Prompting Guide - GPT real time mini

Prompting
,
1

Hey everyone,

OpenAI has launched this amazing prompting guide for real-time 2.0 and 1.5 - Using realtime models | OpenAI API. Is there a similar one for gpt real-time mini? or can we assume that the same guideline can be used?

We also saw this one: https://developers.openai.com/cookbook/examples/realtime_prompting_guide that seems more oriented to the first version of the model.

Does anyone have validated this structure? Any insights on which one is more adequate to the real-time mini?

Thank you

2

Hi @lnobrega,

I’d use the newer realtime prompting guide as a starting point for gpt-realtime-mini, but not assume it transfers perfectly.

gpt-realtime-mini seems optimized for cost and latency, while the newer guide focuses on realtime-agent structure: clear roles, tool rules, confirmations, interruptions and concise responses.

So I’d:

  • use the newer guide’s structure

  • keep prompts shorter for mini

  • test tool-calling and interruption behavior separately

3

Thank you @LarisaHaster .

Do you know anyone actively working on promoting for real time mini?

4

Not specifically, no. I don’t personally know anyone focused only on prompting for gpt-realtime-mini.

What I’ve seen (and what I also do) so far is people testing Realtime prompting more generally, then adapting those patterns to mini and validating behavior case by case.

5

Thanks for raising this @Inobrega, and good context from @LarisaHaster as well.

Right now there isn’t a dedicated prompting guide specifically for gpt-realtime-mini, so most people seem to be starting with the newer realtime guide and then adapting from there.

The main pattern so far:

  • keep prompts shorter
  • simplify tool instructions
  • test interruptions/tool-calling separately

We’re also taking this thread as a feature request for more explicit realtime-mini prompting guidance and examples.

-Mark G.