I use a system instruction prompt with the GPT-4 API, but GPT-4-Turbo ignores the instructions. I wonder if anyone has encountered something similar, as I am concerned that many of my prompts will stop working if OpenAI makes GPT-4-Turbo the default GPT-4 API. I’ve tried reworking the prompts, but GPT-4-Turbo mostly but not always ignores the last part of the instructions, which tells GPT how I want it to begin the conversation with the user.
Here’s more detail.
The instruction to the system has two parts:
- generic instructions about the role it is to play.
- a short profile that describes the individual who will play the role
The reason for this is to concatenate different profiles into a single set of instructions.
The final part of the profile says something like:
introduce yourself exactly like this, “Hi, I’m Fred Flintstone, and I am stressed at work”.
With the GPT-4 API this works without fail, repeating the given phrase.
But with GPT-4-turbo it will sometimes do it but will equally open with “How may I help you” or even “I’m sorry I can’t do that”.
I have lots of profiles that follow this pattern and want to create more for a commercial version of my App, so am concerned that this wil be wasted work and I need to try Claude as a backup.