What do you want the element to model? Build up a series of example problems going from something extremely simple to about as far as you can go with manual calcs. Know what you want the results to be. Read the documentation to find out which elements seem to fit. Experiment with the program until you're satisfied that the elements and are doing what you want and settle on one that way.
I really don't think there is a substitute for that sort of approach.
BTW, this part of the process is worth a lot of time. I just assigned a small research project that should take about 150 hours. I gave him a series of mandatory verification problems that will probably take 50 of those.