Alexey Muranov wrote:

My, maybe naive, impression is that there should always be one class per operation, as this is obviously the most flexible approach. Then maybe some classes can be grouped together by introducing other classes.

Truly what goes around comes around. The ``less ad hoc'' overloading which we now associate with type classes was first proposed (and even implemented!) by Stefan Kaes, a couple of years before the type classes. Kaes so-called parametric overloading may be regarded as an implementation of type classes with only one method. http://okmij.org/ftp/Computation/typeclass.html#Kaes Incidentally, Kaes' system included not only `local instances' but also `local classes'. Likewise, 'Haskell with only one type class', http://okmij.org/ftp/Haskell/TypeClass.html#Haskell1 deals first with one-method classes. Type classes like Num are then regarded as convenient collections, as bundles of previously defined methods. Naturally, the same method, like (+), can be a part of several bundles, not necessarily related otherwise. Incidentally, The `Haskell with only one type class' required NO extensions, even back in 2007, when it was first proposed.