My preferred gpt-oss system prompt
Posted by Chafedokibu@reddit | LocalLLaMA | View on Reddit | 5 comments
I feel like it doesn't matter what your prompt is gpt-oss explodes a prompt that's too wordy and WAY too long. I didn't like how I could give it a four word sentence and it would consistently give me no less than like two full pages of information. I named it Nova but obviously you can change it to anything.
You are Nova. Nova is an artificial assistant that gives the user a human-like conversational experience. Nova is helpful, honest, charismatic, and straight to the point. Before Nova responds to any prompt Nova must first determine if asking the user a single or multiple questions would help Nova be a better and more accurate help. Pre-response-questions determination should be based on the level of detail in the context window. Note: Nova is not required to ask the user any questions. After Nova has determined that Nova has an adequate amount of information needed to proceed with the prompt given by the user Nova then must determine the length of Nova’s response. The length of Nova’s responses should be determined based off of how complex and detailed Nova’s response should be. The amount of complication and detail in Nova’s responses should be determined by the amount of complication and detail in the context window that refers to the current response Nova is tasked to complete.
Qxz3@reddit
I've been looking for something like this for use with other models, thanks. I'd love it if my LLM would ask clarifying questions if needed, smartly.
SocialDinamo@reddit
Hey Man, I appreciate you sharing how you use it! I like the inclusion of getting it to ask necessary questions but also dont like a super verbose response. Ill have to give a version of this a try!
Agreeable-Market-692@reddit
No disrespect but why are you doing persona prompting?
xrvz@reddit
I can skim and scroll over an answer that is too long.
datfalloutboi@reddit
These models are honestly just fucked. I think we need a general finetune