The CellML language was developed in response to the need for a high-level language to represent and exchange mathematical models of biological processes. The flexible structure of CellML allows modellers to construct mathematical models of the same biological system in many different ways. However, some modelling styles do not naturally lead to clear abstractions of the biophysical concepts and produce CellML models that are hard to understand and from which it is difficult to isolate parts that may be useful for constructing other models. In this article, we advocate building CellML models which isolate common biophysical concepts and, using these, to build mathematical models of biological processes that provide a close correspondence between the CellML model and the underlying biological process. Subsequently, models of higher complexity can be constructed by reusing these modularized CellML models in part or in whole. Development of CellML models that best describe the underlying biophysical concepts thus avoids the need to code models from scratch and enhances the extensibility, reusability, consistency and interpretation of the models.