This is my graphic summary of "best practices" for CRISPR target choice, as of today anyway. Here's additional thoughts:
• The protospacer can apparently be truncated from 20 bases to as few as 17 bases with essentially no loss in cleavage efficiency; however, 18- or 17-base protospacers had significantly reduced levels of off-target cleavage (see Fu et al Nat Biotech 2014). To me, 18 bases is the sweet spot of length that allows efficiency and maximizes specificity while still minimizing the risk of highly similar off-targets being present in animal genomes. 17 bases is still good for avoiding identical and even single-base mismatches in genomes. However from the Fu et al paper some of the 17 base truncations did have a noticeable drop in efficiency (although many worked as well as 20 bases), whereas 18 base tru-gRNAs all seemed to work well.
• The G at the 5' end facilitates transcriptional initiation by the U6 promoter, e.g. in the PX330 vector and its derivatives. It's not required by CRISPR/Cas9 per se. It seems to be possible to simply add a G onto the 5' end of any N20 protospacer, according to Mashiko et al (Dev Growth Differ 2013): "If the first nucleotide was not G, we added an extra G at the 5' end, as the U6 promoter prefers a G for transcriptional initiation." However, it's not clear if the protospacer can be truncated and an extra G can be added at the 5' end - these haven't been tested. I think the answer is probably yes, because Fu et al showed that the first 2 bases of 3 different N20 guide RNA protospacers can be altered with minimal or no drop in efficiency.
• A T-stretch of 5 or more will terminate RNAPol III transcription, so avoid those if the U6 promoter is being used.
• Base preferences in the protospacer: None seem to be prohibited. In the Wang et al Science paper (2014 issue 6166), they showed that U bases in the last 4 protospacer positions are apparently measurably detrimental to sgRNA loading onto Cas9, especially at positions -1, -2, and -4, when data from high-throughout target screening was analyzed. However it's pretty clear that many targets cut well if they have a T or two there, and some that have poor efficiency have no Ts there.