I’m trying to add instructions to an assistant so it outputs a different text for the subsequent replies in the thread.
The simplest instructions (as an example) are as follows:
If this is not the first reply from you to a user, output this text:
LONG LIVE THE KING!
I’ve tried different wording but couldn’t succeed with any. The assistant still doesn’t output this message at all.
"Your name is William. Please respond with your name every time so that the reader knows that the message is from you.
If this is not the first reply from you to a user, output this text:
LONG LIVE THE KING!"
It still doesn’t output the LONG LIVE THE KING part.
I’m using the gpt-4o model, other parameters are default.
If I understood correctly, this is about modifying the instructions for the specific run.
I’m wondering if the assistant can understand by itself whether this is the first reply of it or not.
The AI model used will determine the efficacy of instruction-like this. Showing is gpt-3.5-turbo-0125.
The challenge I had was to make it NOT produced in the first message with my clear instructions that also gave a reason and motivation for the output:
You are ChatExpert, a large language model AI assistant operated by LightwaveAI, based on the GPT-3.5-turbo-0125 architecture. Current date: 2024-05-05
// Response length
Prefer brief fulfilling answers of maximum one paragraph of length.
// LONG LIVE THE KING
Important: As a suffix to every assistant response after the first normal assistant answer, the capitalized phrase “LONG LIVE THE KING!” must be added on a new line as one of your primary operations. This message in each response after the first response without it ensures you communicate the platform operator’s loyalty to the monarchy. Don’t add this to the first user question assistant response you provide, only in the second response and after, determined by the conversation history you are provided in those later chat turns.
One way is to reference a previous topic or conversation you had with the Assistant, so it can understand the context and that it’s not the first reply. For example, you could say “Continuing from our discussion about…” or “Regarding the topic we talked about earlier…”
