Our two stage methodology for the study of cue usage coordinates an exhaustive corpus analysis with a system for text generation. Coding of the corpus uses Relational Discourse Analysis, a synthesis of two previous accounts of discourse structure. In the first stage of our study, hypotheses about cue usage are evaluated and refined using the corpus analysis. Several initial results conccrning how cues mark segment structure are presented here. In the second stage of our study, the results of the corpus analysis are used to determine a set of heuristics to be implemented in a system for text generation. The automatic generation of texts is then used to exercise and further evaluate the heuristics for cue placement.