I convert to polys proper for the start of it. I remember reading that Howard did most of a model in polys before converting and thought I would try it and it works very well. I do use poly proxy later but most of the structure and things like cloth creases etc is done initially in polys. I tend to have a little group of duplicates all sitting together which are hidden in layers, so I can see changes and check what looked better before converting.

I am sure this is a terribly messy way of working (!) but I find visual aids are the best thing for me, having a training in the visual medium as a route to doing practical stuff means it's a way of working I am happy with.