"in the top 30% of all quality scores":

gunzip -c input.vcf.gz | awk -F '        '  -v M=` gunzip -c  input.vcf.gz | grep -v -E "^#" | cut -d ' ' -f 6 | sort -k1n | tail -1` '/^#/ {print;next;} { if(($6)>= M*2.0/3.0) print;}'

If you know the minimum quality, you can use vcftools:

vcftools --gzvcf input.vcf.gz --minQ <float> ...

To find the range of your qualities you can use --site-quality option in vcftools, in a separate command.

I wouldn't recommend hard filtering variants. If you are working with human data you can use VSQR to model false positives and false negatives.

Of course I am a huge hypocrite and commonly hard filter VCFs.

Care to tell us a little more about why you want 30%?

I've decided to use the variants with a QD above 30 using GATK select variant:

GenomeAnalysisTK.jar -T SelectVariants -R reference.fa --variant input.vcf -o input_QDAbove30.vcf -select "QD > 30.0"