When we started building Visible over 4 years ago, we decided to optimize for time to market. Without LLMs it mostly meant cutting corners, building fast and iterating on live features. As we matured, we were introducing new patterns but the old code would still deliver value.

However, in 2026, now with tens of thousands of daily users, when the implementation can be optimized for AI agents, the variety of patterns and legacy code does not make things any easier. Agents get lost in complex logic and often invent new patterns. Skills don’t really help keep it in line. We needed a different approach.

In this talk, I'm going to talk about how we approached making our fairly advanced codebase more accessible to AI, and in turn also to human developers. How did we rework our design system to be less verbose, how did we make verification of changes semi-automatic (using in-house QA agent, and converting mobile app to desktop), and where we had to invest more time than we'd like to admit.