They're changing the order (and removing a layer of complexity) by using an earlier part of their model to calculate skill use later. It could certainly lead to further improvements in either efficiency or potentially transfer learning down the line.