You also need to use a fresh incognito instance if you want a result unbiased by previous sessions.
and it would provide more context if the prompt was also shown.
As to jaeger's write-up, it was focused merely on smack/insults/harassment, and shows there would be effort required to train it on context. Not bad at all, but it doesn't focus on other actionable aspects like politics (can talking about history be viewed as "politics?) or vulgarity, whether masked or alluded to, the sexual innuendo/ locker room talk (is that allowable?)
I see SkuzzAi, both forum and in-game, being feasible, just not as a priority at the moment considering the training time that might be required. It has to be reviewed, assessed and tweaked.