Looks like you are targeting 341-805 (V3-V4) — so this could explain part of your issue (the V4 pre-trained classifiers will perform poorly since they only cover half of your amplicons; the full-length classifiers are sub-optimal; a V3-V4 classifier would be best)
Those are the adapters + primers, so contain some non-biological sequence that you want to exclude when extracting reads for training a primer. So your example command would not work — you would not extract any reads because the adapter portion would not match anything.
I have just the thing for you! See this topic:
You can follow those directions to train your own V3-V4 classifier... and please let me know what you find!