My guess is @Ander is referring to this part of the BOLD tutorial, when he mentions that a portion of the workflow "is pretty random":
In the quote above, and the step @Ander is working on in this question, we're trying to find a region that aligns between a primer set and some set of sequences. It's not clear to me from these posts thus far how the user went about subsettting their data to try the alignment of the new primer set (ZBJ-Art) to their own BOLD sequence collection. In the tutorial, I point out that you should start with a subset of your samples, filtering both for sequence length and taxonomic completeness (see here). In that example, I mention filtering these gapped sequences by a minimum length of 660bp; even though the expected sequences is usually a bit less (around 658bp), remember that the expected length is that of the degapped sequence. In the case of the ZBJ-Art primers, you might expect a degapped amplicon of 157bp, but you should probably set a threshold at least that long (or maybe a bit longer). If you pre-filter your gapped alignment file in such a manner, I would be really surprised to see you getting these degapped lengths less than 150bp as routinely as you're suggesting.
For what it's worth, most BOLD sequences should be that ~650bp length, and most were generated using a primer set that occur outside of the flanking regions that ZBJ-Art primers work - see this supplementary figure that indicates where these primers fall on COI (from Jusino et. al's 2018 paper describing their development of ANML primers). Filtering for a length of just 150bp might be a bad idea, and filtering for these more complete BOLD sequences is probably an indication that they aren't junk... no guarantees though 
It would be helpful to get some clarity on this statement from @Ander too:
50 combinations of what parameters? The length of sequences? The taxonomic completeness?
One other thought would be to try a very specific subset of sequences. Say, take only a particular family of moths or beetles or any other frequently represented insect order, filter for just those records with some minimum length, filter for those records with complete names, then see if you can't get your primers to align where you know they should be aligning. If you happen to have a copy of the original Zeale paper, Table 3 highlights that they were able to recover a lot of Lepidopteran samples from bat fecal samples that matched to a BOLD reference - point being, if you're going to test our your samples on some subset of your specific BOLD references, moths is probably a good place to start for this particular primer set.