FYI 2: Prompt size, Pollyanna, and a Plot, three ChatGPT 3.5 hints

My newest composition, (which, as I eluded to in the last post, was called Dt. Doe and the Secret Society but according to ChatGPT, is really) The Mansion on the Hill, was an experiment in having two interlocking story lines, culminating in a conclusion resolving them both. While the story is perhaps a bit simplistic, it came together pretty well and I learned quite a bit from the process. Incidentally, the 9,000 or so word result took about three hours.

The first thing I learned is what to do if a network error occurs despite your best efforts. It could be due to too long of a response (hence you need to use the pause command as I mentioned earlier), a system time out, or even a network glitch. Regardless, you’ll get the dreaded error notice. It happened to me during this session.

But not is all lost. I found that the prompt buffer has no size limitations. That means you can put the entire story you’ve gone through to that point and have the AI digest it in a new thread. Then you tell ChatGPT to take all of into consideration in its responses. I found it helpful to put any command at the beginning, and end with the story itself. Then continue from there.

In fact, I managed to use the entire composition of Dt. Doe and the Secret Society into the prompt which is how I discovered it was supposed to be called The Mansion on the Hill. I prompted:

You are a writer for the New York Times Book Review. You have been given a book to read and asked to write up a book review, discussing the characters, plot, and setting, and then giving a recommendation as to what type of readers might enjoy it. When you get to 500 words, pause unitl I reply with “continue.” Here is the story you are reviewing….

Which brought about ChatGPT’s review of a composition that it created:

In “The Mansion on the Hill,” Detective Jane Doe is called to the scene of a suspicious death at an old, dilapidated mansion. Upon arriving, she finds the victim, a middle-aged man, strangled in his cluttered bedroom. As Jane investigates, she is aided by the victim’s sister, who seems nervous and anxious. Jane also questions the victim’s family and close friends, but their alibis check out and their stories contain inconsistencies. Despite feeling frustrated and exhausted, Jane is determined to solve the mystery and uncover the truth.

The characters in this story include Detective Jane Doe, the victim, and the victim’s sister and family and friends. The plot centers on Jane’s investigation into the suspicious death at the mansion, and the setting is the old, dilapidated mansion itself, which adds to the atmosphere of mystery and unease.

This book would likely appeal to readers who enjoy mystery and crime novels, particularly those with a detective protagonist. The setting of the old mansion and the presence of a mysterious symbol add an extra layer of intrigue to the story, and the search for the killer keeps the reader guessing until the end. Overall, “The Mansion on the Hill” is a compelling mystery with well-developed characters and an engaging plot.

So, I went back and changed the name of the composition. And I learn that you can put 9,000 words into the prompt with no issue.

The second thing I learned was what a Pollyanna view of life ChatGPT can put into its compositions. It has a tendency to want to end on an upbeat note or some “happily ever after” feeling. And sometimes it tries to end the process by saying itself, “The End.” It’s quite comical when that happens. I had to rerun a couple prompts to get slightly different endings. Otherwise everything immediately turned out roses and sunshine and the day was saved.

On the one hand that is a good comment on humanity in a way, since ChatGPT has access to all of the world’s writings up through 2020. Overall I think it says that we humans like happy endings. Or else the high priests over at OpenAI have their “bad ending” filter set a little too high.

The third thing I learned was the need for an actual plot structure. That would definitely help guide the process. In most writing I’m more of a plantser, not as organized as a planner but not quite as open to anything pantser (e.g. writing from the seat of my pants). It definitely pays off to have some idea of where it would go at the beginning.

Also, I’ve been wondering what to call the creations. With MidJourney (a generative AI that makes visual art such as this), I ended up calling them Imagines:

The final term I’ll suggest deals with the most controversial of subjects: what to call the final output of an AI-art generating application. I have steadfastly used the term Imagines. This was coined in part because the most popular AI-art-generator, MidJourney, uses the command prompt /imagine to initiate the process. I submit it is preferable to refer to the images as imagines rather than drawing or painting or digital photographs because while the output of the AI can mimic these quite amazingly (and sometimes, quite atrociously), those other terms bring in a lot of other connotations–including, but not limited to, something that the human artist actually made. In the case of the imagines, it is the AI which is making them. The human artist simply imagined it.

I propose that we use “composing” as the verb when describing what ChatGPT does, and composition as the end result. This wasn’t strictly my idea, I asked ChatGPT for its options:

Q: What are three verbs to describe the creation of a story via ChatGPT?

A: There are many different verbs that could be used to describe the creation of a story using ChatGPT or any other language model, but here are three possibilities:

1. Generate: ChatGPT can generate text based on the prompts and input provided to it. This process involves creating new content based on the input and outputting it as text.
2. Write: ChatGPT can be used to write a story by providing it with prompts and guiding it towards a desired outcome or plot. This process involves using the language model to produce text that is cohesive and flows well.
3. Compose: ChatGPT can be used to compose a story by providing it with prompts and guiding it towards a desired outcome or plot. This process involves using the language model to produce text that is well-structured and follows a logical sequence of events.

ChatGPT when on to note that it is only a tool that can assist in the creative process, and it is not capable of creating a story on its own. The user must provide input and guidance to shape the output of the language model into a cohesive narrative.

I think composed is a great term, and will be editing my prior entries to reflect this.

With regard to the specific story, the initial prompt was as follows:

Write the opening scene of a murder mystery. Start with a shocking or mysterious event, such as the discovery of a body or a cryptic message. Then Introduce a compelling central character who the reader can relate to and root for. Set the scene with vivid and evocative descriptions that establish the mood and tone of the story. Pose a tantalizing question or puzzle that the reader will want to solve. Be sure to use a unique or unusual setting to add interest and atmosphere to the scene. If you get to 500 words, pause until I type “continue.”

I think ChatGPT did a great job starting off the story:

The sun had just set behind the old, dilapidated mansion on the hill, casting long shadows across the overgrown lawn. Detective Jane Doe stood at the foot of the grand staircase, staring up at the imposing structure with a sense of trepidation. She had been called to the scene of a suspicious death, and the atmosphere of the place already had her on edge.

The interaction with ChatGPT was a bit different this time. I was able to get the AI to compose a continuous story for a few session, until it would try to wrap things up. After the initial prompt I had two simple write the next scene. When you get to 500 words, pause until I reply with “continue” prompts, which took me to the arresting of the suspect. At that point ChatGPT wanted to pretty much be done:

As the suspect was led away in handcuffs, Jane couldn’t help but feel a sense of satisfaction and relief. She had finally solved the mystery, and she knew that the victim’s family could now find some closure.

The case was far from perfect, but it was a start, and Jane knew that she had made a difference. She was proud of the work she had done, and she was already looking forward to her next challenge.

You can almost hear the “the end” at the end of that. A couple of times during this project ChatGPT did, in fact, just put The End.

At the conclusion of the above scene, I had to add some additional details to the prompt on my end:

Write the next scene when Jane finds out the suspect had suspiciously died while in custody. When you get to 500 words, pause until I reply with “continue”.

That gave the story a twist and allowed it to continue on.

Having established Dt. Jane Doe, I then turned to the other story line. That of the leader of this Secret Society that the AI came up with. This allowed me to weave in the second story line.

Write the next scene from the perspective of the leader of the secret society who had the suspect killed. Give his name and a description, what his role it and what the purpose of the secret society was for. What does he think about Jane. When you get to 500 words, pause until I reply with “continue”.

I had played around with multiple storylines in some other testing, so knew it could be done. I let the AI do some more creation as seen above. I also went for a subplot of sorts, bringing Ryan into the picture.

Write the next scene from the perspective of Jane. She has been on a date with someone she recently met. Give the person’s name and a description. She is telling the person about the case and what happened to the suspect, when she comes up with an idea. It is a wrong idea, a red herring, that distracts her from the real truth. When you get to 500 words, pause until I reply with “continue”.

And then Mark:

Write the next scene from the perspective of the lieutenant of the secret society, he is the one who killed the suspect. Give his name and a description, and why he is willing to kill for the secret society. Somehow he arranges a meeting with Jane and plans to kill her but she narrowly escapes. When you get to 500 words, pause until I reply with “continue”.

Through the use of a bit of direction I was able to get through the entire composition, building the stories to a climax and then relenting and letting ChatGPT end on a happy note.

There are a couple of things I’m planning or currently working o:

1. My next composition will be using the Hero’s Journey format since is already so well-designed. My strategy is to ask ChatGPT to come up with a synopsis and then work from that structure letting the AI determine the story.
2. Using ChatGPT when playing a game of Microscope. Microscope is a light roleplaying game that is collaborative world building. I’m considering some changes that could be made to the Microscope game and then creating a storyline with it.
3. I’ve been experimenting with putting some of my previously written pieces (such as poetry) and requesting a review on how to improve it, then making the changes and resubmitting it to the ChatGPT in the same thread. It would be interesting to see if ChatGPT would remember it had suggested the changes.

Thanks for reading.

Original published: https://medium.com/@winstoneverlast/fyi-2-prompt-size-pollyanna-and-a-plot-three-chatgpt-3-5-hints-93821dca51d5