In this study, we investigated how musicians use descriptive imagery to perform expressively. The study's procedure was carried out by university musicians alone in a naturalistic practice environment. They recorded their performance of two melodies both before and after receiving imagery-based instruction designed to make performance of the melodies very expressive. Participants read and considered a teacher's expressive instructions for each melody, practiced in light of it, and recorded a final expressive performance, transcribing their thoughts while practicing. Results showed that ratings of expressiveness increased significantly from pretest to posttest, indicating that the imagery-based instruction was effective at improving participants' expressive performance quality. Also, some musicians used a cognitive translation process, by which they explicitly generated plans for the musical sound properties in their performance, but others exclusively reflected on and personalised the provided imagery in their thoughts. Further analysis revealed that scoring high in the use of the cognitive translation process corresponded with greater improvement in expressive performance. We were also able to identify 'cases of mistaken expressivity', i.e. instances when participants believed their performance had effectively improved, but a change in expressive sound properties was not perceived by the listeners who made up our panel of judges.