We present a new region-based approach to process discovery using prime event structures as an intermediate model. We use the prime event structure to generate precise models of the most frequent use-cases captured by an event-log. We start with an event log, apply a concurrency oracle to construct a partial language, and fold the language into prime event structures. We apply the theory of compact tokenflow regions to these structures to synthesize a sequence of Petri nets representing the most frequent partially ordered runs of the recorded behavior. The sequence of Petri nets has increasing fitness but decreasing precision. To highlight the benefits of such an approach, we introduce a plug-in for the tool ProM called Prime Miner, implementing the new concepts.