Training a system to recognize and follow user-specified style choices or constraints that don't change the core task goal.