I'd either use amino-acid fragments with similar properties and consequently the same atom-types for the parameters, as suggested in the MacKerrel paper. Guess you use a graphical program (e.g. quanta or insight) with an in-buildt CHARMM facility. If so, then I believe the topology file can be generated automatically.
An alternative is to use semi - or ab initio methods and perform a gasphase optimization geometry. In that case you'll have to rescale the frequencies to assign the empirical force-constants (i.e. kb, ka and so on) within your molecule, and use existing non-bonded L-J and charges from an analog (protein) residue.