Paper #53: User Engagement With Intermediate Representations in LLM-Assisted Audio Programming

We present MaxPyStudio, a web-based chat interface that generates Max for Live audio devices through MaxPyLang, a human-readable Python intermediate representation of Max patches. The system surfaces two views of every generation—a signal flow patch graph and the MaxPyLang source—each gated behind an explicit toggle, making view engagement an observable act. In an exploratory study of 52 users producing 151 generations, we find that users preferentially reach for the graph view (opened by 55.8% of users vs. 38.5% for code) and engage with it more deeply (24.1% glance rate vs. 64.7% for code). A counterbalanced layout manipulation confirms this preference is robust to toggle position (Fisher's exact p = 1.0; pooled graph-first rate 80.8%, p < 0.01). When generations fail, users reprompt rather than inspect the intermediate representation, with view rates 3.9× higher after successes than failures.