Comment by stavros

The general way I think about it is that it's OK (and usually encouraged) to have a smarter model prompt a dumber model, but never the other way around.