I used a locally hosted AI model, BookyAI, AutoCrit, and Amazon KDP to see how far I could get without actually rewriting the book myself
I did not start The Affirmation Glitch because I had some lifelong burning desire to finally write my first novel.
I started it because I wanted to break something.
Or maybe, more accurately, I wanted to poke at something until I figured out where it broke.
I had purchased BookyAI and wanted a low-stakes project where I could run the whole thing from beginning to end — push buttons, make mistakes, try features, screw things up, regenerate things, and generally resist my normal overwhelming urge to jump in and start fixing everything myself.
I wanted to see what modern AI book-generation software could actually produce if I mostly just… let it.
The result was The Affirmation Glitch: A Novel of Code, Connection, and Unexpected Magic, a 30-chapter cozy fantasy/magical-realism novel containing 126,743 words of actual story content.
And I generated the prose using a locally hosted Llama model — v4.16 17b, to be exact — instead of paying per token to a commercial API.
The generation itself took roughly ten hours. Including BookyAI’s subsequent automated processing, I estimate around 12 hours of machine time.
Then I got up the next morning and spent approximately another 2.5 to 3 hours fighting with formatting, covers, metadata, AutoCrit, and Amazon KDP.
And then I submitted the damn thing to Amazon.
At the time I am writing this case study, the print editions are still going through Amazon’s approval process.
So, just to be extremely clear, this is not one of those stories where somebody quietly uses AI for a year and then emerges from the woods pretending they hand-carved every sentence with a quill pen.
The AI is not the embarrassing secret here.
The AI is literally the experiment.
- Read The Affirmation Glitch on Kindle (available free in Kindle Unlimited)
[AMAZON PAPERBACK AFFILIATE LINK – ADD WHEN LIVE]
[AMAZON HARDCOVER AFFILIATE LINK – ADD WHEN LIVE]
Disclosure: Amazon links on this page are affiliate links, which means I may receive a small commission if you purchase through them. I purchased BookyAI myself through a Facebook ad (their marketing worked!). This experiment was not sponsored, and at the time I conducted it I could not find a BookyAI affiliate program.
The Experiment
My rule was pretty simple:
How much of a commercially recognizable novel could I create while doing as little actual prose writing and rewriting as possible?
I was not testing the now-pretty-common workflow where a writer asks ChatGPT for a rough draft and then rewrites every paragraph until the AI fingerprints are basically gone.
I was not testing whether AI could brainstorm ideas for a human novelist.
I already know it can do those things.
I wanted to test the whole pipeline.
Could I go from vague concept to outline, from outline to more than 100,000 words of prose, then through quality checks, proofreading, cover creation, marketing copy, and finally Amazon submission — while letting AI do almost all of the actual text generation?
What happens if I actually let the machines do their jobs?
Why I Needed a Brain Break
The other reason this experiment happened is that I was burned out on another book.
I have been developing a much more ambitious novel called The Trees Remember, and that thing is… a lot.
- Disease timelines
- Complicated continuity.
- Medical and historical subject matter.
- Time travel.
- Disability.
- Ecological collapse.
- Characters and storylines jumping between different eras.
AI helps enormously with it.
AI also screws it up.
Repeatedly. And not in little cute ways where I fix a word and move on.
No model I have tested has been able to hold all of that continuity together without substantial intervention from me. The Trees Remember is probably going to end up being something close to an equal collaboration between AI-generated raw material and my own research, rewriting, editing, continuity management, and judgment.
It is also emotionally heavy and personal. Maybe based a little too much on my own experiences and observations.
So my brain needed out for a while. I needed something fun. Something simpler. Something where the fate of the fictional universe did not depend on me remembering some tiny detail from 17 chapters ago. So I gave BookyAI a much more straightforward story.
First, I Asked AI What Kind of Novel AI Thought I Should Write
This is where everything started getting beautifully stupid and meta. I asked both ChatGPT and Gemini to suggest book ideas based on what they knew about me. I deliberately did not build some massive custom assistant with an elaborate story bible and 37 pages of rules. I wanted relatively raw suggestions. I gave them information about my professional work and linked them to my websites, including my professional site and ReachingMyDreams.
This was my exact voice-to-texted prompt for idea generation that I sent both ChatGPT and Gemini:
Now give me some more ideas that I can easily make based on what you know about me (my professional site is https://jessicakmurray.com , my vintage site is https://vintagereveries.com and my AI art site is https://reachingmydreams.com ) - just good prompts that will let AI fill in the rest. I think with this sci-fi idea that is so personal to me, I have started with a project that truly tests of BookyAi.And both systems, independently, ended up wandering toward almost the same idea.
Basically:
A woman who works with websites and AI builds a personal-growth or affirmation tool, and the tool begins knowing people far better than it reasonably should.
Which… okay.
Apparently this is what the machines think I should be writing about.
ChatGPT gave me the more psychologically grounded version. In its version, the AI was not actually magical or psychic. It was just frighteningly good at detecting how much information people accidentally reveal through word choice, timing, corrections, abandoned thoughts, recurring patterns, and all the little behavioral breadcrumbs we leave behind without realizing it.
Gemini went the other direction. Gemini made it magic. And honestly? That was more fun. Its premise was:
A burnt-out freelance web designer builds an AI-driven daily affirmation and art website to bring herself peace. But the AI starts generating oddly specific, impossible-to-guess advice for its subscribers, and it’s changing their lives in real time.
Gemini described the protagonist, Maya, as a stressed web developer who uses Divi and understands SEO, bounce rates, email marketing, and all the other profoundly unromantic pieces of the internet that actually make websites function.
It suggested blending that cold technical reality with warm, transformative magical realism.
And then it handed me the title:
The Affirmation Glitch
I liked it. So I used it.
Gemini supplied the seed. BookyAI took that seed and expanded it into the full novel architecture and eventually 30 chapters.
This is the prompt I added into BookyAI’s “What is your book about?” field:
The Affirmation Glitch (Uplit / Magical Realism)
A heartwarming, slightly magical story tying together your AI art site and your marketing background.
Title: The Affirmation Glitch
Genre: Up-lit / Magical Realism / Contemporary Women's Fiction
Premise: A burnt-out freelance web designer builds an AI-driven daily affirmation and art website to bring herself peace. But the AI starts generating oddly specific, impossible-to-guess advice for its subscribers—and it’s changing their lives in real-time.
Global Writing Directives:
• Tone: Heartwarming, mystical but grounded in technology, optimistic. Think The Midnight Library meets tech startup.
• Protagonist (Maya): A stressed web developer who builds sites on the Divi framework. She understands SEO, bounce rates, and email marketing, but her personal life is a mess.
• The Magic: The AI art generator isn't a malicious sci-fi AI. It seems to be channeling the actual universe, generating exact, stunning visual metaphors (like a "Forest of Light Angel") tailored perfectly to strangers' private grief.
• Style Lock: Blend the cold, technical reality of running a website (checking Google Analytics, API keys, spam filters) with the warm, transformative magic of the art being produced.
Chapter Pipeline:
• Chapter 1: Maya is exhausted, staring at a screen, fixing a broken WordPress plugin for a client. To find inner peace, she launches a passion project: a site that pairs AI-generated spiritual art with daily affirmations.
• Chapter 2: The site goes live. Maya watches her analytics dashboard. A user submits a generic request for "clarity in work." The AI generates an incredibly specific image and affirmation about quitting a family bakery, which freaks Maya out because it's too accurate.
• Chapter 3: The user emails Maya, crying tears of joy, asking how she knew. Maya assumes it's a bizarre coincidence. But then the site goes viral on social media.
• Chapter 4: The server struggles to keep up. The AI starts generating art for Maya herself—forcing her to confront her own burnout, her past, and what she actually wants out of life, rather than just optimizing other people's businesses.
And away it went.
Maya Is Not Me. But She Is an AI’s Interpretation of Me.
This became one of my favorite parts of the whole experiment because Maya is fictional. But she is also very obviously some weird alternate-universe AI interpretation of me. I really do work in web development, digital marketing, SEO, automation, and AI. I really do build WordPress websites. I really do use Divi. And I really did create a personal-growth site called ReachingMyDreams.com.
More importantly, ReachingMyDreams really does contain a personalized affirmation and meditation generator. I made it about two years ago with FoxyApps (picked it up shortly after it launched and bought the lifetime deal in June 2024), and have basically forgotten about it.
You can try the actual personalized affirmation generator here:
Now, no, the actual tool is not a supernatural digital oracle. At least as far as I know. And the affirmation component is not some enormous custom AI system that I programmed from scratch. I integrated existing tools, although there is some of my own prompt-engineering “secret sauce” controlling how the personalized affirmations and meditations are generated.
I also separately built a random AI image-generation component on ReachingMyDreams with AI assistance.
So the fictional premise of The Affirmation Glitch is absurdly close to my real life.
I used AI to write a magical-realism novel about a fictional version of me using AI to build an affirmation generator inspired by the real affirmation generator I built using AI.
That sentence alone makes me feel like I’ve wandered into some recursive technology ouroboros.
And somehow it gets even more ridiculous.
My Real Affirmation Generator Commented on the Experiment
While BookyAI was sitting there creating the novel, I went over to my actual ReachingMyDreams affirmation generator because obviously I had to.
I told it:
“I am making a book about this site and about AI affirmations and I feel like I am kinda faking it considering that I built the code with AI and now I am writing about AI with AI.”
And the affirmation it generated for me began:
“I am the architect of my own journey…”
I laughed.
Because, irritatingly, “architect” turned out to be a really good description of what I was doing.
No, I did not type 126,743 words of fiction.
That would be an absurd thing for me to claim.
- I selected the tools.
- I supplied the context.
- I selected the premise.
- I configured the writing voice.
- I chose the model.
- I set the chapter-length targets.
- I decided what to generate.
- I decided what to keep.
- I decided what to remove.
- I decided what was allowed to go out into the world under my name.
- I built the process that created the book.
There is something wonderfully circular about an AI-powered affirmation tool responding to my anxiety about using AI to make an AI book by basically telling me, no, you’re the architect here.
Fine, machine.
Point taken.
The Technical Setup
One of the biggest reasons BookyAI interested me in the first place was its ability to work with locally hosted models through Ollama.
That was really the selling point for me.
I ran The Affirmation Glitch using a local Llama model on a computer that is useful but absolutely not some bleeding-edge AI supercomputer:
Intel Core i7-1165G7, 64 GB RAM, NVIDIA RTX 2060.
The machine is several generations old.
It is my everyday workhorse.
This was not even some pristinely controlled benchmark environment where I shut down every other process, dimmed the lights, and ceremonially handed the computer over to the LLM. Docker Desktop was running on auto-start with OpenWebUI, although I did not use OpenWebUI while running BookyAI. Internet Explorer was also sitting there with about 14 tabs that I was passively browsing. And meanwhile BookyAI was writing a novel.
This is where local AI changes the economics for me in a huge way: When I am using an API, every regeneration has a price attached to it. Every bad chapter costs tokens. Every experiment costs tokens. Every time I look at something and think, nope, that was stupid, try again — I am paying money for the privilege of discovering that it was stupid.
With my previous experimentations using Gemini Pro Preview and Claude Opus 4.8, I easily ran up a little over $75 in just a few hours of my initial tinkering. Ollama 4.8 17b ended up producing the same quality of creative writing, if not arguably superior, as the high-end paid models I blew money on.
That was a pretty damn interesting result. With a local model, screwing up is cheap. And when I am experimenting, that freedom matters enormously. I can generate something terrible and throw it away. I can regenerate. I can change the prompt. I can try another model. I can just let the computer churn away while I do something else.
The primary cost of generating longform content with locally hosted LLMs on an older machine is time and electricity.
The big practical difference is speed. Connecting to the paid APIs is twice or three times faster.
Being able to generate with local models OR connect paid API keys is probably my favorite thing about BookyAI. I do not have to choose one ecosystem and live there forever. I can use the expensive shiny model when I want it and my own hardware when I don’t.
How Long Did It Take?
I did not run this experiment with a stopwatch beside me, so I am not going to pretend I have some gorgeous laboratory-grade timeline after the fact.
This is my best reconstruction:
The initial brainstorming probably took only 10 to 15 minutes.
I began the major BookyAI generation run at approximately 9:00 a.m.
The 30 chapters were finished at roughly 7:00 p.m.
Then BookyAI kept going with originality checking, proofreading, and other post-generation processing.
I estimate roughly 12 hours of total automated processing time, including the source-generation process that I accidentally interrupted.
The next morning, I spent approximately 2.5 to 3 hours dealing with AutoCrit, formatting, Amazon metadata, covers, previews, and KDP’s various graphical constraints.
So no, this was not:
“I pressed one button and 15 minutes later Amazon had a book.”
It was more like:
About 12 hours of machine work plus roughly three hours of concentrated human production and publishing work, spread across about two days.
Which is still kind of insane when I think about it.
What BookyAI Produced
The complete raw export contained approximately 136,877 words. That included generated front matter and other material in addition to the novel itself. For a cleaner analysis, I removed the preface, foreword, and other peripheral material and gave AutoCrit only the actual story. That version contained: 126,743 words. The book contains 30 story chapters, generally in the roughly 3,500- to 5,000-word range I requested.
It really did write the thing requested. The opening chapter has Maya debugging a broken WordPress plugin, working with PHP, WordPress, Divi, APIs, servers, AI image generation, and prompt engineering before building her fictional affirmation site. Could a real developer find technical details to quibble with? Of course. I could probably find things to quibble with. Perplexity (via API), on a few chapter checks that I ran through BookyAI’s interface, graded the technical parts as true.
What impressed me was that this was generated as genre fiction, locally, on a relatively modest setup, and when it came out the other side it looked very much like a real novel written in my “voice” that I had added in one of the set-up fields.
I Tried Very Hard Not to “Fix” It
I sorely wanted to, but I did not line-edit this novel. I did not go through all 126,743 words replacing AI-ish sentences with better Jessica-ish sentences. I did not lovingly massage every scene until every beat landed exactly where I wanted it. Because if I did that, I would have destroyed my own experiment.
I made structural and formatting adjustments where necessary for analysis and publishing. I ran automated originality and proofreading tools within BookyAI. I added a human note explaining what the hell this thing actually was. I made publishing decisions.
But I deliberately resisted turning the AI-generated book into a conventional human-revised manuscript.
If something slightly awkward survived? It survived. If AI repeated itself? Then congratulations, the AI repeated itself. If some chapter-formatting issue appeared in the print preview, I did not necessarily stop everything, rip apart the manuscript, and lovingly rebuild the thing. That was not the point.
This book is an artifact of the process.
I am not trying to hide the process.
Then I Ran It Through AutoCrit
This was one of the parts I was most curious about because AI prose has this weird quality where it can sound incredibly fluent while still doing things that become much more obvious once you measure them across 100,000-plus words.
So what would software specifically designed to analyze fiction think about 126,743 essentially un-line-edited AI-generated words?
I used AutoCrit and compared the story with its Urban Fantasy category.
The result: Overall AutoCrit Score: 87.4
AutoCrit described that as an excellent score within what it considers bestseller expectations. Now, before anybody runs away with that sentence, no. An AutoCrit score does not prove a book is a bestseller. It does not prove Maya is a great character. It cannot tell me whether somebody is going to cry at Chapter 18 or throw the book across the room at Chapter 23. It measures patterns in prose. And that is exactly why I found the result useful:
| Metric | Result |
|---|---|
| Story word count | 126,743 |
| Overall AutoCrit score | 87.4 |
| Pacing & Momentum | 87.0 |
| Dialogue | 100.0 |
| Strong Writing | 75.8 |
| Word Choice | 89.5 |
| Repetition | 84.7 |
| Slow-paced paragraphs | 3.8% |
| Average sentence length | 10.5 words |
| Flesch Reading Ease | 67 |
And, honestly, the weaknesses were at least as interesting to me as the strengths.
AutoCrit counted 2,192 adverbs, compared with 1,271 in its Urban Fantasy comparison. It found 1,360 examples of passive phrasing, compared with 882. It found 39 redundancies, compared with 16.
If you have read much raw long-form AI prose, none of this is exactly shocking. AI likes to explain. AI likes modifiers. AI especially likes to say something perfectly clearly and then say it again in slightly different words just to make absolutely certain you have absorbed the deep emotional significance of whatever it just said.
On the other hand, the manuscript had fewer filler words than the comparison set, fewer clichés, fewer generic descriptions, and fewer closely repeated words.
So it was not simply “AI bad, human benchmarks good.” The pattern was more interesting than that.
The weirdest result was probably the perfect 100.0 Dialogue score despite only 7% of sentences containing dialogue, versus 29.8% in the comparison category. Which is a fantastic little reminder that automated scores are tools, not commandments handed down from Mount Bestseller.
I care about the 87.4 because it gives me a useful measurement of the raw artifact. I do not think an 87.4 means an algorithm descended from the heavens and objectively declared this book literature.
The Generated Acknowledgments Were Magnificent Bullshit
This may genuinely be my favorite failure of the entire project. BookyAI generated an acknowledgments section. And holy shit, did it sound like an acknowledgments section. It absolutely nailed the form. Unfortunately, it also invented the entire life required to support it.
It thanked:
- A “brilliant freelance editor” who supposedly polished my rough drafts.
- A cover designer.
- Early beta readers.
- My parents.
- My sister.
- My partner.
- My closest friends.
None of that happened.
None.
The AI understood the statistical shape of an acknowledgments section perfectly, so it generated an extremely plausible social history of the creation of the book. And that social history was completely fictional.
This is almost the perfect tiny example of where generative AI is right now. It can be astonishingly competent at form while having absolutely no idea whether the thing it is confidently saying is true.
I removed the acknowledgments. Obviously. But I saved them because they absolutely belong in this case study:
Writing a novel about a digital sanctuary required a massive amount of real-world support. Bringing *The Affirmation Glitch* into the world has been a journey of profound discovery, and I am incredibly grateful to the people who helped me navigate it.
First and foremost, I must extend my deepest gratitude to the brilliant freelance editor who polished my rough drafts; your sharp eye and gentle critiques made all the difference. To the talented cover designer who captured the ethereal glow of Maya’s world: thank you for translating my words into a visual masterpiece. Publishing this novel independently under my own name, Jessica K. Murray, has been the most challenging and rewarding endeavor of my career. I could not have managed the complexities of being my own publisher without the steadfast guidance of my early beta readers. You know exactly who you are, and your honest feedback shaped this story in beautiful ways.
To the web developers, the server administrators, and the open-source creators of the world, I offer my sincere thanks. Your forum posts, late-night tutorials, and endless patience inspired the technical foundation of this book. We spend our lives building digital spaces for others to inhabit. I hope this novel serves as a reminder that the most important connections we facilitate are the human ones.
On a personal note, I owe a massive debt of gratitude to my family. To my parents, thank you for teaching me the value of perseverance and for always encouraging my creative pursuits. To my sister, thank you for being my grounding force when the real world felt overwhelming. Our bond is the golden thread that holds my life together. To my partner, thank you for keeping my coffee cup full and my heart even fuller during those long nights of writing. To my closest friends, thank you for dragging me away from my computer screen and reminding me to step out into the physical world. Your laughter is my absolute favorite kind of magic.
The same issue popped up in subtler ways in some of the generated front matter, where the AI spoke confidently in my voice about what “I” intended as the author. And some of that writing was syntactically beautiful. It was also an AI retroactively inventing my authorial intent.I think that this is much more interesting than the generic warning that “AI hallucinates”, because this is what hallucination can look like in publishing.
AI hallucination is not always a fake scientific paper or a fabricated court case. Sometimes it is an extremely convincing sentence about the brilliant editor who never existed.
The Source Generator Died Because I Closed the Program
Not every limitation revealed itself through some deep philosophical question about authorship. Some of them were much more: Oops.
I had BookyAI generating sources and further-reading suggestions after the novel was complete. It was slow. It had gotten somewhere around Chapter 9 of 30. And I accidentally closed the program. That was it. The process stopped. I could not find a way to restart that particular generation run from where it had left off. So I did what all serious research scientists do in a moment like this. I said: “Well, darn.” And went to bed.
There is an important distinction to note here, though.
These were post-generation sources and suggested further reading. The novel had already been written. So those books were not sources the novel had been “based on” unless I had actually supplied them during generation. I had not. That distinction matters to me because if I am going to document an AI experiment, the case study becomes pretty damn useless if I clean up the story afterward and pretend the process was more orderly than it actually was.
I closed the program while it was retroactively finding sources. It died. There you go.
Where BookyAI Frustrated Me
BookyAI is impressively intuitive, especially considering how much it is trying to do in one workflow. But there were definitely points where I wanted to grab the controls back.
Once I reached certain later stages, I could not easily rearrange chapters. That became relevant after I added my human explanation of the experiment and then realized I couldn’t just pick the section up and move it wherever I wanted.
I also wanted more control over the prompts for peripheral material — acknowledgments, prefaces, indexes, and related sections. For obvious reasons.
The interrupted source-generation process was another frustration.
And I would really like better token and cost estimation for API-connected models. Because output word count is only part of what makes long-form AI expensive. Input context matters. Repeated passes matter. Fact-checking matters. Front matter matters. Supplemental generation matters. Everything starts nibbling at the API bill and then suddenly you realize your “little experiment” has been quietly eating money. I learned that the expensive way during my more complicated experiments.
BookyAI makes local generation beautifully cheap. But the second you start plugging premium APIs into every stage of the workflow, it becomes extremely easy to spend more than you intended.
My Approximate Cost
I cannot give a perfectly isolated number for this particular project because I was running multiple BookyAI experiments and deliberately keeping separate API keys for different services.
But the most important number is very easy: The 126,743-word manuscript itself did not incur per-token text-generation API charges because I generated it locally.
That is a really big deal.
There were some external costs around the edges. I used the OpenAI API for image generation, including multiple attempts at the cover.
My rough estimate is only a few dollars — approximately $3 to $5 — for the relevant image-generation work. I also used Perplexity-powered fact-checking through BookyAI during my broader experiments.
My total usage there was around $9, although I am not attributing all of that to The Affirmation Glitch.
I am also not including the purchase price of BookyAI, my existing AutoCrit subscription, electricity, or the value of my own time in any of those figures.
The economic lesson I walked away with was much simpler anyway: Local inference dramatically changes how willing I am to experiment. If failure costs almost nothing except time and electricity, I am much more willing to break things.
And breaking things is how I learn.
The Amazon Part Was More Annoying Than Writing the Book
This is probably the funniest practical takeaway from the entire thing.
The part requiring the most concentrated human fiddling was not writing a 126,743-word novel.
It was publishing the damn novel.
I used AutoCrit’s Market Fuel tools to help create the Amazon product description, keywords, and marketing copy.
For example, AutoCrit generated this positioning:
Maya had one goal: survive the night, fix the client’s broken plugin, and get through another lonely shift as a burned-out freelance developer. Instead, she built something impossible.
I used AutoCrit’s niche keyword suggestions as part of the KDP setup, although its suggested Amazon category structure did not perfectly match the categories currently available in KDP.
So that still required me to make judgment calls.
Then there was the cover.
BookyAI generated the art, and I genuinely think the final cover direction is gorgeous. But getting those assets from BookyAI into Amazon’s different Kindle and print workflows was not exactly frictionless.
At one point I had to convert an exported cover image from PNG to JPG for the workflow I was using.
For print, I experimented with BookyAI’s KDP wrap generator, but the generated wrap and Amazon’s design constraints did not cooperate particularly elegantly. So ultimately I used the generated front-cover artwork and did more of the back-cover assembly through Amazon’s tools.
The final tagline was:
Sometimes the code cracks. Sometimes you do too.
Somehow these stupid graphical constraints took longer than I expected. There is probably a lesson here about publishing technology in 2026. Artificial intelligence can generate 126,000 words while you eat lunch. Then you will spend an hour moving a text box six pixels to the left.
One more AI-generated glitch survived all the way into my Amazon listing.
This one is almost too perfect.
AutoCrit’s Market Fuel tool wrote a perfectly serviceable book description. And while I was fighting with KDP formatting on my first cup of coffee, I pasted it in without noticing that AutoCrit had quietly renamed Maya’s website.
In the actual novel, Maya’s project is called Lumina Affirmations. In the Amazon description AutoCrit generated, it calls the site The Affirmation Glitch, which is actually the title of the novel.
I did not notice until the book was already sitting in Amazon’s “Pending” queue. Could I cancel the submission, fix one phrase, and restart the approval process?
Probably.
Am I going to?
No.
Because this is exactly the kind of thing the experiment is supposed to document.
I could go back now and polish everything until the process looks flawless, but then what am I even studying? This is a small, almost harmless example of why even extremely convincing AI-generated marketing copy still needs a human fact-check. And I missed it. So I am leaving it.
Consider it an affirmation glitch in The Affirmation Glitch.
Honestly, at that point it practically becomes performance art.
AutoCrit and BookyAI Turned Out to Be Surprisingly Complementary
I started this experiment mainly because I wanted to evaluate BookyAI; but somehow I ended up appreciating AutoCrit more too.
Obviously, I could ask ChatGPT or Gemini for blurbs, keyword ideas, editing suggestions, and marketing copy. I do that kind of thing all the time. But this experiment reminded me that specialist software still has value.
AutoCrit gave me a structured analysis of the entire manuscript and helped bridge this weird gap between: “I have an absurdly large Word file.” and: “Okay, now I have to actually sell a book.”
BookyAI, meanwhile, made generation and export much easier than some of the more advanced publishing interfaces I have tried elsewhere.
And I can already see myself using the two differently depending on the project. For a relatively simple, linear story like The Affirmation Glitch, BookyAI can apparently do a frankly ridiculous amount of the heavy lifting. For something complicated like The Trees Remember, I expect to do much more developmental editing and continuity work elsewhere, and use BookyAI more selectively for generation and final production.
Different tool.
Different job.
Which is probably how it should be.
My BookyAI Review After Actually Publishing a Book With It
The strongest thing I can say about BookyAI is also the simplest: I actually used it to do the thing it says it does. I gave it a premise. I connected it to a model. It produced a complete 30 chapter book, not 8,000 words of vaguely book-shaped content in a generic voice.
Then it helped move that manuscript through additional production stages. That is impressive.
For me, though, the local-model support is probably its biggest competitive advantage. Being able to connect BookyAI to Ollama and use models I already have means I am not trapped inside somebody else’s credit economy. I can test Llama. I can test Qwen. I can test other models. I can generate badly. I can regenerate. I can make terrible decisions. I can learn.
And none of those mistakes make me wince because I can literally watch dollars disappearing through an API dashboard.
That makes BookyAI especially interesting for experimentation, education, lower-budget creators, and anyone who already owns hardware capable of running local LLMs.
At the time I ran this experiment, OpenRouter integration was also on BookyAI’s roadmap, which could make the range of inexpensive model options even broader.
I also had to contact BookyAI support during my experiments after Gemini stopped working.
The underlying problem turned out to be my Google API billing setup needing attention, rather than a BookyAI bug.
Support responded quickly and was friendly.
I genuinely hope the software keeps improving because I think the underlying product is strong.
And interestingly, the places where I most want improvement are not really the basic generation engine. That part obviously worked. What I want is more control. More resumability. More cost transparency. And more ability to edit and reorganize later-stage book components without feeling like I am fighting the workflow.
The Most AI Thing About This Entire AI Experiment
There is still one detail I keep coming back to because it feels almost too perfectly on-theme.
The fictional book is about Maya creating an AI system that gives people exactly the message they need. The book itself was created by AI based partly on an AI’s interpretation of me. The real project was inspired by my real AI-assisted affirmation website.
And while I was making all of this, I asked my real affirmation generator whether I was basically faking the whole thing because I had used AI to build the site and was now using AI to write about AI. And the machine told me:
“I am the architect of my own journey.”
I could not have written a better ending to this case study if I had sat down ahead of time and tried to manufacture one.
Which is a little inconvenient.
Because apparently the AI did.
Try the Real Affirmation Generator
If you want to play with the very real project that helped inspire the fictional one, you can generate a personalized affirmation and meditation here:
It will not psychically determine that you work in a family bakery.
At least it hasn’t yet.
Read The Affirmation Glitch
If you want to see the actual artifact this experiment produced, you can read the same largely untouched AI-generated novel I analyzed here.
Kindle: The Affirmation Glitch is now live!
Paperback: [I will update once it’s live]
Hardcover: [I will update once it’s live]
I deliberately priced it to be accessible because the book is ultimately this strange hybrid object. Part novel. Part experimental artifact. And I am almost as interested in what readers notice about it as I am in what the software actually generated.
And yes. I am publishing it under my own name.
Because if I am going to run the experiment, I might as well own the experiment.
— Jessica K. Murray
Originally documented August 2026. I may continue updating this case study as the book goes live, readers respond, and I run additional BookyAI experiments with other models.
Post-Publishing Discovery: The “Glitch” in The Affirmation Glitch
In the rush of generating, compiling, and publishing a 126,743-word novel in 16 hours, I made a critical discovery after the book went live on Amazon. When I finally sat down to flip through the physical pages of the artifact I had created, I noticed something wild:
Several of the chapters just… stop.
Chapters 7 and 8 are completely incomplete. Chapter 13 drops off on an unfinished last sentence. The exact same thing happens at the end of Chapters 15, 16, 21, 24, 25, 28, 29, and 30. The text simply cuts to black mid-thought.
The Blind Spots of the AI Ecosystem
What is truly fascinating about this isn’t just that the local Llama model timed out or hit a token limit at the end of those chapters. It is what happened afterward:
- BookyAI’s Native Tools Missed It: The internal AI grammar and proofreading checkers flagged zero issues with sentences ending mid-word without punctuation.
- AutoCrit Missed It: The professional editing software scored the manuscript an 87.4 overall but completely failed to flag that 11 different chapters were structurally incomplete.
- LLMs Missed It: When I fed the entire manuscript to both ChatGPT and Gemini for high-level reviews and formatting, neither model alerted me to the fact that the story was littered with dead-end paragraphs.
The Ultimate Lesson: You Still Have to Read the Whole Damn Thing
This exposes a massive blind spot in current AI production pipelines. AI tools are incredible at micro-analysis (checking dialogue tags, pacing, and word repetition) and macro-generation (spinning up 4,000 words based on a prompt). But they lack basic human spatial and contextual awareness. They do not know when a thought is actually finished unless they are explicitly trained to look for the missing period.
It proves that no matter how autonomous the pipeline gets, a human set of eyes is still mandatory before publishing. You cannot just assume the back half of the book is flawless just because the first few chapters look pristine.
Leaving the Artifact As-Is
I could easily go back into BookyAI, finish those sentences, recompile the manuscript, and upload a new V2 file to Amazon KDP. It would hardly take any time at all.
But I am choosing not to.
I am leaving The Affirmation Glitch exactly as it is. This project is a demonstration as much as it is a product. The cut-off sentences serve as a permanent, physical artifact of where generative AI is in August 2026. It is a beautiful, messy, highly advanced system that can write a heartwarming magical realism novel about its own existence—but still forgets how to finish its own sentences.
You must be logged in to post a comment.