See also
OTU / denoising pipeline
The input sequences to cluster_otus or unoise3 must be a set of unique sequences sorted in order of decreasing abundance with size annotations in the labels.
The fastx_uniques command can be to find the unique sequences and add the size annotations.
Finding unique sequences is sometimes called "dereplication".
I suggest you use -relabel Uniq so that the unique sequences are labeled Uniq1, Uniq2 and so on. The input to fastx_uniques should be the reads after any quality filtering or length trimming.
It is critically important that sequences are quality filtered and globally trimmed before dereplication.
Samples should be pooled before dereplication
I recommend pooling samples, i.e. concatenating reads for all samples
that were sequenced in the same run. This is important for getting the best
detection of chimeras and
cross-talk, and for getting the best sensitivity to low-abundance
sequences that could be lost if individual samples or subsets of samples are
clustered separately. See pooling samples for
discussion.