The task of optimizing machines to support human decision-making is often conflated with that of optimizing machines for accuracy even though they are materially different. Whereas it is typical for learning systems to prescribe actions through prediction, here we propose an approach in which the role of machines is to reframe problems in order to directly support human decisions. Inspired by the success of representation learning in promoting machine performance, we frame the problem as one of learning representations that are conducive to good human performance. This "Man Composed with Machine" framework incorporates a human decision-making model directly into the representation learning paradigm with optimization achieved through a novel human-in-the-loop training procedure. We empirically demonstrate on various tasks and representational forms that the framework is capable of learning representations that better coincide with human decision-making processes and can lead to good decisions.