> I think it is depends what you wanted to do with the generated structures.
Generally I would like to come as close to the general CASP7 ab initio rosetta protocol as possible. I cannot generate too many decoys with abrelax though so I am thinking about generating some tens of thousands decoys with standard ab initio and then running relax for cluster centroids and the some number of the lowest energy decoys from that clusters (generating single relaxed decoy for each initial crude decoy -nstruct 1). Then, I am going to do re-clustering of the relaxed structures to find the “best” decoys.
However, maybe it is better to select smaller number of starting crude decoys from standard ab initio mode for the next refinement stage and instead use the computational power to generate more decoys with -relax for each starting structure (that is -nstruct N). In other words I could discard some small and low energy clusters and instead generate more refined decoys from the “better” decoys from the first ab initio run.
>It might because the native model has big clash so that it causes the decoy to break down a lot.
Yes, it might be the case, I haven’t tested it thoroughly.
> Generally, the more decoys are generated, the better coverage you will get.
Of course, but I need to compromise this with my computational resources 
> If you want to keep the bond angles of native structure, you can try to use the flag -use_input_bond, which “Minimize or refold using the bond lengths and angles from the starting structure, useful for minimizing a structure without idealizing it first.”
Ok, I will keep this in mind.
Cheers,
Janek