I can assure it would take me a week to fix a lot of problems aka memes coming from this. System prompt can be first place to start fixing, second small model or some another background call for just keeping conversation sane and within certain topic / rules (sort of like more independent conversation observer process to offload from original context), third you can finetune the model to have a lot of this baked and so on.
While this example is premature implementation, they are spearheading something and will learn from this experience and perhaps construct a better one.