This paper presents how non-representational views of cognition can inform interaction design as it moves from traditional graphical user interfaces to more bodily forms of interaction such as gesture or movement tracking. We argue that the true value of these "bodily" interfaces is that they can tap our prior skills for interacting in the world. However, these skills are highly non-representational and so traditional representational approaches to interaction design will fail to capture them effectively. We propose interactive machine learning as an alternative approach to interaction design that is able to capture non-representational sensori-motor couplings by allowing people to design by performing actions rather than by representing them. We present an example of this approach applied to designing interactions with video game characters.