" have a problem with prompting. I spent several days trying to write a reliable prompt. The topic: I run an agency and have an AI agent as a secretary who answers general questions about the agency or schedules appointments. There are two tools: one to check availability and one for the booking function. I kept it very detailed because I noticed that with shorter prompts, the quality fluctuates a lot. I set the temperature to 0.2 and formatted everything in Markdown. I’m using GPT-4o-mini, and everything is transcribed in German. However, I have the problem that my AI agent either doesn’t execute the functions, just reads them out, or doesn’t take the examples in the prompt as context/inspiration but instead copies them verbatim, even though I’ve defined guidelines and described them there. For example, in the prompt, I have something like "**A**: Looks good, they are available at nine, ten-thirty, and eleven-thirty. What works best for you?" but it just copies this instead of executing the function to check availability and then responding with the result. Additionally, there are repeated follow-up questions when I say, for example, "I’d like an appointment tomorrow at 1 PM," and it replies, "Okay, when would you like the appointment tomorrow?" Sometimes steps are skipped, and further actions are taken, even though I’ve noted instructions like "It’s life-threatening if you proceed without considering this step." I tried it with the VAPI GPT, but as soon as I close one gap, another opens up. How do you manage your prompts to have a solid base, and then test and fix errors? I feel like one part falls apart as soon as I build another.