Custom instructions vs. api system prompt

Hey. I’ve been playing with the custom instructions in the commercial ui and then in the api playground. When I put my ai’s “personality” into C.I. the responses are much more aligned with what I wrote. When I write the same instructions in the system prompt using API playground, the results are more vanilla and less like what I described.

I have tried with different settings like temperature, but it seems that C.I. might have something extra added to it, to make it follow instructions more clearly?

1 Like

I have seen the same thing I also noticed when looking at the logs for the request that the instructions field was set to null if you enter the instructions in the large field below the assistants name.

However, if you enter the same text instructions into the run instructions, the request will include the text in the additional_instructions haven’t tested the outputs though I just figured this out myself.

1 Like