions are prompts
3. Resolve, don't dump
4. Return what the agent can write
Tools
1. Resolve tokens
2. Search tokens
3. Get component
4. Search components
5. Validate
6. Get Figma mapping
@timgremore walks through what each one does:
headway.io/events/design-sys…
We don't use Figma Code Connect.
It's one more thing to maintain, and it only lives in Dev Mode. If your developers don't all have a seat, they never see it.
An MCP server gets your UI kit's context to an agent without that overhead.
Thoughts from @timgremore @sweetdesignman
Your design system has users who can't install your React UI kit.
Internal tools on another framework. Decks. Teams with no dev to wire it up. They ship work with your name on it either way.
@timgremore on why a DESIGN.md gives those people something to point an agent at:
headway.io/intent-and-craft/…
Underrated AI guardrail: the anti-patterns section at the bottom of your design system docs.
Turn those into "when not to use" rules in your MCP and stop re-prompting your agent for the right component.
headway.io/events/design-sys…
Headway retweeted
If you've liked the care in my Omarchy tools, that's the same care my team at @headwayio puts into client work.
Got something worth building? 👇
jonkinney.com
Headway retweeted
Want to know what we think about agentic development at @headwayio?
"My vibe coded app works great! Except the problem is it has to download 15mb of JavaScript before it can run..."
headway.io/even-keeled/how-o…
I re-listened to this today. Everything and nothing has changed.
Your sales team has Claude Code now.
Same prompt, two answers, and no way for them to tell which one breaks your design system.
Here's where a design system MCP comes in:
headway.io/events/design-sys…
We run a DESIGN.md and an MCP side by side.
A DESIGN.md won't hand your agent the props on a React component. An MCP will.
One leans design. One leans development. Picking between them is the wrong question.
@timgremore on Intent & Craft:
headway.io/intent-and-craft/…
Headway retweeted
Looking forward to arriving in Austin tomorrow for
@railsworld! A few of us Headway folks will be in attendance. Will I see you there?!
Your AI agents aren't ignoring your design system. They just can't see it.
Today at 1:30 pm CT, @sweetdesignman and @timgremore go live on design system MCPs: what they are, how to stand one up, and what they still don't solve.
Bring your questions.
headway.io/events/design-sys…
"Won't a custom MCP burn through tokens?" came up in the Q&A last time.
Short answer: run it locally. No hosting, no auth, same quality result.
Next live stream with @timgremore and @sweetdesignman is Tuesday, September 22 at 1:30 PM CT. Bring your questions: headway.io/events/design-sys…
Same screen, built twice. Once with Figma's MCP alone. Once with a custom design system MCP alongside it.
Design system informed, versus invented in response to your prompt.
@timgremore walks through the results.
Join us for our next live stream to ask your questions about custom MCPs for design systems with @timgremore and @sweetdesignman.
This Tuesday at 1:30 PM CT: headway.io/events/design-sys…
Five levels of design system readiness:
1. Ad hoc
A Figma file people copy and paste out of.
2. Catalog
A library exists, components aren't composable, code is disconnected.
3. Structured
Figma and code aligned, tokens in place, some governance.
4. Validated
Rules that fail CI, not rules that fail review.
5. Agentic
AI produces shippable UI, humans review the exceptions.
Most systems we look at sit at two or three.
Which one is yours?