Side effect machines are augmented finite state machines with counters on each state. They are used to convert DNA or other string data into numerical features. In this study we examine the effect of imposing shapes on side effect machines. When a standard finite state device is programmed with an evolutionary algorithm there is no restriction placed on the transition function. A shape for a population of evolving finite state machines is a restriction on the possible transitions. We demonstrate that choosing a shape with expert knowledge yields improved performance on a supervised classification task. The shapes used are designed, induced from evolved side effect machines, and designed based on features of evolved side effect machines. The best performance was exhibited by an shape induced from an evolved side effect machine.