The aim of this paper is to present a knowledge-based approach to supervised classifier design. For this purpose, an expert system has been built following the Common KADS methodology. Classifier design is seen as a general design problem and a modified version of the well-known Propose-Critique-Modify method is proposed as a suitable strategy to solve it. In this context, a number of heuristics are used to shrink the search in the space of possible designs. Although the system is evaluated on different datasets, special emphasis is put on a particular problem: Alzheimer’s diagnosis.