"Check if my app has accessible colors." This was my first prompt to Claude to ensure compliance with the European Accessibility Act. The results were catastrophic: hallucinated colors, incorrect contrast calculations, and no UI interaction.

In this talk, we'll follow the journey from a naive "take a screenshot and ask Claude" prototype to a system capable of performing meaningful accessibility color audits on React Native.

Each failure taught us something new. We'll explore the technical decisions that transformed a toy experiment into a practical auditing system, from understanding color accessibility and the surprising complexity of contrast analysis to extracting information from screenshots and accessibility trees.

Although our case study is accessibility color auditing, it applies whenever we need deterministic result from an AI, including end-to-end testing, UI validation, and automated quality analysis.