He's referring to cases 19, 20, 21, 22 , which all look rather alike.
Gotta admit, I struggled with these. The easy cop-out is indeed to AUF until the edge matches the centre, and it's not that harmful, movecount wise. I eventually got good enough at recognition, and no longer considered these cases a problem. It was when I tried to learn pair-making using empty slots that the struggle reappeared. And I never did get it mastered ... too long recognition, too many errors.
As an example:
Starting with a completed first two layers, setup:
F' U' F R' U2 R
The most obvious thing is to recognise that a U will align the edge and centre, then solve it algorithmically, so typically
y' U R' U2 R U' R' U R; = 9Turns in QTM
But that's not really recognising the case, if I could do that consistently, I would see the empty-slot solution:
R U2 R y' R' U R; = 7Turns QTM = better
In this specific case, I can see the solution... but if the setup was F' U' F R' U2 R U' , I would bale out, and start with a U2, giving me a 10T solution, not great.