This paper presents part of a work aiming at building a tool for the detection of graphomotor difficulties involving disorders in the writing of children. We have defined an experimental protocol, containing exercises, like copying figures or writing sentences under different conditions. It allows to measure simple aspects of graphomotor skill up to complex ones. A great number of features were obtained from on-line children productions. We focus here on the method we used to select low-level features that can describe the automation level of graphic activity. It is based on hierarchical clustering of features and a sequential forward selection. Every exercise is represented by two relevant features at least. We show that, in most cases, the selected features allow to recognize the school level of children having a regular schooling but to discriminate children with scholar difficulties as well.