This article argues for a particular understanding of feature class behavior-the recurrent patterning together of certain phonological features, such as place of articulation and laryngeal features. The proposals build on the well-known work of feature geometry in assuming the importance of feature classes in phonology, but differ in that features of a class are targeted directly and individually by constraints (or rules), even when a feature class such as Place is mentioned. Further, constraints mentioning feature classes are gradiently violable. Evidence for this view of feature classes comes from two sources. First, assimilation involving feature classes is sometimes only partially successful; an adequate understanding of such cases requires the proposed view of feature classes. Second, there are broad categories of feature class generalization that require it, including dissimilatory effects usually handled by the obligatory contour principle. Overall, the proposals broaden the explanatory potential of the feature class idea due to feature geometry. At a more general level, the results here suggest that linguistic representations sometimes need to be reconsidered in the context of optimality theory (Prince & Smolensky 1993), since they can effectively function as inviolable constraints and so hinder our understanding of the more subtle kind of phenomena revealed by analyses employing gradiently violable constraints.