This paper aims at finding the optimal combination of verbal instruction and on-line practice for learning a new computer application. Experimental subjects learned commands for an electronic spreadsheet by reading brief user-manual descriptions and working training problems on-line. The form of the training problems was varied within subjects in order to control how much independent problem solving subjects engaged in while learning any given command. There were three forms of practice: (1) Pure Guided Practice, in which subjects were told exactly what keystrokes to type to solve the problems; (2) Pure Problem Solving Practice, in which subjects solved problems without guidance; and (3) Mixed Practice, in which the first problem for a command was presented in Guided Practice form and two others in Problem Solving form. The spacing of the training problems was also manipulated; the problems pertaining to a given command were either Massed (i.e., presented consecutively), or Distributed (i.e., separated by other instructional material). After a 2-day delay, subjects solved new problems on the computer without reference to the instructional materials. The results indicate that problem solving was a more difficult form of training than guided practice, but it produced the best performance at test. Distributing the spacing of training problems during training also improved performance at test. The results have clear pragmatic implications for the design of interactive tutorial manuals as well as implications for cognitive models of skill acquisition.