This article discusses a new method for understanding user intent by breaking down interactions on mobile devices into two stages. By summarizing individual screens and then extracting intent from those summaries, small models can achieve results similar to larger models without needing server processing. The approach improves efficiency and maintains user privacy.