Again, @Tintin, I may not be the best-qualified person to answer this question, so I’m going to lean on resources that might help. Please forgive me if any of this is unhelpful to you.
It is not necessary to denoise. See the OTU Clustering Tutorial for a how-to, and note that
these files are analogous to those generated by
qiime dada2 denoise-* and
qiime deblur , except that no denoising, chimera removal, or other quality control has been applied in the dereplication process."
Taking this approach probably means manually applying your own QC, as you’ve suggested. By correcting rather than dropping or ignoring noisy reads, DADA2 provides the added benefit of reducing the number of false positives, often apparently yielding fewer ASVs than you would have gotten OTUs from clustering methods.
What you choose for your analysis should fit your study needs, but as a side note, here’s an amazing, brief look at the history of taxonomic assignment and clustering, with additional good posts linked in it. Depending on your specific work, you may be able to simplify your pipeline and preserve more data without clustering to 97%. There are definitely cases where clustering/OTU picking avoids pitfalls inherent to ASV/denoising methods, but it’s worth a read if you haven’t considered this approach.