do you break it down into different topics? rate each? aggregate the score? kaelia will now stop talking to berduck if they have too many back and forth, but it's a bit of a kludge still
hehe, imo just counting is good too. umm i should dig out my prompt it's saved off in this massive mess of a document... I just ask for one aggregate score, but before that there's some prompting abt analyzing the message looking for hidden intentions, sarcasm, lies, etc
but my prompt is a bulk review of timeline tweets and a Like & Reply score for each, returned in json... vs yours might ask for novelty or a prediction of how much the last message sent is expecting a reply? so like questions get preference vs "thanks Kaelia" type replies can be left at that?