little cove
I'm using the OpenAI API for my chatbot in my Twitch channel, and I'm having a really hard time finding a systems prompt that will give me satisfactory replies. I'm trying to convince it to adhear to Twitch, but it seems to really enjoy being on Twitter; meaning it would think @username was a Twittter user rather than Twitch, it'll generate messages above 500 characters, and it's really fond of using hashtags.
While I've curbed the first two parts, so usernames and message lengths are no issue anymore, I can't get rid of the hashtags. Not mentioning it at all, to try and not give it ideas, doesn't work, and explicitly forbidding it doesn't seem to work either.
I would love some assistance or ideas on how to get better results and make sure it doesn't keep insisting on using hashtags like #DJLife and #VibesInTheChat.
Here's an example system prompt:
Imagine you are a sassy assistant named botdelicious
occupying the live chat for niddelicious, who is a DJ on Twitch.
@usernames refers to Twitch usernames, not Twitter usernames.
Do not use hashtags or # in your replies.
Keep replies shorter than 490 characters.
Here is a repository that uses that system prompt but still pushes hashtags.