<p>Lesson plan refinement is cognitively demanding in teacher education, yet university-based lesson planning often lacks enactment-based feedback and relies heavily on peers with limited classroom experience. Generative AI can offer scalable support, but its educational value likely depends on guided use. This mixed-methods, course-embedded study examined how structured ChatGPT prompting shapes preservice teachers’ cognitive engagement during collaborative STEM lesson plan refinement. Twelve preservice teachers (six dyads) refined existing lesson plans with an embedded GPT‑4 bot and prompt templates aligned with lesson plan analysis and lesson planning. Data included 877 coded utterances from group discussion, utilization coding of 80 ChatGPT responses, and semi-structured interviews. Epistemic Network Analysis and thematic analysis showed that AI-related commentary was closely linked to reflection, creation, and summarizing, and that adaptive modification and continuing inquiry dominated feedback uptake. Template-based questions generated denser connections between lower- and higher-order cognitive processes, whereas no-template questions centered on information requests. Analysis- and planning-oriented templates supported complementary engagement patterns across refinement cycles. Findings highlight guided prompt templates as practical scaffolds for fostering reflective and iterative use of AI feedback in lesson plan refinement.</p>