Our attempts to improve physics instruction have led us to analyze thought processes needed to apply scientific principles to problems--and to recognize that reliable performance requires the basic cognitive functions of deciding, implementing, and assessing. Using a reciprocal-teaching strategy to teach such thought processes explicitly, we have developed computer programs called PALs (Personal Assistants for Learning) in which computers and students alternately coach each other. These computer-implemented tutorials make it practically feasible to provide students with individual guidance and feedback ordinarily unavailable in most courses. We constructed PALs specifically designed to teach the application of Newton's laws. In a comparative experimental study these computer tutorials were found to be nearly as effective as individual tutoring by expert teachers--and considerably more effective than the instruction provided in a well-taught physics class. Furthermore, almost all of the students using the PALs perceived them as very helpful to their learning. These results suggest that the proposed instructional approach could fruitfully be extended to improve instruction in various practically realistic contexts.