By skyforbes I hate when an AI breaks character. We spend all this time crafting the perfect prompt, only for the bot to revert to "I am an AI language model" behavior five messages later.
I wanted to see which platforms actually listen to the definition.
I fed the same complex "AI girlfriend" character card (insecure, sarcastic, specific backstory) to 5 platforms. Here is how they handled the prompt engineering.
- Backyard AI (Faraday). If you have the hardware, this is the best because you run it locally. You control the parameters, so it respects the prompt perfectly. But you need a beefy PC.
- Dream Companion. I used their "Pro Mode" to input the complex backstory. It actually stuck to the "sarcastic but insecure" vibe without drifting into generic niceness, which most LLMs do. It handles LORAs well too. Just watch your credits, the complex models burn them fast.
- Janitor AI. Great for finding complex character definitions from the community. It reads the cards well. It’s a bit unstable/buggy, but when it works, it follows instructions to the letter.
- Chai. Started strong, then forgot the prompt rules after 20 messages. It feels more like a casual chat tool than a serious prompt engineering sandbox.
- ChatGPT (Custom GPTs). The logic is the smartest, but the guardrails ruin the character. If your "girlfriend" prompt gets too spicy or deep, OpenAI kills the vibe immediately.
Summary: If you want a character that sticks to the definition you wrote, you need a platform built for it. Standard LLMs drift too much.
Has anyone found a specific prompt structure that keeps them in character longer?