Custom Instructions could be better using this technique!

[jochenschultz]

I found out that when we give the model our custom instructions in a structured way of prompt, answer, prompt, answer - it will perform much better in terms of following them.

ChatGPT should give that option instead or additionally of/to the current custom instructions. Means that there should be a custom chat intro that can be used on all chats.

And it should inject the chat after a couple of token. That means the displayed conversation is only between assistant and user and from time to time you add the new custom instructions in an invisible conversation between

moderator and assistant_alter_ego

which gives the chance to remove them from the beginning of the conversation (I really hope they are not embedded).

[jochenschultz]

Playing around with that could lead to beautiful answers like this.

[jochenschultz]

I mean in this conversation I have also added a couple of custom instructions (which are not visible here, but I think you get the point).

I would say at least 20% of my prompts are like “why do you write so much explanation text? If I want more information I will write verbose…” and stuff like that to correct the output.