Typical help processors are invoked explicitly by the user or implicitly when an error occurs. Often a beginner will not know that he needs help because the inefficient use of commands will get the job done without raising errors. WIZARD is an expert system that recognizes beginner misbehaviors and can automatically start a help transaction. The WIZARD processor relies on a special purpose, dynamic, pattern matcher directed by a KL–One based knowledge network. An author studies logs of beginner interaction and develops sequence rules which parse and properly identify misbehaviors. Objects that drive the parser to understand VAX DCL commands are coded into the network and a set of semantic programming utilities is used to perform actual goal recognition. This thesis deals primarily with the implementation of such a goal recognizing expert invocation system. It is the WIZARD documentation and final working report. I discuss the motivations for the design of the system and detail the knowledge base and heuristics that support goal recognition. Some issues of generality are taken up and potential topics for later research are presented which will extend WIZARD'S capabilities.