RNR: Teaching Large Language Models to Follow Roles and Rules