I propose a new approach to mechanism design: rather than assume a common prior belief, assume access to a common dataset. I restrict attention to incomplete information games where a designer commits to a policy and a single agent responds. I propose a penalized policy that performs well under weak assumptions on how the agent learns from data. Policies that are too complex, in a precise sense, are penalized because they lead to unpredictable responses by the agent. This approach leads to new insights in models of vaccine distribution, prescription drug approval, performance pay, and product bundling