The problem
ERGO Group AG acquired Next Insurance, and the product needed a full rebrand onto ERGO's global design language. The existing design system was in rough shape: unmaintained for years, so out of date that designers detached components just to make them work, and no documentation on how to use each component. Seven platforms were using components that looked and behaved similarly, but the code was different and impossible to re-use, all on an aging tech stack.
The old system ran on implicit knowledge. A designer was expected to know the design system well enough to pick the right component every time, and design critique sessions would catch mistakes, but that process doesn't scale for a growing organization, and it completely breaks when using AI coding agents for prototyping or design.
An agent also raises the cost of getting this wrong. It treats whatever it finds in the system as correct and reproduces it in everything it builds, so an inaccuracy that used to mean one mismatched button somewhere now gets copied into every page the agent touches.
My role
I built the new design system with guidance from the design director: Implemented a rebrand in 7 different platforms, created a design system from scratch to mirror an existing brand, and a parallel migration of the legacy code to React, working with rotating engineering teams across the platforms. This project had a harder goal than a normal rebrand because the primary user would be an AI coding agent, not just a human designer, and it needed to be launched in a few months.
AI agents don't understand intent, they read structure and metadata. If you hand them two components that look alike but mean different things, like a button and a near-identical label, they will confuse the AI agents. So I encoded the intent directly into the system by using purpose-driven names with semantic tokens, intent metadata on every component, and an exact 1:1 mapping from each Figma component to its real counterpart in code. So the agent understands the intent from documentation instead of guessing.
That documentation lives in the repo and in Storybook, next to the code, because that is what the agent actually loads. An agent starts every session knowing nothing about your conventions, the way a new hire does, and a separate documentation site might as well not exist for it.
