I’ve been exploring how Pando can work with AI—from creating designs with natural language to checking work against the design system.
The goal is pretty simple: Figure out where AI can actually help without losing the human experience.
I’ve been leading the exploration of how Pando can work with AI—trying out different tools and approaches and working with our development team to connect AI directly to the design system.
A lot of the work has been experimenting with Claude Design, Claude Code, Figma, and Figma MCP, then testing what we get back against the actual Pando UI Kit to see what works and what doesn’t.
AI is quickly becoming part of how digital products are designed, built, and managed. Instead of treating it as something separate from our design process, I wanted to see how Pando could work with it.
The goal was to start bringing Pando into that conversation and figure out what it would take to make the system understandable and usable by AI tools.
I don’t see AI as a replacement for the designer or as something that should take over the experience. I see it as another tool, one that can help us work faster, explore ideas, and solve problems we might not have been able to approach as easily before.
But like any tool, you have to understand how it works before you can really use it.
For me, that meant learning how AI works, where it makes assumptions, how to give it the right context, and where human thinking still needs to be part of the process.
The goal isn’t to let AI own the experience. The goal is to use AI to help us create better experiences for real people.
Before AI can work with a design system, it needs to understand what the system actually is.
Pando isn’t just a collection of colors, components, and a UI Kit. There are connections between components, patterns, behaviors, standards, guidelines, and the design decisions behind them.
This raised a bigger question:
Could I teach AI to understand Pando well enough to actually work with it?
That became the starting point for the experiments that followed.
I started by taking the Pando Figma file and bringing it into Claude Design.
The first attempt didn’t work the way I expected. AI made a lot of changes and assumptions as it tried to fill in gaps where it didn’t understand the system.
Instead of getting frustrated and stopping right there, I started digging into how Figma, AI, and Claude Design worked together. I experimented with different ways of inputting my commands and explaining what I wanted Claude to do so it had a better understanding of Pando.
After one too many misfires and do-overs, I was able to get it to a point where it could start building designs that followed Pando much more closely.
Once Claude Design started understanding the system, I began testing it with actual commands.
For example, I could simply tell it: “Build a Content Call to Action.”
Claude Design was able to create the component along with the appropriate guidelines and interactions.The goal wasn’t just to have AI reference the design system. I wanted to see if someone could interact with Pando through natural language and have it create the right design and behavior.
The next part of the project took us in a different direction.
Working with our development team, I’ve been helping build an agent in Claude Code that can work directly with the Pando Figma file.
We originally created a collection of .md files to give the agent the context and rules it needed.
As we learned more, the development team found a better approach: connecting the agent directly to the Figma file through a Figma MCP.
That changed things quite a bit. Instead of giving AI a static set of information, it can now work directly with the actual source of truth.
The current workflow
Pando → AI → Designer Testing → Feedback → Pando → AI
We’re currently testing the agent against Pando components and any new work being created by the designers and developers.
This new agent can review new components and designs and check whether they’re following Pando’s standards and established direction.
Instead of relying on manual review, the agent can help identify where new work may not align with Pando and give designers and developers another way to check their work.
The designers are using and testing Claude Design, which lets us see where it understands Pando and where it still gets things wrong.
I also update the Pando Figma UI Kit every week. Those updates can then be brought into Claude Design so it stays up to date with the actual system.
That gives us a pretty simple feedback loop:
Pando
The system isn’t being handed over to AI. We’re figuring out how to make it work better with AI.
The biggest thing I’ve learned so far is that making a design system AI-ready is more than just uploading a Figma file.
AI will make assumptions when something isn’t clear. It needs to understand the rules, relationships, and how everything fits together.
A good design system needs to make sense to the people using it, but now it also needs to make sense to the AI tools we’re using to design and build.
And the more I learn about AI, the better I understand where it can help and where the designer still needs to make the call.
We’re continuing to test the Claude Code agent against Pando and see how much of the system it can understand and validate.
I’m also testing Claude Design with the Pando UI Kit as it evolves and learning where Pando needs to be clearer for AI to use it.
The goal isn’t just to make Pando work with AI. I want to see what happens when both people and AI can work with the same design system.
I’m still learning, so I wouldn’t present this as a finished product.
What I have right now is a working project exploring how AI can interact with Pando in two ways: creating with the system and validating against it.
More importantly, working through this has changed how I think about Pando itself. It’s not just a collection of components and guidelines anymore. I’m starting to see Pando as something AI can both understand, use, and help us and me evolve.