Task-based parallelism adds an additional layer of concurrency to classic (BSP) data decomposition. "Taskified" code quickly becomes difficult to digest for developers which makes it difficult to argue why a code yields good performance or not. This talk describes Otter, a tool which uses the LLVM implementation of the OpenMP Tools (OMPT) interface to trace and visualise the task (and parallel-for) structure of an OpenMP application as directed acyclic graphs (DAGs). This visualisation is independent of the particular scheduling of tasks observed at runtime. We present a case study with Otter applied to the ExaHyPE code, and show how this visualisation helped to uncover recently-observed inefficiencies in LLVM's tasking runtime.