Cufflinks can assemble novel transcripts, so make sure to also supply the -G option. One thing to think about is the size normalization and dispersion estimation steps in cuffdiff. With only 10-20 transcripts, I would worry that at least the library size normalization step wouldn't work very well. If you have enough replicates that you don't need to share information across genes, then the dispersion calculations should be fine. The simpler method, btw, would be to just use the full dataset for all of the estimates and then just subset it to the transcripts of interest afterward (you can then readjust the p.values).