Keeping Pace with AI: Building LUNAi
An AI design assistant that kept evolving as the tools changed around it
Role: Concept, UX, prototyping & knowledge architecture
Team: AI x Design, Northwestern Mutual
Timeline: March 2026 – present
Tools: Figma, Figma Agents & Skills, VS Code, GitHub Copilot, YAML, Markdown
Example of LUNAi being used right in the Figma canvas
The Starting Line
Our design knowledge was scattered.
Guidance for Luna, our design system, lived across multiple websites, files, and team conversations, along with our content standards and established patterns. Getting an answer usually meant leaving Figma to dig through documentation or ask a teammate.
With AI showing up everywhere, that felt like a barrier we didn't need.
What if designers could get answers without ever leaving the canvas?
That question became LUNAi, an AI-powered Figma assistant that brings design system knowledge into the workflow. It started by answering questions and grew into a tool that generates Luna-compliant UI, offers guidance, and helps designers stay aligned while moving faster.
This case study follows how LUNAi evolved as tools like GitHub Copilot and Figma AI changed what was possible.
On Your Mark · March 2026
LUNAi didn't start as an AI assistant.
The first concept was a design review companion that would flag issues and leave annotations on the canvas before designers shared their work. While exploring the problem, I noticed designers spent as much time searching for information as reviewing designs. Instead of tackling both at once, I scoped an MVP around access to design knowledge and kept AI-powered review as the long-term vision.
The MVP had two main features:
Ask LUNAi Prototype
Ask LUNAi
A conversational interface where designers could ask questions about Luna, content guidelines, and design patterns directly within Figma.
Smart Search
A faster way to find the source of truth, linking designers directly to the relevant documentation instead of searching across multiple resources.
Smart Search Prototype
The hypothesis was simple: if design knowledge lived in the canvas, would designers use it?
Get Set, Go... Find an LLM
With the MVP taking shape, I turned to what I assumed would be the easy part, adding the intelligence.
Around then, I was invited to present LUNAi at an internal AI Design Lab to designers, design leaders, and others exploring AI across the company. The response was strongly positive. People saw the value right away.
The open question was how to get an approved LLM behind it. I explored options including local models, and reached out to an engineer after hearing him speak about them at a company-wide AI session. His read was honest: my approach would take months of approvals, funding, and stakeholder buy-in, with no guarantee it would happen at all.
AI tools were moving so fast that I expected a better path to open up soon. The goal had never been a specific technology. It was making design knowledge easier to reach. So I went looking for another route.
Changing Lanes · July 2026
After moving away from the original LLM approach, I started exploring the tools already available to me. Around the same time, my manager challenged me to think about how designers could work beyond the Figma canvas, which pushed me to rethink what LUNAi could become.
Using VS Code and GitHub Copilot, I reorganized Luna knowledge into a structure that was easier for both AI and humans to navigate. Engineering resources, design knowledge, and UI definitions each had their own space, with YAML helping make the system more approachable for designers without coding experience.
My goal was simple: create a repeatable setup where designers could use a single prompt to generate prototypes, explore interactions, and reference Luna guidance without needing to understand the underlying technology.
To test the concept, I asked a designer with little coding experience to use the workflow on their own. Using a prompt connected to SharePoint documentation, markdown knowledge files, and the YAML structure, they were able to navigate the system and generate useful outputs with minimal guidance.
The biggest challenge was generating YAML from scratch, but the feedback was clear: it felt far more approachable than working directly with HTML or JavaScript.
The experiment reinforced an important lesson: the value wasn't the AI itself. It was organizing knowledge in a way that made it easy to access and use.
Example of a more designer-friendly file structure
Example of a designer-friendly YAML file that can be edited without coding, with changes reflected throughout the front-end experience.
Final Lap: Figma Agents · September 2026
While I iterated, Figma was moving too. Its first agents felt limited, but later updates added Skills, which let agents be grounded in domain-specific knowledge.
The work I'd already done structuring Luna knowledge carried over directly. As a Figma agent skill, LUNAi can now:
Generate Luna-compliant UI
Answer design system questions
Surface content guidance
Annotate designs with recommendations
Create new frames grounded in established patterns
That annotation feature brought back the design review idea from March.
The question shifted from how do we get AI into the workflow to how do we make it useful by giving it the right knowledge.
Cooldown Lap: A Reflection
LUNAi changed shape several times: design review tool, design system assistant, knowledge layer for Copilot, and now a Figma agent skill. Each change came from rethinking the solution while keeping the same problem in focus: helping designers find answers and make better decisions without leaving their workflow.
What I took away:
Adapt the solution, not the problem. In AI, the finish line keeps moving. Staying effective means continually testing and rethinking what's possible.
Knowledge quality determines AI quality. The best AI experiences aren't defined by the model behind them, but by the knowledge they make accessible.
Keep humans in the loop. AI-generated work should be reviewed and validated, never trusted blindly, especially for accessibility, content, and design-system standards.
LUNAi isn't meant to replace designers. It takes on repetitive lookups and small interruptions so designers can focus on bigger, more meaningful problems.
What's next:
LUNAi is moving from prototype to team-ready tool. The next lap focuses on three things:
Making it sustainable. As Figma introduces pricing for agents, we're evaluating cost against value to decide where AI assistance earns its place in our workflow.
Focusing on what matters most. Working with the design team to prioritize the use cases that save the most time, rather than trying to do everything at once.
Earning trust before scaling. Partnering with content and design system stakeholders to review LUNAi's outputs, fix gaps in the skills, and refine them until they're reliable enough for the whole team.
The finish line is still moving, but the goal is the same: a tool designers trust enough to use every day.