Twenty-seven mentally retarded employees of a sheltered workshop were trained on five discrimination tasks (wires, hardware, moldings, capacitors, and fasteners) using five combination of preference, prompt, and task agreement: all agree, preference different, prompt different, tasks different, and all different. The data revealed that when the learners were prompted in the dimension of the target discrimination (a) fewer errors were made, (b) training time was reduced, and (c) fewer training trials were required to reach criterion. Preference for a particular dimension (color, shape, or size, as measured by a screening test) did not significantly affect performance on the discrimination tasks. Findings were discussed in terms of the functions of prompts as feedback, reinforcement, and symbolic parts of a compound stimulus.