I Wrote a 90,000-Word Novel With AI in 2 Days. You Probably Shouldn’t.
AI is great for some things. Not all things. Join me as I take on this experiment.
Welcome to the Scarlet Ink newsletter. I’m Dave Anderson, an ex-Amazon Tech Director and GM. Each week I write a newsletter article on tech industry careers and tactical leadership advice.
Free members can read some amount of each article, while paid members can read the full article. For some, part of the article is plenty! But if you’d like to read more, I’d love you to consider becoming a paid member!
Meta comment: Holy cow, this article was a bit crazy to write. It’s a hard and complex process to summarize. At one point it was double the maximum acceptable length for an article (it was over 10k words), but I really disliked the idea of making this a two-part article. So I’ve spent a lot of time tightening it up to be short enough to post. Actually, I’m not 100% sure this is short enough, but I’m going to try posting it.
Like everyone in technology, I’ve been playing with AI. But also, like everyone in technology, most of my experience is related to coding.
“Can AI make a decent UI for me?”
“Can AI refactor this complex code?”
“Does AI write efficient code?”
“Is AI stealing the work of other engineers/designers?” (Answer is yes, but that’s a different discussion.)
This focus on coding has put me into an AI bubble, like many in the industry. “Wow, AI is magic!” While coding is important for various reasons, it’s only a small subset of what humans do.
AI influencers (and certain AI tech CEOs) have repeatedly proclaimed the end of human labor in a few short years. Yet Amazon still uses humans to put books into boxes after hundreds of millions in AI and robotics investments. Why? Because many tasks are hard as heck to automate.
I recently had a brainstorm that went like this:
I should do a naïve test of AI. As in pretending I know nothing about AI development. Which isn’t a terribly hard thing to imagine, considering I’m not anywhere close to an expert.
I should start from scratch rather than an existing project. That would allow my audience to follow along much easier.
It shouldn’t be coding. Everyone does AI coding tests. There are literally thousands of posts about using AI to code. I wouldn’t be adding value if I made another AI coding post.
It should be at the intersection of my interests and skills. It should be in my personal Venn diagram, at an intersection of “Can this perhaps be done with AI?” and “Dave has judgment and interest in this topic.”
That got me thinking. Kindle Unlimited is (unfortunately) infamous for AI slop books appearing in their listings. I’ve repeatedly read articles about slop farms generating AI garbage books to generate free money before they’re caught.
Their AI slop books are garbage, obviously computer-generated to anyone with a meat brain. But is it possible to make an enjoyable book with AI yet? Even more challenging, can I get AI to write a book that isn’t immediately identifiable as AI-written?
I realized this was perfectly in my Venn diagram. Not only do I write a newsletter, but I’m an avid reader, and I want to write a fiction novel someday. This is a perfect experiment.
Hypotheses
Overall plot - AI will do an acceptable job of creating a formulaic plot. It won’t be surprising or unique, but it will be acceptable (like the majority of mainstream fiction books). I think the skeleton of the story will be acceptable.
Chapter breakdown - AI will do a poor job of breaking the plot into chapters. It won’t know what plot points make for good chapters, or how to chronologically organize a story.
Narrative - AI will do a horribly poor job of writing narrative. Every line will sound like AI drivel.
Quality - Beyond the problems above, I believe it will run into obvious continuity problems, like re-opening open doors, people telling information they don’t actually know, etc. Because AI doesn’t understand the world.
What does complete success look like? A slightly enjoyable ~80-100k word novel. If my brief experiment could reach the high bar of not bad, it likely means it’s possible to write acceptable fiction for mass consumption. I’ll admit I hope this fails, because the idea of more AI trash content being generated makes me a little sad.
The setup!
I’m going to use Cursor Pro to write my book. It’s a lovely platform, easy to install, and collaborative file editing is easy.
I decided on a rule that I won’t edit any files directly because this is about AI writing. In my mental model of “what happens next,” if AI can actually write books, they will be mass-produced by the thousands. This requires zero human touch.
This will, on the other hand, be human-directed. I want to use AI to write my book. So I will direct it, like a composer of an orchestra. I will come up with an idea, build a framework for AI to follow, and then direct AI to work through the writing tasks.
Here’s a key controversial limitation. I will use auto for the model choice rather than picking the latest and greatest model. I suspect that Opus 4.6 or another bleeding-edge model might do better. But it would cost a fortune to test, and I feel like testing what I would call a normal model. Plus, everything I’ve read about the latest model releases is that they’ve spent a lot of time optimizing code writing. When I see a release stressing that they’ve optimized English writing, I’ll consider testing this again.
To start, I’ve created an empty project in Cursor.
Creating a writing framework.
When you’re coding, you generally start with a framework. You have packages you import, you choose a coding language, coding frameworks, data storage, file structure, etc.
Writing is more wishy-washy. There are people who sit down, write a story from beginning to end, and then edit it. No process there. But there are absolutely frameworks you can follow. I feel like an organized approach is absolutely necessary if you’re going to have AI write.
The core approach I’m picturing in my head is a snowflake style method approach. We would start with a premise, break it into a plot, and then chapters, and then narrative.
Core components of writing a novel.
Premise - AI needs to know the main arc of the book.
Characters - I need characters with stable personalities. I don’t want a cheerful Olaf style comic relief character to suddenly become a moody Edward Cullen.
Overall Plot - The whole book needs to follow a predictable and planned path.
Structure - There’s the famous 3-act structure, the hero’s journey, etc. I should use a well-trod structure for the story.
Chapters - I will need the book broken down into bite sized chunks for writing narrative.
Voice - A book needs a consistent voice. You can’t have Tolkien writing one chapter and Rowling the next chapter. You need a consistent point-of-view, because swapping between first and third person would be dizzying. You need consistent word choices and sentence length. Critically, in all examples I’ve ever tested, AI business writing is obviously AI. I want to find a human-like voice.
The high-level plan.
All that being said, here’s my plan.
Explain to Cursor what we’re doing.
Ask it to setup folders and files for various purposes.
Explain the premise.
Explain a voice to use.
Explain that the structure we’ll use is the hero’s journey (it’s a favorite of mine).
Explain the high-level plot, and ask it to write the rest.
Ask it to create chapters from the plot.
Ask it to write chapters one by one.
Ok, here we go!
