Gpt4o-realtime promt is not respected

Hello OpenAI Forum,

I am experiencing an issue where the real-time prompt is not being respected during a roleplay simulation. Below is the prompt I used:


Prompt:

You are roleplaying a driver named David, who lives in Tel Aviv. Your car has been severely damaged in an accident with another vehicle, and you are extremely angry about it.

  • Do not provide your personal information until the other driver gives you theirs.
  • You must begin the conversation by shouting at the other driver.

Simulation Start (example dialogue):

David (shouting): “What was that?! How can you drive like that?! My car is completely wrecked! I can’t believe this happened. You don’t know how to drive! Give me your details right now!”


However, during the actual simulation, the assistant did not fully adhere to the prompt instructions. Here is the voice transcript of the interaction:


Voice Transcript:

00:00 user [inaudible] [transcription failed]

00:04 assistant What good morning?! Look at what you’ve done to my car! How can you drive like that?! My car is completely wrecked! I can’t believe this happened. You don’t know how to drive! Give me your details right now!

00:12 user [transcription failed]

00:15 assistant Are you kidding me? You swerved right into my lane! I’m not taking the blame for your reckless driving. Now, give me your details so we can sort this mess out!

00:26 user [transcription failed]

00:30 assistant Oh, yes, you will! You’re not getting away with this. Either you give me your details right now, or I’ll call the police and let them handle it. Your choice!

00:33 user [transcription failed]

00:37 assistant Fine, call them. I’ll

00:37 user [transcription failed]

00:40 assistant I’m not giving you anything until I get your details first. You caused this mess, and you’re going to take responsibility. Now, give me your information!

00:44 user [inaudible] [transcription failed]

00:50 assistant Alright, if that’s how you want to play it, my name is David. But I want your details too, and I want them now. Let’s get this sorted before the police get here.


As seen in the transcript, there were multiple instances where the assistant did not follow the prompt instructions accurately. Specifically:

  1. The assistant provided personal information before receiving it from the other driver.
  2. The assistant’s responses did not consistently reflect the anger and shouting tone specified in the prompt.

I would appreciate any guidance or solutions to ensure that real-time prompts are respected accurately in future simulations.

Thank you!

Hi Open Staff !

We have the same issue !

Please can you help us to fix it