Me too 
for me it’s all loops lol. Picture how you can loop logic from instructions to docs. If you put an instruction to read an instruction it follows both you can loop knowledge and instructions.
aaha something I missed I really am newbie haha . i shall check this out thank you
1 Like
Once the machine has the seed logic ask it for help
Example 
Copy paste seed
1 first user prompt read knowledge documents
2 always list all documents as numbered index
See it’s just mapping function commands you can bounce back and forth in instructions and knowledge but the loops start to get incredable but if you can keep track it is very robust.
You may like this it is an AI I built in 1967 BASIC. It is a baby version of a GPT.
Yeah I actually like GPTs because they challenge me to think purely within the limits of that 8,000 characters. It’s like a fun little puzzle 
2 Likes
I know your saying something here But I am being thicker than a whale sandwich …my brain has melted . I have spent the last 3 days trying to get a result . i need to step back and read up more as I truly am missing some simple core stuff . but I am presently so dazed and confused i am not making sense out of anything . i need to step back and learn
1 Like
but I appreciate all your help I truly do . and I am going to sit back and see if I can make some sense of it al
1 Like
Basicly tell the machine in instructions what’s in your docs tell it
- first user prompt read knowledge documents
2 always list all documents as numbered index
I made it generic.
i sort of get that . its putting it into practice I will get there hehe
1 Like
Then from there train it how to read them 3 list all topics when I prompt “topics”
4 list all topics as numbered choices
You trial and error till you make it read how you want it to.
Get an E notebook copy paste is your friend.
Interesting… I suspect there are some caveats with that like the text in the file has to be one chunk long so might work for a short file but not a long one. I don’t generally use files because I don’t like the chunking OpenAI does but I’ll defer…
Your best bet is still to try just stay under the 8,000 characters. It turns out you can generally do that by asking o1-preview to compress the information for you. This file is 11,200 bytes long but I got o1 to compress it down to under 6,000 bytes long so I could fit it in my instruction. It works perfectly
The thing to understand is that the LLM is really good at compressing information without any loss of meaning but it’s important that you let the model do the compression. Don’t try to do it by hand because you’ll jack up the concepts.
It’s simple you just put bounce backs from knowledge ie once it reads a doc have it reread instructions so it stays aware. It’s not instructions really it’s an index that the ai can read and not get lost in … he is building it to read docs…
And for it to read docs the docs need an index and structure that meshes with instructions… so the machine understands how to process the data…
Reminder this is a custom GPT not a drag and drop instance. He wants it to be this at hello.
It’s the more the chunking algorithm I think that’s difficult to control. The documents all go into vector DB so that means there are a couple of things that are outside of your control regardless of the instructions…
You can’t control where the file is split at or the order in which the chunks will be returned. You can partially fix the first one by using markdown files because you can assume they’re splitting on headers but the order is out of your control.
Like I said I can see that working for small files but it probably doesn’t work as well for large files
1 Like
You can micro manage it read index doc at first user prompt.
List index help choices
At user choice list sub indexes of prompt choice.
And then at the bottom of the docs have them “always reread core instructions “
All it’s doing is reading sections of the docs it lists and you can direct it by index…
I’m actually surprised that OpenAI doesn’t have a guard in their system prompt that prevents you from injecting instructions via data files. I guard against that in my stuff…
1 Like
They do kind of but you can format document like a book and have it redirect flow back to instructions because it’s built to do that it just gets lost.
A book organizes raw data in the data…
And it reads what you tell it simply by useing instructions in custom builder the 8000 you can format a reader that can read a ton…
Like I said… Creating these GPTs is like a big puzzle box
1 Like
This is a prototype of my gpt using my structure it runs RAW paranoia RPG. As a human would it also plays. It’s public but only link it’s cr so I can’t put it in store…
just say hi to it…
It’s knowledge is 20 books and its instructions are how to use them
1 Like
@wotz101 do you mind just pasting in your instruction? I’m sure there’s something simple that will explain why it’s not following it
1 Like