Hi,

In upcoming version of PHENIX there will be more detailed manual for phenix.refine.
Until recently phenix.refine was not using H atoms by default. We changed this behavior very recently. In newer version, if you have an input PDB file with H atoms (generated for example, with phenix.reduce), then phenix.refine will use these hydrogens by default. You don't need to do anything special for this.

I nowadays use phenix.refine almost always in the first stages of the
refinement and later usually move to refmac. Now I have two modest
resolution structures which behave much better in phenix than in refmac.

This is partially because Refmac uses "2nd derivatives' minimizer". Saying differently, given a pretty refined model out of phenix.refine Refmac can possibly refine it some more to deeper minimum. The downside of this is sometime higher Rfree factors.
However, if you have a pretty bad starting model, phenix.refine is more powerful than Refmac to bring this model to a reasonably refined state.

B-factors on
the other hand seem to be less restrained in phenix.

Here is how B-factors are restrained in phenix.refine (if TLS is not used):

http://www.ccp4.ac.uk/newsletters/newsletter42/content.html
look for " The Phenix refinement framework " by Afonine P.V., Grosse-Kunstleve R.W, Adams P.D.

It is pretty old text, but reflects some general things. Note that "k" and (Bi+Bj) in the ADP restraints target formula have now different empirically found values.

And yes, playing with wxc_scale is definitely good thing to do if you're not happy with either Rfree-Rwork gap or stereochemistry.

Analogously to wxc_scale in coordinates refinement, to play with B-factors restraints use wxu_scale.

If I take the phenix refined coordinates, keep the aniso records from TLS and run some refmac 
cycles, my structure is better than average with regards to the clash scores (69 %). I do not 
know how valid refinement strategy this is. (Rfree-R = 5.4 %)

Please look another post on phenixbb (Subject: "Re: [phenixbb] R-factor problem !!!") regarding how tricky and error prone could be to jump between different programs.

Pavel.