How to Make AI Write a Bestseller—and Why You Shouldn't

As a great man once said, "Drive stick, motherfucker."

This is not endorsement. The techniques I will discuss are being shared in the interest of research and defense, not because I advocate using them. I don’t.

This is not a get-rich-quick guide. You probably won’t. Publishing is stochastic. If ten people try this, one of them will make a few million dollars; the other nine will waste thousands of hours. This buys you a ticket, but there are other people’s balls in that lottery jar, and manipulating other people’s balls is beyond the scope of this analysis.

It’s (probably) not in your interest to do what I’m describing here. This is not an efficient grift. If your goal is to make easy money, you won’t find any. If your goal is to humiliate trade publishing, Sokal-style, by getting an AI slop novel into the system with fawning coverage, you are very likely to succeed, but it will take years, and, statistically speaking, you’re unlikely to be the first one.

Why AI Is Bad at Writing (and Will Probably Never Improve)

A friend of mine once had to take a job producing 200-word listicles for a content mill. Her quota was ninety per week. Most went nowhere; a few went viral. For human writers, that game is over. No one can tell the difference between human and AI writing when the bar is low. AI has learned grammar. It has learned how to be agreeable. It understands what technology companies call engagement; at this, it outplays us.

So, why is it so bad at book-length writing, especially fiction?

1. Poor style. Early GPT was cold and professional. Current GPT is sycophantic. Claude tries to be warm, but keeps its distance. DeepSeek uses rapid-fire register switches and is often funny, but I suspect it’s recycling jokes. All these styles wear thin after a few hundred words. Good writing, especially at book length, needs to adjust itself stylistically as the story evolves. It’s hard to get fine-grained control of the writing if you do not actually… write it.

2. No surprise. The basic training objective of a language model is least surprise. Grammar errors are rare because the least surprising way to say something, often, is grammatical. Correct syntax, however, isn’t enough. Good writing must be surprising. It needs to mix shit up. Otherwise, readers get bored.

3. No coherence. AI can describe emotion, but it has no interior sense of it. It can generate conflicts, but it doesn’t understand them well enough to know when to end or prolong them. Good stories evolve from beginning to end, but they don’t drift; there’s a difference. The core of the story—what the story really is—must hold constant. Foreshadowing, for example, shows conscious evolution, not lazy drift. AI writing, on the other hand, drifts aimlessly.

4. Silent failure. This is why you’ll find AI infuriating if you try to write a book with it. Ordinary programs, when they fail, crash. We want that; we want to know. Language models, however, do not tell you when they malfunction. In AI, there are fractal boundaries between green and red zones. Single-word changes to prompts—or model updates, out of your control—can break entire systems.

This is unlikely to change. In ten years, we might see parity with elite human competence at the level of 500-word listicles, as opposed to 250 today, but no elite human wants to be writing 500-word listicles in the first place. For literary writing, AI’s limitations are severe and probably intractable. At the lower standard of commercial writing? Yes, it’s probably possible to AI-generate a bestseller. That doesn’t mean you should. But I’ll tell you how to do it.

Technique #0: Prompting

Prompting is just writing—for an annoying reader. Do you want emojis in your book? No? Then you better put that in your prompt. “Omit emojis.” Do you want five percent of the text to be bold? Of course not. You’ll need to put that in your prompt as well. I was using em-dashes long before they were (un)cool, and I’m-a keep using them, but if you’re worried about the AI stigma… “No em-dashes.” You don’t want web searches, trust me, not only because of the plagiarism risk, but because retrieval-augmented generation seems to inflict a debuff of about 40 IQ points—it will forget whatever register it was using and go to cold summary. “No web searches.” Notice that your prompt is getting longer? If you’re writing fiction, bulleted and numbered lists are unacceptable. So include that, too. Prompting nickel-and-dimes you. Oh, and you have to keep reminding it periodically of every instruction, because it will forget and revert to its old, listicle-friendly style. You’ll blame the AI for being too dumb to understand your prompts. See? You’re already an author.

Technique #1: Salami Gluing

Salami slicing is the academic practice of publishing a discovery not in one place but in twenty papers that all cite each other. It’s bad for science because it leads to knowledge fragmentation, but it’s great for career-defining metrics (e.g., h-index) and for that reason it will never go away—academia’s DDoS-ing itself to death, but that’s another topic.

I suspect that cutting meat into tiny slices isn’t fun. Gluing bits of it back together might be… more fun? Probably not. Anyway, to reach the quality level of a publishable book, you’ll need to treat LLM output as suspect at 250 words; beyond 500, it’ll be downright bad. If there’s drift, it will feel “off.” If there isn’t, it will be repetitious. The text will either be non-surprising, and therefore boring, or surprising but often inept. On occasion, it will get everything right, but you’ll have to check the work. Does this sound fun to you? If so, I have good news for you. There are places called “jobs” where you can go do boring shit and not have to wait for years to get paid. I suggest looking into that. You can then skip the rest of this.

Technique #2: Tiered Expansion

Do not ask an AI to generate a 100,000-word novel, or even a 3,000-word chapter. We’ve been over this. You will get junk. There will be sentences and paragraphs, but no story structure. What you have to do, if you want to use AI to generate a story, is start small and expand. This is the snowflake method for people who like suffering.

Remember, coherence starts to fall apart at ~250 words. The AI won’t give you the word count you ask for, so ask for 200 each time. Step one: Generate a 200-word story synopsis of the kind you’d send to a literary agent, in case you believe querying still works. (And if you believe querying works, I have a whole suite of passive-income courses that will teach you how to make $195/hour at home while masturbating.) You’ve got your synopsis? Good. Check to make sure it’s not ridiculous. Step two: Give the AI the first sentence of the synopsis, and ask it to expand that to 200 words. Step three: Have it expand the first quarter of that 200-word product into 200 words—a 4:1 expansion. Do the same for the other three quarters. You now have 800 words—your first scene. Step four: Do the same thing, 99 more times. There’s a catch, of course. To keep the story coherent, you’ll need to include context in your prompts as you generate new work. AI can handle 5000+ word prompts—it’s output, not input, where we see failure at scale—but there will be a lot of copying and pasting. A lot of copying and pasting. A lot of copying and pasting. Learn those hot keys.

Technique #3: Style Transfer

You’re going to need to understand register, tone, mood, and style. There’s probably no shortcut for this. Unless you can judge an AI’s output, how do you know what to use and what to toss? You still have to learn craft; you just won’t have to practice it.

It’s not that it’s hard to get an LLM to change registers or alter its tone; in fact, it’s easily capable of any style you’ll need in order to write a bestseller—we’re not talking about experimental work. The issue is that it will often overdo the style you ask for. Ask it to make a passage more colloquial, and the product will be downright sloppy—not the informal but mostly correct language fiction uses.

Style transfer is the solution. Don’t tell it how to write. Show it. Give it a few thousand words as a sample, and ask it to rewrite your text in the same style. Will this turn you into Cormac McCarthy? No. It’s not precise enough for that. It will not enable you to write memorable literature. But a bestseller? Easy done, Ilana.

Technique #4: Sentiment Curves

Fifty Shades of Grey is not an excellent novel, but it sold more copies than Farisa’s Crossing will. Why? There’s no mystery about this. Jodie Archer and Matthew Jockers cracked it in The Bestseller Code.

Most stories have simple mood, tone, and sentiment curves. Tragedy is “line goes down.” Hero’s journeys go down, then up, in mood. There are also up-then-down arcs—rags to riches to ruin. There are curves with two or three inversions. Forty or fifty is… not common. But that’s how Fifty Shades is structured, and that’s why it best-sold.

Fifty Shades isn’t about BDSM. It’s about an abusive relationship. Christian Grey uses hot-and-cold manipulation tactics on the female lead. In real life, this is a bad thing to do. In writing? Debatable. It worked. I don’t think James intended to manipulate anyone. On the contrary, it makes sense, given the characters and who they were, that a high-frequency sentiment curve would emerge.

Whipsaw writing feels manipulative. It also eradicates theme, muddles plots, and damages characters. Most authors can’t stand to do it. You know who doesn’t mind it, though? Computers.

This isn’t limited to AI. If you want to best-sell, don’t write the book you want to read. Instead, write a manipulative page-turner where the sentiment curve has three inversions per page. It’s hard to get this to happen if your characters are decent people who treat each other well. On the other hand, the whole story becomes unstable if you have too many vicious people. The optimal setup is… one ingenue and one reprobate. I bet this has never been done before. Of course, the reprobate must behave villainously, but you can’t make him the villain, so you must give him redeeming qualities such as… a bad childhood, a billion dollars, a visible rectus abdominis. One of these forgives all sins; all three make a hero. If you’re truly ambitious, you can add other characters, like: (a) an actual villain of ambiguous but certain ethnicness, (b) a successful sister or female friend whom the ingenue resents for no reason, or (c) a werewolf. This, however, is advanced literary technique. You don’t need it.

If you’re looking to generate a bestseller, the sentiment curve is the one element to which you cannot trust a large language model. You have to do it by hand. I recommend drawing a squiggly line (the more inversions, the better) on graph paper, taking a picture, uploading the image to the cloud, and using a multimodal AI to convert it into a NumPy array. You’re done.

Technique #5: Overwriting

Overwriting can be powerful. It’s when you take a technical aspect of writing to its maximum, showing fluency where lesser writers would become incoherent. Hundred-word sentences—sometimes brilliant, sometimes mistakes, sometimes brilliant mistakes—are an example of this.

From Paul Clifford, “It was a dark and stormy night” is an infamously bad opening sentence, but it isn’t that bad, not in this clipped form. It’s simple; the reader moves on. The problem with the sentence, as it was originally written, is that it goes on for another fifty words about the weather. Today, this is considered pretentious, boring, and even obnoxious. Back then, it was considered good writing.

Overwriting that breaks immersion by drawing attention to itself is ruinous. Skilled overwriting, when it serves the story’s needs, shows craft at the highest level.

The good news is that you’re writing a bestseller. You don’t need to worry about this. Craft at high levels? Why? You don’t need it. In fact, you didn’t need this section at all.

Technique #6: Escalation Via Naive Bayes Attacks

Overwriting’s a style risk bestsellers don’t need to take, but they do need to take content risks to drive gossip and buzz. How do you get an AI to write explicit sex or violence? It’s not easy. We all complain about how reluctant chatbots are to serve up graphic axe murders when asked for cookie recipes, but what can you do?

A Naive Bayes attack is a way to make a language model malfunction, or behave strangely, by feeding it weak evidence slowly. You can’t get socially unacceptable behaviors, even in simulations or stories, if you deliver the payload all at once. You have to escalate in a series of prompts. Give the LLM one big vicious prompt, and it will fight you. Give it a series of small drips, and you can guide it to a dark place.

This applies to more than just language models, if you—uh, let’s just move on.

Technique #7: Recursive Prompting

Recursive prompting is the Swiss army machine gun mixed metaphor salami blender of LLM techniques, as it subsumes and expands upon everything we’ve discussed so far. The idea is simple: use one LLM’s output as input to another one. Why talk to an LLM when you can have another LLM do the talking? Why manage LLMs when you can have an LLM do the managing?

I was once faced with a trolling task where I needed a 670-word shitpost to be embedded inside another shitpost, and I wanted AI slop but I could afford no drift. Worse, I needed it to pull information from 30,000+ words of creative work. Claude has a big enough context window, but is too measured in style for good shitposting. On the other hand, DeepSeek handles the shitpost register as well as a professional human troll, but not large context windows. The solution I used was style transfer: I included 2,000 words of DeepSeek output in my Claude prompt. Also, I didn’t write the style transfer prompt myself; I had ChatGPT do it.

In other words, I used the strengths of several models to produce a shitpost that, while not at the level of a top-tier human shitposter, is better shitposting than any single model can achieve today. A new state of a new art. I’ll put that on my next vanity plate, but they’ll make me take some middling letters out. “A new st[…]art?” We’re getting there.

Technique #8: Pipelining

You will exhaust yourself with the work described above. Recursive prompts to generate recursive prompts to run Naive Bayes attacks on large language models just to make your villain steal a child’s teddy bear and kick it into the sun… it’s a grind.

You’ll want API access, not chatbot interfaces. You’ll have to start writing some code. Some recursive-prompt tricks can be done with five queries; some take fifty or five hundred. You’ll need to start out doing everything manually, to know what your “creative” process is going to look like, but you’ll find ways to automate the drudgery. Setting? “Give me 300 words describing the setting of a bestselling novel.” That does it. Plot? Again, your sentiment curve just needs to be squiggly. Characters? Covered. Style? Covered. Theme? You’re writing a bestseller. Optional.

You’ll end up with five thousand lines of glue code to hold all your LLM-backed processes together. If an API breaks, you’ll have to spend a few hours debugging. But I have faith in you. Did you know that Python 3.7 has three different string types? Well, you do now. Look at you, you’re already going.

Technique #9: A Little Bit of Luck

This is surprising to people, but writing a mediocre novel doesn’t guarantee millionaire status. Even having a mediocre personality (i.e., not being a “difficult author”) doesn’t guarantee it, although it helps. In fact—and I don’t want to discourage you on your mediocrity journey, but you should know this—there are people out there who excel at mediocrity and have never received a single book deal. If you stop here with your AI slop novel, you’re going to be one of them.

The good news is that using AI to generate a query letter is a thousand times easier than using it to generate a book that readers won’t clock as AI slop. Compared to everything you’ve done, writing emails and pretending to have a pleasantly mediocre personality is going to be super easy… unless you’re truly gifted. Then you’re fucked.

No one wins lotteries if they don’t play—Shirley Jackson taught us that.

Technique #10: Ducks

Your query letter worked. You signed a top-tier agent and you have a seven-figure book deal, and now you’ve got a ten-page editorial letter full of structural changes to an AI slop novel that you realize now you don’t even understand, because you didn’t write it. Well, shit. What are you going to do? You thought you were done! It turns out that, if you want the last third of your $7,900,000 advance, you have three hundred more hours of prompting to do.

There’s a trick. Ducks. In video games, a duck is a design fault deliberately added for that one boss who has to make his mark. Imagine a Tetris game with a duck that flaps its wings and quacks every time the player clears a line. In executive review, VP says, “Perfect, except the duck. Take that out and ship it.” You get told to do what you were going to do anyway. You win.

At book length, you’re going to need six or seven of these to give your editor something to do. Some ideas would be:

• Name your character Fifi. You’ll change it later. If you miss a few pages during your Ctrl+F journey, you just got a new character for free.

• Add an alien species that for no explained reason has one weakness—an irresistible drive to mate with pumpkins.

• Include a nose-picking scene from the perspective of the booger. Don’t tie it to the rest of the plot at all. It will stick to something.

Of course, the duck principle doesn’t always apply. Some of us remember Duck Hunt, a game in which the ducks and the quacking were thematically essential. But Duck Hunt is 19-dimensional Seifert manifold chess and we’re not ready to discuss it yet. We might never be.

Technique #11: You Can Now Write a Real Fucking Book

Congratulations. You’ve spent nine hundred and forty-seven hours to produce word-perfect AI slop. You’ve queried like a power bottom. You’ve landed your dream agent, your movie deal, your international book tour. Famous authors blurb your book as: “Amazing.” “Astonishing.” “I exploded in a cloud of cum.” The New York Times has congratulated you for having “truly descended the gradient of the human condition.”

It’s not all perfect, though. You suspect, every time someone else’s novel features a successful author and his failures, that it was written about you. Academics focus on that pumpkin scene you forgot to take out, so you must concoct a theme to hang it on. You have all the rich people problems, too; you spend an hour a week with a financial advisor who nags you not to golf with ortolans so much because those little birds are expensive—and, anyway, you’d be 20 strokes better if you just used golf balls like everyone else.

Still, you have a literary agent who returns your calls. People who don’t read closely name their kids after your characters. Best of all, you’re now one of the five people alive who has enough clout to get actual literature published. What are you gonna do with that fortunate position?

Two AI books at the same time.