Human-computer interaction (HCI) has gathered many guidelines for interface design and has discovered the strengths of direct manipulation as an interaction technique. However, to date it has been generally assumed that these guidelines apply generically across al applications. This article challenges this assumption ni relation to educational software. First, it is considered why educational software is different from other products, such as databases or word processors. Then, a variety of evidence is presented that suggests that designing for learning might be harder than designing for use. Current guidelines are good for optimizing current performance, not future performance. The experiments presented show how good performance can be associated with low rates of transfer, how poor initial performance can give rise to more robust knowledge for harder problems, and how command giving rather than direct action may produce better learning. A full psychological explanation is still awaited, but one can speculate that (a) interfaces that encourage planning, rather than situated action may produce better learning; (b) learners need a “sense ofengagement” (Hutchins, Hollan, & Norman, 1986) with the learning materials, not the interface; and (c) hte interface, the learning context, or the tasks may individually or together cause changes in strategy from implicit to explicit or vice versa.