The thesis describes a design for how a tutoring system can enhance the educational capabilities of a conventional knowledge-based system. Our approach to intelligent tutoring has been conceived within the framework of the KNOWLEDGE-LINKER project, which aims to develop tools and methods to support knowledge management and expert advice-giving for generic applications. Biochemistry, more specifically experiment planning, is the current reference domain for the project. The selected tutoring paradigm is a computer coach, based on the following central concepts and structures: instructional prototypes, an intervention strategy, teaching operators and instructional goals controlling the instructional strategy; error descriptions to model common faults; and stereotype user models to support the user modelling process. The tutoring interaction is planned using the instructional prototypes and constrained by an intervention strategy which specifies when the user should be interrupted, for which reason, and how the interruption should be handled. The set of instructional prototypes and teaching operators can be used to describe individual teaching styles within the coaching paradigm; we propose one way to represent them using discourse plans based on a logic of belief. The case data may be either generated by the coach or specified by the user, thus making possible using the coach both for instructional purposes as well as job assistance.