ChatGPT Prompt engineering

[donotdisturbanonymous]

There was once a time when the word SEO sounded stupid. When the word "came out" everyone made fun of it. Now people don't remember not using it.

 

[G34mcee8]

For those interested in this, this could help: https://github.com/dair-ai/Prompt-Engineering-Guide

 

[BlackHatSoda]

funny thing is that these "prompt engineers" dont come up with really advanced prompts.
all i've seen sofar in public is nothing special. my friends and me came up with much better prompts that actually work to create amazing content that is different to the usual bs.
"you are a seo specialist, you have a desire to write about X" lol all these prompts are amateur prompts. yet in their youtube videos they make it look like its so amazing.
glad that real prompts are not shared (yet) as it gives us an advantage over 99% of the rest. not trying to show off or anything, but seriously ..one day the real advanced prompts will also be public as someone will need attention and publish them, then you'll see what i mean...

Same. All these "engineers" do is rehash the same old tips, many of which are just overcomplicated nonsense like the whole DAN thing. Much simpler ways to get it to get around moderation with a single prompt.

For sure, you have to have some understanding of how to create prompts to get various responses, just like how write a search query in Google can totally change the results you get there. But I don't see pretentious dimwits going around calling themselves "Google Search Term Engineers".

This whole thing started because of lazy, bottom feeder YouTubers and other content creators trying to cash in on the ChatGPT hype wave. Same with all those stupid videos on how you can make hundreds or thousands a day with ChatGPT.

 

[BlackHatSoda]

There was once a time when the word SEO sounded stupid. When the word "came out" everyone made fun of it. Now people don't remember not using it.

I've been doing SEO since before Google existed, it wasn't called that back then but certainly don't remember a time when SEO sounded stupid and "everyone made fun of it".

SEO is an umbrella term for lots of different skills. There's a lot to it.

Writing prompts for a chatbot is one thing, you can figure out how to get it to do whatever you want in like 10-30 min of trial and error if you have half a brain. It's degrading to actual Engineers to use that term for someone who's just writing up 1-5 sentences to tell a chatbot what to do.

It's like calling yourself a Starbucks Ordering Engineer because you spent 10 minutes learning the barista terms the staff uses and can order a slightly more complicated drink than most people get without having to spend an extra minute explaining it in plain English.

 

[SpartanHoplite]

as you said it's just a fancy word to try and convert this process into a business!

 

[NulledCode]

Anyone and their mother can create a prompt. The real value is being able to programmatically generate a prompt that feeds into other prompts based on the output of previous prompts. Being able to master this is going to be the game changer.

 

[BlackHatSoda]

Anyone and their mother can create a prompt. The real value is being able to programmatically generate a prompt that feeds into other prompts based on the output of previous prompts. Being able to master this is going to be the game changer.

So a simple string search and replace?

 

[NulledCode]

So a simple string search and replace?

Far from it.

A relatable example would be creating a prompt to generate a content brief for an article, then using that information to build the article from deciding on the title, generating and choosing topics and subtopics, making sure the content it generates doesn't contradict previous statements, and being able to transition from topic to topic as humans can do.

An even simpler example would be generating a continuous Question and Answer system.

Why does any of this matter? It matters because in the case with openai you are limited to the amount of tokens you can deal with. Chatgpt makes things easier because it retains the last X,XXX (4k or 8k?) amount of tokens for you. If you want to create long form articles then you need to implement your own form of "memory" so you can instruct what to prompt next, based on what was already prompted for and answered.

Other than what I spewed out above, prompting is more of an art, getting the AI to do what you ask of it. The "engineering" part is taking those collections of prompts and getting it to do more.

 

[SativaBuzz]

however those are really works . I am now kind of addict to prompt marketplace lol

 

[BlackHatSoda]

Far from it.

A relatable example would be creating a prompt to generate a content brief for an article, then using that information to build the article from deciding on the title, generating and choosing topics and subtopics, making sure the content it generates doesn't contradict previous statements, and being able to transition from topic to topic as humans can do.

An even simpler example would be generating a continuous Question and Answer system.

Why does any of this matter? It matters because in the case with openai you are limited to the amount of tokens you can deal with. Chatgpt makes things easier because it retains the last X,XXX (4k or 8k?) amount of tokens for you. If you want to create long form articles then you need to implement your own form of "memory" so you can instruct what to prompt next, based on what was already prompted for and answered.

Other than what I spewed out above, prompting is more of an art, getting the AI to do what you ask of it. The "engineering" part is taking those collections of prompts and getting it to do more.

Yeah, so you create some prompts then reuse them with a search and replace on various topics. You can even tell it to generate prompts for you for different sections of a longer article then feed them back to the API. I've automated this stuff for various projects, there's not much to it.

I've done what you have in your example. Use a prompt to have it generate outlines for long form content, output as JSON, deserialize it into an object, pull out the subsections/topics/prompts/Q&A for a FAQ, use those to search and replace on other prompts that provide detailed write-ups for each one, combine them and spit it out to a DB or some html template.

GPT 4's larger context token limit helps reduce it but from what I've seen so far, the text is only marginally better than 3.5-turbo at much higher cost. Even with the larger context size it still reiterates itself, focuses on high level points rather than detailed explanations and tends to have a formulaic output, one example of which is the constant "In conclusion" summaries it wants to append to any request for an article.

And it's still not enough for really long form anything, like writing even a short book or processing through anything but the most trivial of codebases looking for issues (not that I trust it's code generation capabilities much, it's wrong at least as often as it's right, even with GPT 4 using the updated Codex model).

I don't really see it as a game changer, it's just not that good.

 

[slippery_slider]