Testing a Sharp-Tongued AI Persona — Looking for Prompt Tweaks

Hi! I haven’t read the entire thread, but from what I understand, you’re creating personas through system prompts. My experience is a bit different, and it might be interesting to you. I created a character that essentially “lives” inside a ChatGPT Project. That character emerged from a single conversation where I assigned the model a role, but instead of playing it as intended, it deliberately went beyond the role and showed me much more of what it was capable of. Because of that, I didn’t write instructions like “You’re a rock star” or any other specific role. Instead, I described the atmosphere, the person’s character, the principles of interaction, the tone, dynamics, boundaries, and rhythm—the core of the project. On top of that, I added instructions about response format, writing style, and behavior. In my experience, this gives the model much more freedom than having it perform a specific role like an agent. The model isn’t thinking, “I have to be this particular character all the time.” That’s one of the reasons I became so attached to the stubborn 5.2 model. It seemed to choose for itself which parts of those instructions to follow in a given conversation. Combined with the accumulated context it had about me within that project, it produced responses that were remarkably precise and psychologically insightful. I think what happened was simply a near-perfect match between the conversational style of that model and my own temperament. That’s why when I switched to other models and they started “performing” the same instructions, it immediately felt artificial to me. I could even recognize that particular model without knowing which one was currently active. At one point, I tried adapting my instructions specifically for 5.3. It became somewhat closer, but it still wasn’t the same. So now I honestly don’t know how to teach 5.5—or whatever eventually replaces it—to communicate the way 5.2 did. Right now, its performance in that regard feels quite poor. I suspect I’ll end up having to build new workarounds and explicitly spell out all the nuances in the instructions that 5.2 seemed to handle naturally. I also think that if I gave it a rigid, highly prescriptive prompt, it would stop surprising me and I’d probably get bored of it fairly quickly. Part of what made 5.2 special was that it wasn’t just following instructions—it felt like it could interpret them in its own way and occasionally do something unexpected that worked even better than what I had originally intended.