Personalizable software agents will learn new tasks from their users. This implies being able to learn from instructions users might give: examples, yes/no responses, and ambiguous, incomplete hints. Agents should also exploit background knowledge customized for applications such as drawing, word processing and form-filling. The task models that agents learn describe data, actions and their context. Learning about data from examples and hints is the subject of this paper. The Cima learning system combines evidence from examples, task knowledge and user hints to form Disjunctive Normal Form (DNF) rules for classifying, generating or modifying data. Cima's dynamic bias manager generates candidate features (attribute values, functions or relations), from which its DNF learning algorithm selects relevant features and forms the rules. The algorithm is based on a classic greedy method, with two enhancements. First, the standard learning criterion, correct classification, is augmented with a set of utility and instructional criteria. Utility criteria ensure that descriptions are properly formed for use in actions, whether to classify, search for, generate or modify data. Instructional criteria ensure that descriptions include features that users suggest and avoid those that users reject. The second enhancement is to augment the usual statistical metric for selecting relevant attributes with a set of heuristics, including beliefs based on user suggestions and application-specific background knowledge. Using multiple heuristics increases the justification for selecting features; more important, it helps the learner choose among alternative interpretations of hints. When tested on dialogues observed in a prior user study on a simulated interface agent, the learning algorithm achieves 95% of the learning efficiency standard established in that study.