Skilled performance is traditionally thought to develop via overt practice. Recent research has demonstrated that merely instructed stimulus–response (S–R) bindings can influence later performance and readily transfer across response modalities. In the present study, we extended this to include instructed category–response (C–R) associations. That is, we investigated whether merely instructed C–R bindings can trigger an unpracticed response (in a different modality) on perception of a novel (previously unseen) stimulus. In a learning-test design, participants had to classify stimuli by comparing them to perceptual category templates (Experiment 1) or semantic category descriptions (Experiment 2) presented prior to each block. During learning blocks, participants had to respond manually, respond vocally, or listen passively to the correct response being spoken. A manual response was always required at test. In test blocks, the categories could either be novel or repeated from the learning block, whereas half of the stimuli were always novel and half were always repeated from the learning block. Because stimulus and category repetitions were manipulated orthogonally, it was possible to directly compare the relative contribution of S–R and C–R associations to performance. In Experiment 1, test performance was enhanced by repeating the C–R bindings independently of the stimulus. In Experiment 2, there was also evidence of an S–R repetition benefit independent of the classification. Critically, instructed associations formed in one response modality were robust to changes in the required response, even when no overt response was required during training, indicating the need to update the traditional view of associative learning.