The emergence of Generative Artificial Intelligence (GAI) has caused significant disruption to the traditional educational teaching ecosystem. GAI possesses remarkable capabilities in generating human-like text and boasts an extensive knowledge repository, thereby paving the way for potential collaboration with humans. However, current research on collaborating with GAI within the educational context remains insufficient and the methods are relatively limited. This study aims to utilize methods such as Lag Sequential Analysis (LSA) and Epistemic Network Analysis (ENA) to unveil the "black box" of the human-machine collaborative process. In this research, 22 students engaged in collaborative tasks with GAI to refine instructional design schemes within an authentic classroom setting. The results show that the participants significantly improved the quality of instructional design. Leveraging the improvement demonstrated in students' instructional design performance, we categorized them into high- and low-performance groups. Through the analysis of learning behavior, it was observed that the high-performance group adhered to a structured GAI content application framework: "generate [right arrow] monitor [right arrow] apply [right arrow] evaluate." Moreover, they adeptly employed communication strategies emphasizing exercising cognitive agency and actively cultivating a collaborative environment. The conclusions drawn from this research may serve as a reference for a series of practical applications in human-machine collaboration and provide directions for subsequent studies.