Due to lack of documentation and discussions available online regarding the Configure section of the GPT Builder, I propose to share here our findings, doubts or anything relevant we came up with while constructing our GPTs.
Here’s my 2 cents:
For some type of tasks it’s easier to build a GPT from scratch tailored to your needs then search for an existing one or create a dedicate script that can achieve more or less the same result of the GPT. Having said that there are lots of unclear and unreliable aspect to the thing;
They’re mainly composed by:
Name: Doesn’t seem to impact too much the GPT behavior, unless your naming isn’t consistent, that seem to confuse it a lot. My other finding is that if you use a trademarked name, you won’t be able to publish the GPT.
Description: Here you can use trademarked words without problems, I also think these words have an impact on google searches if you publish your GPT.
Instructions: One of the confusing aspects of the builder. Sometimes I’m convinced the least number of words I use the better results, other times I have to write a template thought process. I’m well confused.
Conversation starters: Usually the one suggested by the Create Assistant are useless. For instance, they often add squared brackets in the prompt as if it could take variables.
Knowledge: Incredibly slow, and I managed to successfully retrieve only PDFs so far, tried CSV or JSON but didn’t seem to be reliable at all. Haven’t tried in a couple of weeks though.
Actions: at first I thought it was the most complicated part, turns out it’s pretty easy to find a custom gpt that can help you writing the schema out of API documentation.
I’ll keep updating this if I find something else.