While computer games have become more sophisticated, most computer characters follow a limited static script in their interaction with the user. We advocate a knowledge acquisition approach to building characters and use modelbased classification techniques as our learning mechanism. Model-based classifiers can be readily integrated with existing rule-based approaches for building computer characters. We demonstrate our idea using a simple simulated game of predator and prey, examine the types of rules created by the classifier and discuss how to develop adaptable characters. This approach facilitates creating and debugging characters and we believe it will lead to more interesting games.