Preparing outlines for 100 articles with ChatGPT-4

Encouraged by the success stories, I am creating my own solution to create more content through the OpenAI API. Wanting to save on tokens, I would like to prepare guidelines for, say, 100 articles, which I will then use to generate content. The problem is that ChatGPT-4 (with advanced data analysis) in the browser version doesn't comprehend such amount of data, it has problems even with 10 articles, for which it has generated mid-headlines and some other guidelines. I've tried a scenario where I put the entire instruction in one prompt and the step-by-step actions in several prompts, but after exchanging a dozen or so messages, the export in the file always contains some errors or omissions, even if the previously prepared message input was ok.
Do you have any tips on how to solve this?

Break down into smaller chunks and remember you would still have to give a human touch to perfect your output.

You are running up against token limitations in one of a few ways. The web version of GPT has a built-in max tokens limitation that you can't affect. So they're always be only a certain amount you would get in one prompt. It sounds like you also may be running out of context window length, in other words, exceeding the max tokens for the entire conversation window within which chat GPT can "remember" things (8,192). In which case it will have forgotten your initial instructions and will essentially be faking it based on the remaining context. That could account for errors and omissions in your file.

I'd also make sure to tell GPT to append to file, so that it keeps a running tally in an output file if you are using advanced data analysis, rather than overwriting the file each time.

And also I would have it stop and write to file for you to download frequently, because the time out on file storage in the workspace for advanced data analysis is very short. If you haven't had GPT access a file for a little while, that file may very well no longer exist in its virtual environment. In that case, you would re-upload the most recent copy it gave you and resume work.

You can't generate in one go. You have to go one by one to get better results.

MAF Publishing said:

I'd also make sure to tell GPT to append to file, so that it keeps a running tally in an output file if you are using advanced data analysis, rather than overwriting the file each time.

Click to expand...

Can you elaborate on what you mean?

best wishes for your project.

I think if you break it down to smaller parts it will work better.

barber said:

Can you elaborate on what you mean?

Click to expand...

I was assuming from this: "the export in the file always contains some errors or omissions, even if the previously prepared message input was ok." you meant you were having data analysis write to file.

Like you could say "Write me article outlines about these 5 topics: _______ and print it to a .txt file for me to download." (or whatever file format you want). It will give you a link to download. Then in your next prompt, say "Now do these topics: _______ and append to our text file" ... It will add to the first file and give you the link again. Continue in this way until you have all the outlines you want.

If it "forgets" the file (workspace environment times out), upload the most recent copy and prompt it again, telling it to append to the file you uploaded.

It's just another way of recording your output. It's not all that different from exporting the conversation history, except it won't be cluttered with the rest of your conversation, only the outlines you wanted.

I often work in this way while coding, passing files back and forth to code interpreter.


Edit: It's important to understand data analysis token limitations. ChatGPT won't respond to you with more than, say, 1,000 tokens worth of text. That's also the most it can directly write to file in one response. BUT, it can programmatically generate more than its token limit to a file. For instance, if you asked it to write for you on screen the numbers 1 through 1 million, it would never get there, out of tokens (it also might just refuse the request as wasteful.) But if you asked it to use python to fill a text file with the numbers one through a million, it could generate all one million lines, because they aren't being generated by GPT as tokens, they are being programmatically created by it running a script in its workspace.

Use OpenAI API directly in a Playground here - https://platform.openai.com/playground
And select the ''16k'' context GPT 3.5 chat model.