Knowledge representation schemes for the automatic generation of action-related natural language must account for the relationships commonly described in naturally occurring text. Previous work representing action in task-related discourse (Pollack 1986; DiEugenio 1993b; Balkanski 1994) has used a representation based on Goldman’s (Goldman 1970) work in the philosophy of action. These approaches have been successful in accounting for text describing the relationships between two steps, but when viewed in the context of the plan in which those steps are embedded, their representation can be seen to lack specificity. Furthermore, its limited scope cannot account for a wide range of relationships typically expressed in task-related discourse. In contrast to the work based on Goldman’s approach, researchers in AI planning have concentrated on the development of planning algorithms and the data structures that make the production of sound plans computationally manageable. There is a correspondence between the inter-step relationships expressed in utterances and the structure of plans produced by recent work in hierarchical and partial-order, causal link (POCL) planning. This paper describes work-in-progress that relates the representational requirements of task-related natural language generation to recently developed models of plans.