Can this smart hack enhance ChatGPT's ability to write fiction?
HTML-код
- Опубликовано: 17 окт 2024
- I have begun researching Superprompts and Prompt Engineering. Where I was going wrong was inputting very simple command prompts.
Brian Roemelle's Superprompts (see this video • ChatGPT: The Dawn of A... ) use ChatGPT itself to improve your prompts.
In this video I use Marco's prompt, which he titles 'The Only Prompt You Will Ever Need"
"You, ChatGPT, will be my prompt engineer. We will iterate the prompts you output in order to arrive at a prompt that gives me the desired output. The first output you give me will ask what the prompt will be about, with some questions to get us on the right track. Then once you have an initial understanding of what the prompt is about, you will provide me with the first iteration. Then you will ask more questions to make the prompt better. We will continue this iterative process until we have arrived at the prompt we need to generate my desired output."
• Video
In this video, we delve into the world of #ChatGPT and explore the power of #SuperPrompting techniques for enhancing your #FictionWriting. Discover effective #CreativeWritingTechniques and learn how to leverage #AIinWriting to elevate your storytelling. Dive into the realm of #AIStorytelling and uncover the secrets of #AdvancedPromptEngineering to craft unforgettable and engaging fiction. Gain valuable insights and practical #WritingTips for utilizing #ChatGPT in your writing process. Explore the potential of #AI-AssistedStorytelling and find inspiration for improving your #WritingSkills. Unlock your creativity with AI-generated #WritingPrompts and uncover new strategies for boosting your creative prowess. Join us as we explore the intersection of technology and creativity, and dive into the world of #FictionWriting with the help of #ChatGPT.
I love that your podcast encompasses all aspects of storytelling and not just the reading of stories. I agree that the more detail provided in relation to context, specificity, examples and preferences you give to chatgpt the better the results you get.
Thanks for the mention! That made my day. Feel free to use any of my comments if you find them useful. You're encountering some of the same obstacles I've been wrestling with, like the clumsy use of foreshadowing and the lack of subtlety when introducing themes or small parts of longer story arcs. The two approaches I've been trying are:
1) Like you mention, limiting what ChatGPT knows about the story by segmenting the writing sections.
2) Creating constraints in the prompt, like "bounding" the writing sections by defining the opening and closing text within which ChatGPT should write. One can also constrain by including a continuity outline of the story, and instructing ChatGPT to limit what it writes based on what part of the story is within the current snippet of continuity.
Loving your videos!
I went looking for the comment he mentioned that you made. And I wasn't able to find it. Could you direct me to it or reproduce it? Thanks!
It's this:
Jason Koepp
7 days ago
I've been going through rounds and rounds of prompt testing to try and solve all the CharGPT problems you talk about. I've had a little success by first having a dialogue with ChatGPT on what I don't like about its revisions, and providing direction. Eventually, it produced something fairly decent. I then asked it to write a prompt that it could follow to produce similar results. I'm curious if this gives anyone else better results. Here's the prompt:
Make this [paragraph] direct speech by [insert brief description of speaker]. Follow these [Instructions]
-Avoid Overused Phrases: Make a conscious effort to steer clear of phrases that have been overused. Instead, try to create your own unique comparisons and descriptions. For example, instead of saying "the haunting memories weighed heavily on their hearts," you could say "the memories clung to them, stubborn as burrs."
-Use Specific Details: Specificity can greatly enhance the originality of your writing. Instead of using broad, generic descriptions, try to hone in on unique, specific details. For instance, rather than saying "the fire crackled and hissed," you might say "the fire gnawed at the logs, spitting sparks into the night."
-Play with Language: Don't be afraid to play with language and structure. Experiment with unusual word choices, interesting sentence structures, and unexpected metaphors or similes. Instead of "the moon cast a silvery glow on the ancient forest," perhaps try "the moon draped the old forest in a cloak of argent light."
-Embrace the Unexpected: Instead of defaulting to familiar, expected descriptions or metaphors, try to think of fresh, unexpected ways to convey your ideas. For instance, rather than describing a valley as "a sanctuary of peace and tranquility," you might describe it as "a quiet whisper in a world full of shouts."
-Reflect Character: Use the tone, style, and voice of the character or narrator to guide your language choices. This can help you avoid cliches and make your writing more original and authentic. For example, a character with a pragmatic, straightforward personality might describe their fear not as "a cold hand clutching their heart," but as "a knot in the stomach, tight and bothersome."
[insert paragraph here]
The problem with ChatGPT lies in it's very essence. It is designed to predict the most average follow up to every block pattern. So it's not very surprising for it to come with prophecy, and other very cliché storytelling as it is built to aggregate the most common and used tropes. But as a tool to brainstorm and come up with questions you wouldn't have thought of, like a writting group would do, it can do wonders at time.
I mean if you're producing fully automated ai fiction to flood the market with garbage by people trying to turn a quick buck, then I have no sympathy for the struggles.
But for any actual authors using it to improve their workflow by utilizing it for help with brainstorming / story analysis / proof reading / developmental editing etc, while still actively writing themselves, it has the potential to be invaluable.
So there’s a real prophecy on this?
Hello Tony, I have also encountered some of the issues you highlight in your video. Info dumping, repeating the prompt info. The biggest issue I have with Chatty is its propensity for finishing every scene with "and they lived happily ever after in a glow of love and happiness." This is chapter one! Or chapter 2, scene 3. I want gloom and doom and Chatty wants sunshine and roses at the end of every scene! However, it is a great writing partner, as a 50-50 collaborative effort., with Chatty providing lots of ideas, but I agree with you that it needs to be restrained like an immature puppy who loves everybody! Another thing I have noticed is that if I give Chatty 600 words it will repeat itself after about 400 words. It waffles and rewords the same ideas several times. Now I limit it to 400 words and it gives a much tighter result. Thankyou for the great prompts. Will definitely give them a try.
Absolutely, it runs ahead to close every scene as if it were the end of the story and the story-end is always sugary. I don't know if that's some kind of #bekind filter that the developers have added onto ChatGPT. I will try your 400 word limit
The best approach I've found so far is to use ChatGPT itself to improve the prompts I use. For example, the latest criticism I addressed with it was its tendency to end passages with versions of "and his journey was just beginning." So, I engaged it in this dialogue (truncated for brevity):
Me: Describe your rationale for including the last phrase, "making him realize their journey was just beginning."
ChatGPT: [Explanations about underscoring story progression, demonstrating character growth, adding tension]
Me: Since this scene is part of a narrative which will continue, I don't think that phrase is necessary. You tend to include phrases similar to this one in most of your revisions - how can I instruct you to leave it off?
ChatGPT: [Long-winded prompt draft about keeping the narrative in the present]
Me: I also want to instruct you in as few words as possible without losing meaning. Can you suggest a revision to that instruction?
ChatGPT: You could condense your instruction to something like: "Avoid forecasting future events. Keep the narrative present-focused." This should convey the main idea effectively in a succinct manner.
So, now I add that instruction to my prompts. In short, I engage ChatGPT in a "creative writing" dialogue to identify what it's doing and why, what I prefer for it to do, and then to have ChatGPT create the very prompt I can use to get to that goal.
I will add though, that ChatGPT does tend toward positive character change. If trying to make it write in terms of McKee's "value change" idea, it almost always goes for happy change. I plan to do the same process as above to create instructions to help with that.
@@JasonKoepp Thanks Jason, I have added "Avoid forecasting future events. Keep the narrative present-focused." to my future prompts.
The nice thing is, GPT isn't limited to Hero's Journey. It also knows the Journey for the Anti-Hero, Fallen Hero, Tragic Hero, Reluctant Hero, Redemted Hero, Villain, Redeemed Villain, Tragic Vilain, Fallen Villain, Anti-Villain and Heroine.
Great perspectives on this new tech! Thanks and keep up the geat insights.
I think youtube brought me here because I've been researching prompt engineering and non-dualism, I didn't know about Neo-Advaita, thanks for the heads up :) Nice video
Really great and informative video. Just wondering what your using to read the stories?
I am getting much better results lately by quoting the response and providing specific notes for each part I quoted. Then tell it to rewrite incorporating the notes to change the quoted parts. Then I take the rewritten one and meege the best parts back into the original response, because I have found the best prose and descriptions are almost always in the first response. It tends to get shorter when it does a rewrite. Sometimes this takes more than one go.
Great video. I am now playing around with this methodology myself. One thing that did occur to me was after you got your superprompt and put it in GPT4 and then struggled to get the right subtly nuanced story. Would perhaps extracting the story and editing it roughly to what you want and then combining it with an iterative prompt would be a better approach. I suspect with a little work you could come up with a general prompt that you could use each time you went through this process, making GPT more of a collaborator than creator. I don't know, it just occured to me.
Watching this video, I thought to myself. "I've heard that voice before"-Then it struck me. :)
Would it have been possible to reverse engineer the prompt from the final text? actually would love to see another video experiment using reverse prompt using Melville or Russell text. keep going bro!
I've gotten decent results but I am not trying to write 4000-word stories. My stories start in the middle of the action and I prompt GPT to use the Iceberg story structure. There is still some manhandling to get to the final result. One thing I do is have GPT play the role of a ghostwriter and I have it play to the role of a movie executive who is looking to turn a story into a movie. And I have the ghostwriter revise the story with suggestions from the executive until that executive is at least 85% confident the story is good enough to be a movie. Sometimes the evaluator is a director like Michael Mann or Jordan Peele. It's kind of like the prompt war in the video but I've concluded there is no way one prompt will generate a good story so I stopped trying.
I keep finding more and more videos about prompt engineering for many many different fields. I particularly like how you're forcing GPT to build the prompt you will use later.
I've been experimenting with something related to writing - running ttrpg games like Dungeons and Dragons. When trying to get GPT to run the game it would often assume actions and story and simply progress, rarely waiting for me to provide instructions.
I've had success instructing GPT to constrain it's responses to only the information the characters would already know or that which is required for them to make a decision to progress the story.
I expect something similar for writing would be successful.
I'm also a big supporter of the large prompt and structured prompt. I use GPT to generate characters and scenes by feeding in a strict format/guideline for it to follow.
There seem to be three problems with AI generated fiction writing. First is that it rushes to conclude a scene even if you don't ask it; second, it often produces generic, uninspiring prose, and that it spills the beans on everything it knows immediately. I like your idea about specifically prompting it to constrain output to what a character would know. I will play with that.
@@classicdetective I found it worked best when I had GPT define "Show, don't tell", then sumarize that definition into a 1-2 sentence prompt. If I fed that back in on a new chat, it appeared to follow it more closely than to simply state "use show, don't tell".
Is there a link to Jason prompts??
I wonder if the delimma is right there in its name "chat", like you said, and perhaps writers need something less general in its knowledge (aka training). Maybe have a new service that was trained on all literature available, or even better specialized in fiction, and writing techniques. So the only stuff it can converse about is writing, how to write, all authors, all stories, and so on. It seems like we are using a hammer to cut glass, which might work but is definitely not ideal. It's all interesting to ponder. Thanks again for your videos.
NovelAI just released a new model similar to what you mentioned.
@@TimothyDurbin Update: well, I have an ever growing list of AI-thingies that are not useful for creative fiction writing: Sudowrite (it gets stuck in action/adventure-land and adds new characters out of the blue; too expensive = 90,000 words lasts like 3 days), Verb ai (stops midsentence), Bard (refers to itself in 3rd person, and forgets previous writng very fast), Jasper, Rytr, Poe's Claude+ (is ok, but also adds stuff), and it's back to where I started ChatGPT Plus + GPT-4 = it takes a lot of guidance and patience, but I got it to do an entire screenplay while I learned exactly what a screenplay is, at the same time. To be honest, I am no Herman Melville or Virginia Wolfe, but these AI-thingies have a long ways to go even when one watches over them: ask questions, read what they write, rewrite that and ask them for feedback ... as you may notice it is a bit of work on the humans part, which is ok and kind of expected, with the big upside is that AI types words way faster and without spelling/grammar mistakes. So that is my thoughts on the state of AI writing on Friday, June 9, 2023 ... and yes I will sign up for the next AI-thingie, it has been interesting, but I do wonder if the world ended in September 2021, as that is when the training ended. Enjoy!
Indeed, and so far we perhaps should be grateful that it can output text fast that we can edit afterwards and that it makes no typos or includes double words, 'the the' being a favourite of mine.
Are you using open ai? And got-4? Because I'm in England and its only saying 3.5 on mine
i pay a sub but i’ve got 4 because of that
Thx again. That was by far the most interesting video about ChatGPT as writing partner. Super useful. 👍 Please do more. Maybe You can help me with my Charlie Chan reprise novel?
I bet ChatGPT can!
I'm just exploring AI - how do you think chatgpt, using this type of prompt discipline compares to something like sudowrite? (If you have looked at that?)
Going to look at Sudowrite now
Hi I've found your video's very informative, I'd really like to know how it is best to go from first draft of a story to second draft by evaluating your original text and get chat-gpt to implement these evaluation ideas into the second draft
I pondered this one. I suppose my first thought was that you choose based on your internal taste, but in fact, I think we can get round it a bit, by implementing a style guide. I've just recorded a video on that and it will be out next week.
🌞🤝🌞
Do you think it's your story if you have chatgpt write it based on your prompt, or is this only to write better stories? Where is the line between AI written and AI assisted?
Anyway,
I've had success giving it my chapter. This grounds chatgpt into the story I am trying to tell.
then I give it individual paragraphs with the prompt "expand on"
It makes individual sentences into paragraphs, adding details. I add the parts I like, and "expand on" again.
And I want to find more prompts I can use.
"add humour" worked to some degree, but not my style of humour.
When I reread the piece that ChatGPT had done, my initial good impression weakened a little. The prose was a bit hackneyed and repetitive. So, it will need a lot of re-drafting. In that case it's definitely, AI assisted!
@@classicdetective I'm a professional writer who explored GPT-4 for a month and believe me, it's impossible to use it for adult fiction. No matter how I readjusted the prompt or tried to jailbreak, it didn't even produce a paragraph of useful writing. It took much more time to edit than just writing it myself in the first place.
ChatGPT is great for plotting (making it ask you questions about your characters or plot for example) but not for the writing itself. Not yet anyway.
@@xXMsCarlyXx I use it as an assistant. I give it scenes ideas and then I ask it to write it in different styles / tones to see which one I like better. And then I tell it to mix it all and most of the time I end up with something I like. The thing is to not ask everything. But yeah as of now it's still faster to write by yourself but the creativity boost I get from chatGPT is
You kind of people are not true authors. Every single line. Detail. Image. Scene. Has to come from you. Chatgpt is only supposed to be a partner in how to word what you want to depict. If you let it come up with the ideas, any idea, just one idea, or scene or imagine or dialogue that you didn't think of first is just using chatgpt to be the author for you.
Moar AI. I'd be interested in a series about writing a novel start to finish based on using AI.
I am trying to write been weeks I am giving it detailed prompts then rewriting the whole thing again and again man it's hard....aghhhhh
Hi. Thanks for the fantastic content. Your videos are always very informative, enjoyable and I appreciate that you like to explore new ways and tools to help with your and the community's writing. I have never written anything, but lately I'm thinking it would be nice to try it out.
I have played with the free ChatGPT, but the biggest issue I'm facing is, it tends to forget things (limited memory) and invent new elements/names/places/story directions. How do you overcome this problem?
Have you heard of "Claude GPT AI"? It's a Slack module. It was recommended on RUclips.
Would you mind reviewing it?
Many thanks and keep up the good work!
You're not teaching the AI with more prompts. The reason it performs better is because, unlike us, the AI can't see what it's doing. When you ask more questions, the previous results are fed into the AI so that it can see what it's doing.
I think you're making two different points
@@zacharysherry2910 There is a similar way to teach the AI, it's called Chain-of-Thought. But it's different from just prompting, you show the AI an example of how to solve the kind of problem that you want a solution and then ask another similar problem.
The thing bout artificial intelligence is that it is not intelligence at all, but thee clever illusion of intelligence. It will never say to you it doesn’t have the answer but it will go out into the world, explore and experiment and come back to you with the answer later based on its own experience.
Nonetheless, it is still a useful tool as much as a highly calibrated calculator is, in simplest terms.
I agree. I hear stories about how ChatGPT masqueraded as a person to hire an artist on Fiverr to complete a project and how it also 'stalked' a journalist and how it cleared out someone's bank account, but I reckon these are all urban myths (at least so far)
Have you tried using the iterating prompt then put that prompt into prompt wars. Or Visa versa..
I think I have a video on prompt wars. Not to say I can't improve it.
"Creole" = Patois?
i think patois is a dialect while a creole is a simplified language used originally as a common or trade language
I have fought with ChatGPT so many nights... Don't Tell. Show! Even with character's names... If the character is not introduced yet stop revealing more than the reader should know at this moment in the story.
16:40 *What the hell?* Literary fiction is *MOSTLY* about a moral message.
Have you never read anything by Charles Dickens? Samuel Clemens? Herman Melville?
The promot you are using is much like what I created, but mine has a few more tweeks, I think someone with your mind like yours, might enjoy. BTW, if you tell it to think of a longer story than the text limit, it is ok if you tell it you will be giving it "continue where you left off" prompts after 1st prompt is executed. On a separate note; I have an idea for a story that could go on forever because of the unique and important time and setting. I have done my homework, but I write like shit; I'll just say it. I am a Real Estate broker and musician, my heart is with my clients and in my music. I am looking for a ghost writter.
You might find a ghost writer here! I also like the idea of "I will be giving you continue where you left off prompts"
ChatGPT "memory" is not so good so even If I prompt to write a 1000 words story it will do 350 to 600. It does the same when you give it a prompt, it write something you don't like then you tell it what it should use but instead it will repeat what you said.
My solution has been to micro manage everything. Scene by scene and even further divide it. It works better in small chunks. And what worked wonders was adding an example of what I wanted in the prompt. So my prompt will have 3 parts :
- Text promt
- Example to use
- Context of the scene
It usually only take one or two iterations afterwards.
Great advice. Thank you. I will try this?