GPT-3.5 & 4
Getting ChatGPT (3.5 & 4) to work with you requires prompting, a lot of prompting. So much prompting that it takes multiple simultaneous chat sessions, because the AI can only remember a limited amount of the conversation. I try not to muddy up my conversations. To continue long conversations, you have to summarize, compile, condense, paraphrase, or simply repeat the ideas into a new chat session. I had multiple notepad documents open to help me copy and paste back and forth.
To start off, I knew what I was looking for, but I had to get ChatGPT up to speed and “thinking” about the topic. The first prompt was just to get the AI started with the idea and helped me recognize how my system could be perceived based on the AI’s feedback…
My first prompt:
I have an idea for a role playing game. [gave simple examples of various games systems.] I would like to create an RPG that is [described the details of what I had in mind]:
In return GPT-4 gave me a list of general RPG design ideas with very standard tropes. A nice start, but I needed to tweak the ideas to line up with what I was thinking and correct a few assumptions.
I kept up this prompting until I could consistently get a proper response from GPT-4. Once I had the basics on how character creation worked, I needed a generic D&D character to convert.
New Chat Prompt 3.5
My next prompt:
Create a 1st level D&D character named Fidget.
Again, I had to post lots of prompts to get the final product. The AI started with an average rogue, a couple skills, and then got lazy and said “etc.” Perhaps I should have prompted “create a complete D&D character”.
Before the AI starts to forget, I ask GPT-4 to summarize what we had so far. I pasted the summary (with a few corrections) into a new Chat thread to continue the conversation.
New Chat Prompt 4
Then I pasted Fidget and my new rules into a new Chat thread with the prompt:
Let see how well this works. Convert the following D&D character to this new system: [Fidget character sheet]
Once again, I had to coax the AI to finish the conversion. It also showed me holes in the text that I needed to fill in and clarify.
A Thousand Characters
Once the rules get beyond 1000 characters, you have to use a special multi-prompt method to keep the AI from muddying the conversation. I found this one online somewhere. Not sure it really works well after 5 pages of text, given the limited memory of GPT-4.
New Chat Prompt to “simulate uploading a document”
Prompt: `Act like a document/text loader until you load and remember content of the next text/s or document/s.
There might be multiple files, each file is marked by name in the format ### DOCUMENT NAME.
I will send you them by chunks. Each chunk start will be noted as [START CHUNK x/TOTAL],
and end of this chunk will be noted as [END CHUNK x/TOTAL],
where x is number of current chunk and TOTAL is number of all chunks I will send you.
I will send you multiple messages with chunks, for each message just reply OK: [CHUNK x/TOTAL], don’t reply anything else, don’t explain the text!
[START CHUNK 1/3]
A bunch of text from the document goes here… about a 1000 characters, because you can’t post too much at once or you get an error. bla-bla-bla
[END CHUNK 1/3]
Reply with OK: [CHUNK x/TOTAL], don’t reply anything else, don’t explain the text!`