If your density is good enough (i.e. some SS elements at least are clear) and the resolution permits (generally 3A or better, but sometimes even 3.3-3.5A) then my favorite strategy is to autobuild parts of the model, hand-validate the builds (i.e. prune the autobuit model such that only high confidence portions remain) and then incorporate the built bits into DM (I usually employ Solomon via Sushi, using experimental + partial model phases). After a few 'mega cycles' of this, the density is almost always traceable w/o too much pain. If density sucks too much to auto-build, a hand-built starter usually helps, and you can do this in hand-built mode throughout :)
An extreme example of this happened just recently: 3.5A data, 4000 residues in A.U. (different conformations for monomers), rather poor phases for initial 2-wavelength Se MAD (crystal perished on the third pass, and it was the only good crystal from a pool of over fifty screened). I was able to hand-build a bunch of helices and a few sheets, then fed the partial model back into a bunch of DM software (Solomon, Resolve, DM, Parrot/Pirate) then hand picked more elements using the resulting maps, and recycled a few times. Long story short - the structure is all built now and it's a beauty :) Some day maybe I get to publish it...
Artem