We present a taxonomy-driven approach to requirements specification in a large-scale project setting, drawing on our work to develop visualization dashboards for improving the quality of healthcare. Our aim is to overcome some of the limitations of the qualitative methods that are typically used for requirements analysis. When applied alone, methods like interviews fall short in identifying the full set of functionalities that a visualization system should support. We present a five-stage pipeline to structure user task elicitation and analysis around well-established taxonomic dimensions, and make the following contributions: (i) criteria for selecting dimensions from the large body of task taxonomies in the literature,, (ii) use of three particular dimensions (granularity, type cardinality and target) to create materials for a requirements analysis workshop with domain experts, (iii) a method for characterizing the task space that was produced by the experts in the workshop, (iv) a decision tree that partitions that space and maps it to visualization design alternatives, and (v) validating our approach by testing the decision tree against new tasks that collected through interviews with further domain experts.