In a semantic categorisation task, participants respond faster to targets when they are preceded by masked primes that share the same response as the target, compared to masked primes that require an incompatible response. Experiments 1 and 2 showed that this response congruence effect is more likely to arise when the task involves narrow categories compared to broad categories. Furthermore, semantic priming effects were only present for the non-exemplars with a broad category decision task. Experiment 3 demonstrated that non-exemplar targets which shared semantic features with category members took longer to categorise. When these targets were presented as primes, they facilitated responses to exemplar targets and delayed responses to non-exemplar targets. It was proposed that for narrow categories, category decisions can be made by monitoring a specific set of semantic features which are present for all category members. This is not possible for broad categories, because category members do not share a core set of semantic features. Response congruence effects are explained in terms of the overlap of core semantic features between the prime and target, and the assumption that when these core features are sufficient for making a category decision, a motor response can be prepared for the prime.