This study investigated the development of lexical complexity, sentence complexity, accuracy, and fluency in the English writing of 22 Chinese university students from the perspective of Complex Dynamic Systems Theory (CDST). Compositions were assigned 30 times over the course of one academic year through Pigaiwang, an online platform that automatically evaluates writing. A modified retrodictive modeling approach was adopted. Specifically, a longitudinal cluster analysis was used to examine emergent prototypes. A moving correlation analysis and retrodictive interviews were conducted to study the signature dynamics that produce each prototype. At each collection, the 22 student compositions were classified into two clusters. One cluster contained those students who performed better than average in accuracy, but worse in the other three variables. The other cluster comprised those students with the opposite performance. Students moved continuously between the two clusters; and their change trajectories can be categorized into three prototypes: a continuously stable type, an initially variable and then stable type, and a continuously variable type. Case studies of three students representing the three emergent prototypes indicated that the signature dynamics for the three prototypes were related to dynamic interactions among different variables and dynamic changes in affect-related elements in the form of writing interests, motivation, and strategies. The initial conditions and the feedback from Pigaiwang acted as key control parameters in shaping the prototypes. The continuously variable prototype developed their writing proficiency to the greatest extent and had the most variability. Based upon the findings, implications for teaching L2 writing are discussed.