I think that agrees very much with my comment, but is sort of orthogonal to the comment above - which I think would recommend reimplementing even the right abstraction if the existing implementation is sufficiently poor.
One way I've always heard it referred to is: refactor the code so your task is easy to do. If making a new abstraction does that (just had this happen last week for me) then so be it.
I think that is incomplete. One should be aware of what other tasks a refactor is making more difficult or more error prone; it is not purely a matter of whether it makes the current task easy.
