Most people would agree that a formalisation of granularity in natural language should involve notions of abstraction and detail. However in reality, this concept has proved difficult to capture formally. In this paper we examine the role of granularity in constructing explanations. This role is clarified by providing several dimensions along which any explanation task can be characterised. We argue that a combination of these dimensions can be used to define a notion of coherence in texts. This notion of coherence gives rise to search heuristics when solving nontrivial language engineering problems such as reference resolution.