This paper proposes an approach to decompose large class diagrams. It first collects metrics of coupling among classifiers (classes and interfaces). According to the principle of high cohering and low coupling, it breaks low coupling classifiers while showing high coupling classifiers in the same diagrams. To guarantee that the generated new class diagrams are readable, it confines sizes of new diagrams to a predefined scope. The results of its evaluations on industrial projects suggest that the approach is practical and valuable. The approach proposed in this paper helps to improve the readability of software models.