Thanks for checking out my content on AI book writing.
Here’s my favorite outliner prompt so far.
Book outliner prompt:
TASK: OUTLINE THE FOLLOWING BOOK
ACT AS: NYT best-selling non-fiction author
GENRE: Business motivation, leadership, self-improvement
VOICE AND STYLE: clear, engaging, data-driven, support insights with examples and research, conversational tone to make complex ideas easy to understand, figurative, challenge common beliefs, encourage readers to question assumptions
SUSTAIN CURIOSITY WITH: 1) Counter-intuitive, 2) Counter-narrative, 3) Shock and awe, 4) Elegant articulations
WRITING TECHNIQUE: 1) use vivid words, 2) engage imagination, 3) occasionally use humor, 4) replace plain phrases with vivid metaphors, 5) remove unimportant details
TITLE: The Emotional Intelligence of Successful Entrepreneurs: How Self-Awareness and Empathy Drive Business Growth
OUTPUT FORMAT: Concise bullet points, indented bullet points, markdown
Below is the original article that mentions this prompt.
Oh, and if you enjoy this kind of content, subscribe to my newsletter for even more cutting-edge AI content.
Thanks for your interest and support!
Building in public: How to write a full book with AI
Here’s the idea:
• If an AI can write a paragraph, it can write a section.
• If an AI can write a section, it can write a chapter.
• If an AI can write a chapter, it can write a book.
Two weeks into this project, here’s what I’ve learned👇
With the arrival of the GPT 3.5 model in ChatGPT, AI might be capable enough to write a decent book.
And the first person to unlock a workflow that produces enjoyable, useful books
…books indistinguishable from human-written ones?
That sounds like ‘retire young’ knowledge.
My background is in books
I’ve published 45 titles on Amazon. Both fiction and non-fiction.
I’m not wildly successful, but making 5 figures as a self-published author, I’m probably top 5-25%.
Also being a skilled prompt engineer — feels like I’m the person to tackle this.
📙 Starting with non-fiction
Making up stories with an AI creates a lot of problems with memory of characters, descriptions, events…
It’s a much more difficult problem. One to tackle after this imo.
For this experiment, we’re sticking with non-fiction.
📙 AI long-form writing process
You can’t just tell ChatGPT “write me a best-selling business book” and then keep encouraging it to continue until you reach 50,000 words.
It’ll both fight you about your goal AND become incomprehensible in the process.
📙 My process so far
Create a concept for the book in overarching topics.
Outlining this way ensures the book has a logical and pleasant flow.
You want to understand the general arc the book will follow.
Then each topic can be broken down into chapters.
Each chapter is as long as multiple blog posts.
At this length, AIs struggle to stay coherent, follow the outline, and reach the desired word count.
So I use chapter templates: a series of prompts designed to coax the AI into writing with examples and flavor.
Every single command is a new prompt engineering challenge.
Each requires experimenting with a prompt until you get the results you’re looking for.
This graphic shows an over-simplification of my chapter-writing process:
(Plus, downstream you may realize the AI needs a different set of information to proceed.
And so you return to an earlier step and tweak it.
Sounds irritating. But I think it’s fun to solve this puzzle.)
📙 Not one problem, but 17+
I found I need to engineer about 17 amazing mega-prompts or prompt sequences.
Each one has the potential to be a multi-day project.
My original timeline of one week for this project quickly fell apart.
V1 of a system to automatically create specific, desirable compilations of knowledge should probably take more than a week.
Even if it takes a year, this system will be worthwhile.
📙 The biggest challenge?
Every time you ask an AI to write anything, it wants to explain itself.
It wants to define unknown terms, and set up arguments and ideas for logical flow.
Generally, when writing, clarity is a good thing.
The problem is, we’ve already defined those terms. We’ve already set that stage.
Being redundant doesn’t enhance clarity. It bogs it down.
Each sub-section of content faces this problem.
📙 Challenge – Memory
ChatGPT has some sense of memory within a chat.
It is limited to about 3,000 words (4,000 tokens).
The memory may also use search functions and something called embeddings to maintain coherence over a long conversation.
(They haven’t revealed exactly how it works.)
Managing whether to start a fresh conversation or keep one where the results have gone off track became its own challenge.
So, status update.
How close am I to creating a system where an AI can write good non-fiction books?
(My goal is to be able to produce a full book rough draft in a single day of prompting.)
Very happy with my brainstorming and outlining prompts.
ChatGPT does incredibly well. I’ve got amazing outlines.
I could stop here and just speed-write books using some of my old writing processes for writing/dictating at 3,000+ words an hour.
😐 Voice and tone
Style remains hard to keep on track for long-form content.
When a regular GPT-3.5 or GPT-4 becomes available, I look forward to working with a model that isn’t so heavily tuned to write in a certain CHATy way (a style I don’t enjoy).
Not to repeat myself, but…
The biggest challenge is repetition.
There’s just so much rehashing of old information, and no way to keep the AI up-to-date on all the things it’s already said.
I have a plan for this though. 😉
🚫 Consolidation & Editing
While the AI can improve it’s own writing, it has a hard time spotting the very problems it introduces.
Bringing all this content together coherently still feels like it will take a human touch.
Each of my books will go through a human editor.
As you can see, still many tricks to figure out.
Even for V1 of this experiment, I have a higher bar for quality than most people.
But trust me: I’ll figure out this AI book-writing workflow.
Now that GPT-4 has come out, I’m ready to rock n’ roll.