System prompt weirdness
I'm running with a system prompt set in llama-cli. It is clearly aware of the system prompt, but in think tags it says stuff like:
"The system message says we are ChatGPT, a large language model. The developer message says: "You are ChatGPT..." There's no explicit instruction about refusing content. The user is just greeting."
or "Now, we need to follow the developer's role: "You are ChatGPT, a large language model trained by OpenAI." The user didn't ask anything that violates policy."
But it ALSO references my custom system prompt in the think tags.
Anyone else seeing behavior like this?
Yes, I did. Running with llama.cpp and simply opening up Claude code with the model selected (not even sending any chat), I see messages like this in the llama.cpp console logging. Where are they coming from? I also see messages I expect to see, you are claude code from anthropic ...
its probably a distill from chatgpt im guessing, im getting the same stuff...