After many, SO many attempts at trying to get more intricate GPT’s to be consistent. I may have had a breakthrough, that I wanted to share.
Now I have tried it a decent amount of times but only time will tell if it does stay consistent.
The prompt was made in a way similar to a legal document. I thought to myself what is the most precise thing I’ve read. Oh, Privacy policies. With there legal jargon it’s very precise. I had GPT actually make this document for me with a section I can fill in. It overly stated beyond insanity to follow this section. Emphasis on this section and that I’m referring to this GPT in question.
This prompt structure made a night and day difference.
There was a user here a while ago who achieved incredible consistency between images. Unfortunately they got banned and their stuff taken down, but I think the gist was that they had a fairly comprehensive template that they used to get the output they desired.
I think this particular approach can work for either case. Although I do see scenarios where it breaks down, It’s probably a very good starting point.
I don’t get it, what do you mean? Can you give me an example?
An when you talk about consistency do you mean over time?
As you might know, ChatGPT has a context window, ie. it forgets over time if it’s a long chat. My “hack” for this is that I ask it to create a sun emoji at the end of each answer, when I no longer get the sun emoji I know the context window has been used up.
In a sense, yes. It was focused on the logic of it, presenting numerous ways to express a single idea to avoid any misunderstanding—almost like a privacy policy. It specifies things in detail, such as describing which sections or elements are being referred to. In my case, I wanted to let this GPT know that I was referring to it and that it should pay attention to that particular section. I then proceeded to fill in the section. Sorry for the very late reply; I don’t visit here often.