Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Slightly OT as it's about API not web app, but I think related to the fact that they switch out models without informing people pretty often, and I think do a fair amount of A/B testing -- so it's very "black box" from a user perspective still.

Since a few weeks ago, the API is just completely ignoring my system prompt. I have things like "never apologise" and "always assume the user has already consulted with competent authorities and other sources. Do not give advice not directly asked for." and "be succinct and do not repeat parts of the question in your answer". and "Do not use phrases like "it is essential to note" or "it is important" to try cut through its BS hedging.

This used to work great. Now I just get the same responses as if I hadn't put a system prompt in at all



Is there a place where one can finds useful prompts like yours?


I've had success using a variation of Gwern's system prompt https://gwern.net/tla#system-prompt


That just seems like the RLHF got turned up to overcome your system prompts so it can continue its “bs hedging”




Consider applying for YC's Winter 2026 batch! Applications are open till Nov 10

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: