As human speakers naturally adapt their linguistic styles to one another, voice user interfaces that prompt similar linguistic adaptations can augment human-like interaction. In this study, we leverage a corpus of human instructions to model the effectiveness of incremental instruction generation in artificial agents. Participants interacted with agents that guided them in selecting virtual puzzle pieces, varying the amount of information provided in each instruction. Through an empirical examination of the Gricean maxims in utterance construction, our initial perception study highlighted the significance of adaptive instruction generation. By employing mouse movements as a proxy for user understanding, we developed computational models that enabled agents to detect user uncertainty and refine instructions incrementally. Comparing speaker-based and listener-based models, we found that agents encouraging linguistic adaptations were preferred by users. Our findings offer new insights into the value of mouse movements as indicators of user comprehension and introduce a methodological framework for developing adaptive interactive systems that generate instructions dynamically.