The paper might explain why it is so. Routine-setting in your brain runs separately than the execution of the activity itself, meaning you need to train the two separately, at least from what I understand.