The role of prior knowledge in skill acquisition is to enable the learner to detect and to correct errors. Computational mechanisms that carry out these two functions are implemented in a simulation model which represents prior knowledge in "constraints". The model learns symbolic skills in mathematics and science by noticing and correcting constraint violations. Results from simulation runs include quantitative predictions about the learning curve and about transfer of training. Because constraints can represent instructions as well as prior knowledge, the model also simulates one-on-one tutoring. The implications for the design of instruction include a detailed specification of the content of effective feedback messages for intelligent tutoring systems. Fifteen figures and three tables illustrate the discussion. (Contains 104 references.) (Author/AA)