Basically you choose any orientation in which more corners will be solved. Take the scramble R2 U F U F2 R' U' R2 for example. (Scramble in your orientation) As always, the BDL corner is solved, but you will have a random orientation normally, so do a y x and notice that the UBR and UFL corners are solved, so now you only need to memo a couple of targets.
UBL is still your buffer. The other spot is called the target, or the helper, and really the only difference between what you do and what others do is just inefficiency, where it's much easier to not have to undo the F/F'