You can use the "developer" role which can go on any message throughout the conversation

https://model-spec.openai.com/2025-10-27.html

However models tend to follow instructions best at either end of the prompt, so the beginning is still the best place for "system prompt" type instructions https://arxiv.org/abs/2307.03172

So I'd recommend making your system prompt something variable that you can adjust as new information comes in. You could consider storing it in Redis and fetching on each message, then when an event happens you add the event info to the stored system prompt. Or you store a list of events and your system prompt is constructed by fetching them as bullet points

Thanks you for the response.

I did consider appending them to the original system prompt, just wasn’t how the model might react to updates being at the start of the conversation instead of at the end, when they cronologycally occured, but I will try it out.

I also didn’t know about the developer role. I don’t think conversation length will be an issue in this app, as conversations are quite small and simple, but I’ll try both approaches and see which one works best.

Thanks a lot!

@jacobparis Hey just checked and it doesn’t seem like the AI SDK supports the ‘developer’ role in UIMessage or in ModelMessage. Is there something I am missing? How do I set the role to developer without getting a type error? Thanks

Hey @jacobparis, just wanted to follow up on this. Any thoughts?