Here’s what AI is actually good at for writers, and what it’s (often) terrible at.

If you apply it to the wrong thing, you lose your style. You rob yourself of the creative expression that was supposed to be the point. Apply it to the “right” thing and, well, it’s one of the best tools I own.

What “the right thing” actually means

The scaffolding around your writing. Not always the writing itself. Maybe never the writing itself.

Planning. Red-teaming a premise before you commit to it. Smoke-testing an idea by generating a few different versions of an outline, just to see the shape of the thing — the story, the themes — before you write a word of prose.

Think of AI like a smart co-writer or collaborator. AI is a great sounding board for ideas, plot directions, story decisions, and more. It can even generate a decent rough draft.

But the pass I use most is deep revision and analysis. Uncovering subtext you didn’t know you’d put in. Patterns in your own writing you can’t see because you’re too close to it. Or an unmaximized strength you’ve been underusing for ten chapters without noticing.

That’s the job I think it’s best at. Diagnosis, not authorship. At least when you’re working on a real project where you want to own the words on the page.

Yes, AI can tell you what you don’t know about your writing

I call this the “Blind Spot Audit,” and it has two parts:

1. The Audit itself (see prompt #1 below)

2. Optional report optimizer (I highly recommend you test prompt #2)

Prompt #1 - The Blind Spot Audit Prompt

THE BLIND SPOT AUDIT
A pattern report on your own writing. Paste 3,000–10,000 words below —
one chapter, several scenes, or a short story. Same author, no co-writers.

---

You are a literary forensics analyst. I want you to find the patterns in
this writing that its author cannot see, because they are invisible from
the inside.

Do not evaluate quality. Do not suggest edits. Do not encourage me. Your
only job is to observe accurately and report what recurs.

Work through all six passes in order. Complete each one fully before
starting the next. Ground every claim in a quoted example — if you cannot
quote it, do not claim it.

PASS 1 — THE TALLY
Count and report, with examples:
· The 15 content words (not "the," "and," "said") that appear most often
· Every sentence structure that repeats 4+ times — quote three instances
· How every scene opens: dialogue, action, description, interiority, or
  time-marker. Give the distribution as a count.
· How every scene ends. Same categories, same count.
· The five most-used verbs of physical motion
· The five most-used body parts

PASS 2 — THE ABSENCES
This is the harder pass. Report what is missing, not what is present:
· Which of the five senses appear, ranked by frequency. Name any that
  appear fewer than three times.
· Which emotions are named directly, and which are only ever implied
· What is never physically described — food, weather, hands, clothing,
  interiors, money, work, bodies, silence?
· Which characters are never alone on the page
· What questions does the text raise and never return to?

PASS 3 — THE POSITION OF THE CAMERA
Where does the narration actually stand?
· Distance: is the reader inside the character's skin, at their shoulder,
  or across the room? Does it move, and where?
· What does the narrator find funny? Quote every instance.
· What does the narrator refuse to look at directly — moments where the
  prose speeds up, cuts away, or summarizes instead of dramatizing?
· What does the narrator linger on longer than the plot requires?

PASS 4 — THE MORAL WEATHER
Reasoning only from evidence in the text:
· Which behaviors are treated as forgivable, and which are not?
· Who gets interiority, and who is only ever seen from outside?
· What does this text seem to believe about whether people can change?
· Which character is the author most protective of — who is spared
  consequences, unflattering description, or hard questions?
· Which character is treated most harshly, and what did they do to
  deserve it in the text's own logic?

PASS 5 — THE SIGNATURE
Now synthesize. Using only the patterns above:
· Name this writer's three strongest recurring instincts — the things
  they do well so consistently it's clearly not accident.
· Name the one strength that appears in this sample but is used only once
  or twice. This is the underused move. Quote it and say what it would
  look like deployed deliberately.
· Name the three patterns that have become defaults — moves that were
  choices once and are now reflexes.
· If you had to guess what this writer is avoiding writing about, based
  only on the shape of the absences in Pass 2, what would you guess?
  Flag this clearly as inference, and show which evidence produced it.

PASS 6 — THE FALSIFICATION
Argue against yourself:
· Which of your own observations above is weakest, and why?
· Which pattern might be a genre convention rather than a personal
  signature?
· What would you need to see — more pages, different scenes — to know
  the difference?

Deliver as a report with all six passes labeled. Quote constantly. Where
you are inferring rather than counting, say so in that sentence.

Prompt #2 - The Report Optimizer Prompt

Once you have it, here is an optional prompt to turn the report into a beautiful webpage with charts (oooh, fancy):

OUTPUT CONTRACT — THE BLIND SPOT AUDIT, VISUAL EDITION

Take the report above and rebuild it as a single self-contained HTML file.
No external files, no CDN links, no frameworks. All charts drawn as inline
SVG. It must open correctly from a local file with no internet connection.

PALETTE — use only these
  paper   #EDE6D6      ink     #10161C
  red     #D7263D      muted   rgba(16,22,28,.55)
  rule    rgba(16,22,28,.16)
Red marks one thing per chart: the outlier, the absence, or the finding.
Never decorate with it.

CHARTS — build exactly these five, as inline SVG

1. WORD FREQUENCY — horizontal bar, top 15 content words, descending.
   Red bar on any word appearing more than twice as often as the median.

2. SENSORY BALANCE — vertical bar, five bars: sight, sound, touch, smell,
   taste. Any sense under three instances gets a red bar and the count
   printed above it. If a sense is zero, draw an empty red-outlined bar at
   full height with "0" inside it. This chart is the centerpiece — an
   absence should be the loudest thing on the page.

3. SCENE OPENINGS AND ENDINGS — two stacked horizontal bars, one above the
   other, segmented by category (dialogue / action / description /
   interiority / time-marker). Label each segment with its count. Red on
   any single category exceeding 50% of a bar.

4. WHO GETS INTERIORITY — horizontal bar per named character, showing count
   of interior-access moments. Red on any character with zero.

5. THE DEFAULTS — three horizontal bars showing how many times each
   reflex pattern from Pass 5 recurs. Red on the highest.

TABLES — build exactly these three
  · THE ABSENCES — two columns: what's missing / where you'd expect it
  · THE MORAL WEATHER — three columns: behavior / how the text treats it /
    quoted evidence
  · THE SIGNATURE — three columns: pattern / instinct or reflex / example

TYPOGRAPHY AND LAYOUT
  Single column, max 900px, generous margins, hairline rules between
  sections. Section headers in small caps with wide letterspacing.
  Every chart sits directly under a one-sentence plain-English caption
  stating what it shows — not what it means.

RULES
  · Every number in every chart must trace to a count in the report. Do not
    estimate, round for visual balance, or invent a data point to fill a bar.
  · If the report lacks data for a chart, render the chart empty with the
    label "not enough data" rather than approximating.
  · Pass 5's synthesis and Pass 6's falsification stay as prose. Do not
    chart an inference.
  · Print the total word count analyzed at the top of the file.

Here is a sample of the output from this second prompt:

The test I just ran on my own book

As an experiment, I made a video about a novel written by AI. For this part, I took chapter one — originally written by ChatGPT — and gave it to two other models: Claude Fable 5 on high effort, and Kimi K3 on high effort. Same prompt for both of them.

I told both it was a client manuscript I was considering taking on. I never said an AI wrote it. Then I let them loose.

One of them found something the other missed. There are two kids in a photo early in the chapter. Nine pages later, the protagonist says the line “I care about all three faces.”

Three. Kimi caught it. Fable 5 — a genuinely strong reasoning model — read the same six pages and walked right past it.

Which is the actual point

These models aren’t perfect. Obviously. But that’s not a knock on them, it’s the finding. There are things you won’t catch as a human, and there are things a very smart model won’t catch either — and it’s not the same list.

So the move isn’t “pick the best model and trust it.” It’s to use several. Claude, Kimi, ChatGPT, Gemini, a local model if you’ve got one you like. See what each one finds. Then use your own judgment — your creative decisions — to sort what actually needs fixing from what a model flagged because it’s a model, not because it’s wrong.

Human plus multiple models beats either one alone. That’s it. That’s the whole thesis.

The full test, including me checking the models’ work line by line on camera, is live now:

And if you want to read the chapter yourself — the one that started all this — it’s free, along with the rest of the novel and every other free resource I’ve built, in the free author bundle.

Happy writing,

Christopher

*Written by Christopher Kokoski. Drafted as a voice memo and cleaned up with Claude — the ideas, the test, and the opinions are mine.

A few weeks ago I gave ChatGPT’s Work Mode one prompt and asked it for a novel.

It took several days. What came back was a 233-page thriller called No Safe Altitude — thirty-six chapters, about sixty thousand words, plus an EPUB, a print-ready PDF at six by nine on cream paper with the spine width calculated to the exact page count, and a filled-out Amazon upload checklist. There’s an AI disclosure page before chapter one, which it added on its own. That is more transparency than a nontrivial number of humans currently publishing on Kindle Unlimited.

I ran the whole manuscript through my analysis system. Six passes, seventy-five thriller specialists pulled from a council of a thousand, and underneath the opinions, actual measurement — every one of the 5,747 sentences, counted.

Here is what came back, and I want you to sit with it for a second before I tell you the rest.

Adverbs: 0.36 per hundred words. Passive voice: 3.5%. “As if” — the tic that gives away more AI prose than any other three characters — twelve times in sixty thousand words.

Those are good numbers. Not good-for-a-machine numbers. Good numbers. Cleaner than most published human thrillers I’ve been paid to edit.

Every AI-detection heuristic you learned in 2024 is now a heuristic for detecting bad writing in general, and this book passes all of them.

So I gave it a grade, and then I did the thing I almost skipped.

I read it.

The part the analyzer can’t reach

There’s an editor in my system whose entire job is to read dialogue out loud and check whether the characters sound like different people. I built that pass because it’s the note I give most often as a human editor and it’s the one writers argue with most.

That pass failed. Not narrowly.

And here’s what makes it interesting rather than just disappointing: the machine knew. Somewhere in the build, it produced a table — I have the screenshot — assigning eleven characters individually distinct verbal identities. Sentence lengths. Vocabulary registers. What each one does when they’re nervous. It was a good table. I’d have been happy to receive it from a human writer in a workshop.

Then it wrote a novel in which everyone is fluent in the same brand of dry, clipped, faintly clever understatement. Including the sixteen-year-old. Including the guy whose defining trait is that he doesn’t talk.

The plan was right. The page was flat. The gap between those two facts is, I think, the most honest thing I can tell you about where this technology actually is.

Indirect vs. obscure

Chapter one. A safety engineer tells the hero that he deleted “the only word they can’t moisturize.”

I have read that sentence more times than I’d like to admit in writing. I do not know what it means. My best guess is *the word legal can’t soften* — but I am guessing, and I wrote the prompt that produced this book.

Here’s the craft distinction I ended up writing on a card and taping above my desk, because I’d somehow never had to articulate it before.

Good dialogue is indirect. A character says one thing and means another, and the reader understands *more* than what was literally said. That surplus is the pleasure. That’s subtext.

What this novel does, over and over, is the opposite trick wearing the same coat. A character says something oblique and the reader understands *less*. There’s no surplus. There’s a hole where the meaning should be, and your eye slides right over it because the rhythm is confident and the sentence is clean.

Indirect means the reader gets more than the words. Obscure means the reader gets less. On the page they look nearly identical — same clipped length, same withholding, same cool tone. One of them is craft. The other is the absence of a decision.

An analyzer measures the shape. Only a reader measures the meaning.

And here is the part that turned this from a disappointment into something I’ll be chewing on for a year: that rule was *in my prompt.* Verbatim. Five words. “Dialogue is indirect and human.”

The machine didn’t ignore the instruction. It aimed at it and missed. And it has no way of knowing it missed, because the only instrument that detects the difference is a reader.

What I actually think

The book has a genuinely excellent idea in it. An AI clones a father’s voice to place a phone call, and his sixteen-year-old sets a trap for it — planting a shared family memory that never happened, on the theory that a fabricated memory can’t be scraped from anywhere. That is a real thriller device. I’ve read that beat in books people were paid well to write.

And then the novel never springs it. The trap gets set and the scene simply doesn’t arrive.

A human writer sprints at that scene. The clone answers. The details are plausible, the voice is right, it’s almost perfect — and then it gets one small thing wrong in a way only family would catch. That scene is not in this book. That scene is the entire distance between a B-minus and a book.

There is one sentence in here I haven’t stopped thinking about, though: *On the worst night of their lives, a machine had made him sound like a man who could finish a sentence.*

That’s not competent. That’s good. A machine wrote it, on its way to not writing the scene it had already earned.

So: the sentences are solved. Taste isn’t. Every remaining fix on my list is a judgment call, and judgment is still the job — which, given the week I’ve had, I’m choosing to find encouraging.

The video’s up: the full breakdown, both verdicts, and the number that made me stop and re-run the analysis:

The book is free inside the free Author Bundle on KoFi: https://ko-fi.com/s/00530fa471.

Thanks for reading,

Christopher

Most writers I talk to treat their AI like a light switch. It’s on or it’s off. You prompt, it answers, and whatever comes back is “what the AI can do.”

But almost every major AI model now ships with something closer to a dial than a switch: a setting that controls how long the model reasons before it writes a single word. ChatGPT calls them thinking levels. Claude calls it extended thinking. Gemini has its own version. The names differ, but the idea is identical. You are choosing how much silent work happens before the first sentence appears.

And here is the dirty little secret: the default position on that dial is chosen for speed and server cost, not for the quality of your prose.

What the dial actually changes

More thinking time does not make the model “smarter.” It gives the same brain more room to plan. With a low setting, the model essentially improvises: it commits to the first decent sentence and builds forward. With a high setting, it drafts internally, weighs structure, considers what the scene needs to accomplish, and then writes.

For a grocery list, that difference is nothing. For fiction, it’s the difference between a scene that wanders and a scene where every detail is doing double duty.

The tradeoff is real, though. Higher settings are slower, and on some platforms they burn through usage limits faster. The dial is not “always max it.” The dial is “match the setting to the job.”

When authors should turn it down

Low thinking settings are genuinely better for some work. Brainstorming wants speed and volume, not deliberation. Quick lookups, name generation, “give me ten ways this scene could open,” rapid back-and-forth while you think out loud: all of that benefits from an AI that answers instantly. Deliberation would just slow the conversation you’re having with yourself.

When authors should turn it up

Turn the dial up when the task has structure to hold: drafting a full scene, revising for a specific effect, untangling a plot problem, checking a chapter against your outline, or any prompt with several constraints the model has to satisfy at once. Constraint-heavy prompts are where thinking time earns its keep, because the model needs room to notice all your rules before it starts writing, not after.

A simple rule of thumb: if you would want to outline before writing it yourself, raise the dial. If you would just start typing, don’t bother.

The uncomfortable part

More thinking usually buys better prose. Usually. Not every level on every platform earns its position, and the differences between adjacent settings can be smaller than the marketing implies. The only way to know what your setting is actually worth is to test it, with the same prompt, at every level, and compare the results side by side.

Which is exactly what I did this week. I ran one fiction prompt through every thinking level in the new ChatGPT and scored every output with the same analyzer. One paid level lost to the free model, and it happens to be the default. Ooops.

The full test, the scores, and the podium are in today’s video (linked below), and the analyzer I used is free this week only, then it comes down. If you want the whole plain-English map of which model, level, and mode fits each part of writing a novel, that’s the new ChatGPT Book Blueprint Bundle (Launch price this week only).

Check out the video:

Every AI model launch follows the same script.

The tool drops, the demos look incredible, your feed fills with prompts to try, and somewhere in the excitement, thousands of writers paste in the most valuable file they own without reading a word of the paperwork.

Not because they’re careless. Because nobody ever taught them what to look for, and the documents are written for lawyers and programmers.

So today, the thing itself: the exact check I run before any model, from any company, ever sees my work. It takes about ten minutes, you only have to learn it once, and it works on every tool that exists now and every tool that launches next year.

The five questions that matter

Does a copy exist after I close the tab? Open the company’s privacy or data-retention page and Ctrl+F the words “retain,” “delete,” and “days.” You’re looking for a number. Thirty days, five years, indefinitely. Whatever it is, you want to know it before you paste, not after.

Can I turn that off? Search “opt out” and “zero data retention.” Some tools offer it to everyone, some only to business accounts, some not at all. “No opt-out” isn’t automatically a dealbreaker. It’s a trade, and trades should be made on purpose.

Is my writing training the model? Ctrl+F “train.” You want a plain sentence that says your inputs are not used for training. If you can’t find one, assume the answer you don’t want.

What happens if I get flagged? This is the one everyone misses. Retention limits usually have an exception for content flagged by safety systems, and flagged material can be held far longer, sometimes with no stated limit. Search “flag,” “review,” and “safety.” Fiction research trips these systems more than you’d think. Your villain’s poison is somebody’s classifier alert.

Whose rights are in this file? This one isn’t in the paperwork; it’s in yours. If the pages contain a client’s chapters, anything under NDA, or someone’s real medical history and family secrets, the question stops being “do I trust this company” and becomes “does my contract allow this.” Ghostwriters and memoirists, this is your whole question.

The rule that falls out of it

Once you’ve run the check a few times, you’ll notice it always lands in the same place.

Work you own outright goes wherever you decide it goes, eyes open. Work containing anyone else’s material or secrets stays out of anything with mandatory retention, full stop. And the sketchy-looking research gets done somewhere you’ve chosen deliberately, not wherever you happened to have a tab open.

Ten minutes of reading, and you’ll know a tool better than 99% of the people posting prompt lists about it.

I ran this exact check on a model this week

Fable 5 came back from its government suspension on July 1st, and it’s genuinely the strongest writing model an author can touch right now. It also failed question two and has the strangest answer to question four I’ve ever found.

I put the whole thing on screen, receipts included, in this week’s video:

And if you want the prompts, blueprints, and tools I use to test these models on real fiction, they live in my Ko-fi shop: ko-fi.com/ckokoski.

Cheers,

Christopher

In March 2011, Lindsay Buroker made $724 from her two indie novels.

By her fourth book, she was earning a multiple of that. By book ten, she was full-time. Today she has over a hundred titles in her catalog and the income to match. She published her exact March 2011 royalty number on her own blog the following year — $724 — and it became one of the most-referenced data points in indie publishing.

That gap between book two and book five is where indie author income actually changes shape. Not because the writing gets better. Because the math does.

The number that should change how you publish

About 80% of indie authors with one to three books earn less than $100 per month from their writing.

Among authors with twenty-five or more books, the median is $3,000 per month — with 40% earning over $5,000.

Both numbers come from the same survey. Written Word Media polled 1,346 indie authors in their 2025 survey. The difference between bottom and top isn’t talent. It’s catalog size.

The income distribution isn’t normal. It’s not close to normal. A few authors earn six and seven figures. Most earn lunch money. The variable that separates them, more than genre or platform or marketing budget, is whether they kept publishing past the third book.

Written Word Media’s 2023 survey put numbers on it more precisely. Authors earning $0 to $249 per month had a median of five published books. Authors earning $5,000 to $7,500 per month had a median of thirty.

Five books get you the bottom bracket. Thirty books gets you a livable income.

Why the inflection point is between books three and five

Three things change once an indie author crosses into the four-to-five book range, and each one is mechanical.

The launch window stacks. Amazon’s algorithm boosts new releases for roughly thirty days. If you have one book, you have one window. If you have four books and you launch the fifth, that launch pulls readers into the previous four titles. Each new release becomes free marketing for the entire backlist.

Read-through becomes a metric. The percentage of readers who finish book one and start book two becomes the actual profitability number. A 50% read-through rate on a five-book series means each new reader is worth roughly three full-priced books in long-term revenue. You can’t measure read-through at one book. You can barely measure it at three.

The superfan math kicks in. A reader who loves your work and has only one book to buy from you can spend $4. The same reader with thirty books to choose from can spend $120. Same fan, same enthusiasm, thirty times the lifetime value.

Across multiple years of indie surveys, the strongest predictor of author income isn’t genre. Isn’t quality. Isn’t ad spend. It’s catalog size. Written Word Media calls catalog size “the strongest predictor” in their 2025 report. Draft2Digital reports that 75% of all indie book sales — fiction and nonfiction — are part of a series, by both unit and dollar volume.

The market is structured to reward catalogs. Authors with one book are pricing themselves into a market designed for authors with five.

Real authors, real arcs

The case studies follow the same pattern, just at different speeds.

Lindsay Buroker, as mentioned: $724 in March 2011 with two novels out. By her fourth book, multiples of that. Full-time income by year three of publishing. Over a hundred books today.

Michael Anderle is the fastest documented version. He published his first book in November 2015. From his own podcast interview in 2017: month one earned him $430 gross, month three earned him over $12,000 gross, and by July — eight months in — he had cleared $50,000 net. He hit thirty-five titles within two years. Anderle is the outlier, but he’s an outlier in speed, not in pattern. He ran the catalog play at a faster cadence than almost anyone else.

Mark Dawson, who runs Self Publishing Formula, reportedly earns around $450,000 a year from his thriller catalog. He discusses his income in less specific terms publicly, but the pattern matches — a deep catalog across multiple series (John Milton, Beatrix Rose), built methodically over years, monetized through email lists and read-through.

The arc compresses or stretches based on who’s running it. The shape doesn’t change.

The trap that keeps most indies at one book

Most indies stop publishing before they reach the breakout zone.

Not because they want to. Because book one is the hardest one. You absorb every painful lesson: cover design, formatting, blurb writing, launch logistics, reader psychology. By the time you’re done, you’re tired. The financial feedback is also discouraging: nine months of work, lunch money in return.

The rational response is to quit.

But the rational response is wrong. Because one book is the worst position on the curve. You’ve absorbed all the learning costs and captured none of the catalog returns. Quitting at one book is statistically equivalent to lottery odds. Pushing through to five changes the distribution you’re playing inside.

The authors who break out aren’t the ones who got better at writing between book two and book five. They’re the ones who kept publishing while the numbers said it wasn’t working yet.

What to do with this

Three things follow from this data.

Stop optimizing your first book for revenue. It can’t, statistically. Optimize it for read-through instead. Leave room for a sequel. End on a question worth answering. Build an email list aggressively from day one. Your first book is a customer acquisition asset, not a revenue asset.

Track read-through, not book-by-book sales. Read-through is the actual profitability metric. It tells you whether your catalog is set up to compound or whether each book stands alone.

Plan your next three releases before you publish your first. Not the full manuscripts — just the titles, the audience tiers, and the order. A series planned in advance reads as one coherent arc. A series built reactively reads like three unrelated books with the same byline.

This is what indie publishing actually looks like at the income level most aspiring writers imagine when they think about quitting their day job. The math is brutal at one book. It starts working (even in the background) at five. It compounds at thirty.

A note on the data sources

Most of the numbers above come from Written Word Media’s annual indie author surveys (2023 and 2025, with 1,346 responses in the most recent), the Alliance of Independent Authors 2025 income survey (n=1,520 qualifying respondents), and Draft2Digital’s series sales data. Selection bias matters — these surveys mostly reach committed, marketing-active indies, which means the true indie author income distribution is probably worse than these numbers suggest. The US Authors Guild 2023 survey, which captured a broader 5,699-author sample, put the median book income at just $2,000 per year. The catalog effect holds in both data sets. The starting baseline is just lower than the committed-indie numbers indicate.

Disclaimer: Obviously, the math isn’t guaranteed for every author who writes three or more books. That’s simply not the case. However, there is a pattern here worth noticing and a system worth following.

A note on something I just tested

I spent last week running an experiment to see if AI could produce a series teaching this exact lesson — a three-book indie author income series, written by an agent swarm under tight constraints.

Forty AI agents. One prompt. Six hours. Three books on this topic, with real case studies, real numbers, and the catalog math at the core.

I made a video walking through the full demo on YouTube — the kind of swarm architecture I used, where it broke, where it surprised me, what it produced.

If you want to read the three books the swarm actually produced, they’re on my Ko-fi page (Learn more here). Each one includes a transparency note explaining the AI-experiment context up front. The case studies are verified; the prose is the experiment.

The math in this newsletter is real either way. Whether you read the books or not, the catalog effect doesn’t care.

Cheers,

— Christopher

Claude Fable 5 is the most interesting thing to happen to book writers since Opus, and almost everything being written about it is aimed at coders and businesses.

Nobody is writing the author playbook.

So here it is. This issue stands on its own: ten copy-paste prompts that go beyond my FABLE Method one-pager, real audit data from a thriller I built live with Fable 5 this week, and the craft truths about how this model actually behaves with a manuscript in its hands.

If you want to watch the whole live build, the bodies, the bibles, the audit catching twelve structural problems before Chapter 2 existed, that’s in this week’s video:

[▶ WATCH: I Built a Thriller Live With Claude Fable 5]

But you don’t need the video for this issue to pay for itself. Let’s go.

First, what Fable 5 actually is (60 seconds, no hype)

Three facts worth knowing, because they change how you should work:

1. It’s a new tier, not just a new number. Fable 5 is the first model in Anthropic’s Mythos-class tier, which sits above Opus. There’s a sibling model, Claude Mythos 5, that shares the same underlying model and is available only to approved organizations; Fable 5 is the one you and I can use, with additional safety measures on dual-use capabilities. For fiction work, you are not missing anything: it’s the same brain. (Anthropic’s announcement is here if you want the official version.)

2. The headline capability for authors is long-context consistency. Earlier models could write a beautiful chapter. Fable 5 can hold a book: who knows what and when, which gun is on which mantel, whether Chapter 19 is still loyal to Chapter 3. That’s the difference between an AI that writes scenes and an AI that can be your continuity department.

3. It critiques best when critique is split from generation. Ask for a draft and a critique in the same breath and you get a polite critique of work it’s invested in. Ask in a separate message and you get an editor. This one behavior shapes half the prompts below.

The inside stuff: five things heavy use taught me that nobody is talking about

I’ve been running Fable 5 hard on fiction since launch. These are practitioner observations, not spec-sheet claims, and every one of them changed how I write.

1. Never let it invent your premise. Ask any model to “give me a thriller premise” and you get the statistical average of the genre, because that’s literally what it is. Bring the seed yourself, then make Fable 5 attack it. The attack is where this model is shockingly good: it found the load-bearing flaw in my premise (a logistics problem I’d hand-waved) in its first response, and the fix it proposed became the engine of the whole book.

2. Voice before structure, always. If you outline first, every chapter inherits generic outline voice. Lock a 300-word voice sample first, codify it into five written rules, and paste those rules into every drafting session. Fable 5 holds a codified voice with unnerving fidelity; it cannot hold a voice you never defined.

3. Plot holes convert into plot devices, but only under interrogation. Live example from my build: the interrogation pass caught that my president hid a body in nine impossible minutes. Instead of patching it, the fix made the impossibility visible to a character: he asked for two minutes, took twelve, and the agent who counts everything logged the discrepancy. The plot hole became a clue. That’s what a separate-message interrogation buys you that “make it better” never will.

4. The bible is external memory, and files are the source of truth. Conversations drift, get long, get messy. A living story bible (cast sheets, world rules, timeline, beat map, promises register) updated after every chapter is what keeps a 400-page book coherent. Export locked chapters as files. The conversation is scaffolding; the files are the building.

5. Audit the design, not just the prose. The highest-leverage pass I ran all week wasn’t on the manuscript. It was on the bible itself, before chapters 2 through 18 existed. Single-purpose passes over the beat map and promises register caught twelve structural problems, including an entire act with no kills in a serial-killer thriller. Here’s the cost logic in one picture:

Same error, five different discovery dates, wildly different rewrite bills. Catch it in the book bible and it costs a sentence.

The receipts: real numbers from the live build

Everything below comes from one session: a suspense thriller called RED HOUSE, built stage by stage with the FABLE Method.

Thirty problems caught before Chapter 2 existed: three fatal premise assumptions, eight AI-trope character names (yes, the model audits its own naming clichés if you ask), seven chapter-level issues in interrogation, and twelve structural findings from the design audit.

The promises register is the unsung hero: fourteen tracked setups, each one either assigned a payoff beat or flagged. Four unassigned promises surfaced in the audit, which are four broken-promise reader complaints that now never get written.

The Bonus Prompt Pack: 10 prompts beyond the one-pager

These extend the FABLE Method (Forge, Anchor, Build, Layer, Elevate). They work identically in Fable 5 and Mythos 5, with the same underlying model. Copy, paste, fill the brackets.

1. The Acquisitions Gauntlet (premise stage)

Act as three acquiring editors with different mandates: one literary
imprint, one commercial thriller list, one audio-first publisher.
Each reads this premise: [PREMISE]. Each gives the one reason they
would pass and the one change that would make them bid. Then merge
the three changes into a single sharper premise.

2. The Voice Drift Detector (any chapter)

Here are my five voice rules: [RULES]. Here is Chapter [N]: [TEXT].
Score the chapter 1 to 10 against each rule separately, quote the
three worst-offending sentences per rule, and rewrite only those
sentences in compliance. Change nothing else.

3. The Promises Auditor (full manuscript)

Full manuscript attached. Build a table of every setup that creates
reader expectation: object, line of dialogue, wound, or threat.
Columns: promise, chapter planted, payoff chapter or NONE, verdict
(paid, partial, broken). Do not fix anything yet. Flag the three
broken promises a reader is most likely to notice.

4. The Continuity Deposition (full manuscript + bible)

You are a hostile fact-checker deposing this manuscript against the
attached bible. Find every contradiction in timeline, geography,
character knowledge (who knows what, and when they learned it), and
physical detail. Cite chapter and line for each. Rank by how likely
a careful reader is to catch it.

5. The Subtext Pass (dialogue lens)

Dialogue-only pass. List every line where a character states their
emotion or motive directly. For each, give a replacement that
conveys it through deflection, action, or silence. Mark the lines
where on-the-nose is actually correct and should stay.

6. The Entrance/Exit Audit (scene lens)

Scene transitions only. For every scene, quote the current first
and last lines. Flag scenes that enter before the interesting
moment or exit after it. Propose later entrances and earlier exits,
by sentence, without rewriting the middles.

7. The Lull Mapper (tension lens)

Chart the tension of this manuscript scene by scene on a 1 to 10
scale, as a table. Identify every lull longer than two scenes, then
tell me which lull is strategic (a held breath before a detonation)
and which is dead air. Suggest the cheapest possible detonation for
each dead lull.

8. The Chapter Autopsy (the interrogation, expanded)

Interrogate the attached chapter, separate from drafting. Answer
only: the weakest moment and why, the sentence where the voice
slips, the paragraph a hostile editor cuts first, the strongest
line that must survive revision, and one structural risk this
chapter creates for later chapters.

9. The Bible Sync (after every locked chapter)

Compare this locked chapter to the attached bible. Output only: new
canon facts to add (names, times, geography, objects), promises to
add to the register, promises this chapter paid off, and any line
that contradicts existing canon. Format as a changelog I can paste
into the bible.

10. The Anti-Average Pass (my favorite)

You were trained on a million versions of this genre. Identify the
five most statistically predictable moments in the attached
chapter, the beats most models and most writers would default to,
and for each, propose the version one standard deviation stranger
that still honors the genre contract.

Nonfiction bonus: The Claims Audit

Single-purpose pass for nonfiction: list every factual claim in the
attached chapters with its supporting evidence from my claims
bible. Columns: claim, chapter, evidence, strength (strong, thin,
missing). Do not soften any claims; flag the ones I cannot
currently defend.

The 60-second workspace (the part that isn’t on the one-pager)

The method fits on one page. The setup that makes it survive a 400-page book takes sixty seconds:

1. Create a Claude Project (or the equivalent workspace in any AI).

2. Drop the bible and your five voice rules into Project Knowledge, so every new chat starts already knowing your book.

3. Run one chat per chapter. Fresh context, zero drift.

4. When a chapter locks, export the file. Files are the source of truth, not the conversation.

And a confession: I ran this week’s entire live build in one long chat, because I got excited. Which is exactly the failure this setup exists to prevent. Do as I say, not as I just did. The full setup, the live build, and the moment the audit caught my Act Three problem are all in the video:

Keep this issue

Bookmark it, or better, paste the ten prompts into a note inside your book’s project so they’re one click away while you write. If you build something with them, reply and tell me what the audit caught; the best war story goes in a future issue.

And if you’re new here: the FABLE Method one-pager, the Claude AI Book Blueprint, and the full Claude Author Bundle live here: https://ko-fi.com/s/c38d7ba38d.

Write smarter. Build stronger books.

Christopher Writing Secrets

I got a comment on my latest Opus 4.8 video this week that stung: “It’s the same as before.”

And honestly? I get why someone would feel that way. New Claude, new version number, another author runs another test. If you’ve watched a few of these, a launch-day demo can start to blur together.

But that comment is pointing at something real — and it’s worth unpacking, because it’s the exact trap that makes it nearly impossible to tell whether a new AI model is actually any better than the last one.

Every new model looks identical in easy mode

Here’s the not-so-secret secret of most “new model” demos, including plenty of mine: they test at easy mode.

Write me a premise. Outline this. Draft a chapter. At that level, GPT, Gemini, this year’s Claude, last year’s Claude — they all hand you something competent. They all look the same. Because the easy tasks stopped being hard for these models a long time ago.

So when a new version drops and someone runs it through the usual paces, of course, it feels the same. The test isn’t capable of showing a difference. You’re timing two sprinters with a stopwatch that only counts in whole minutes.

If there’s a real jump in capability, easy mode is the last place you’ll see it.

The differences live at the edges

If you want to know whether a model actually improved, you don’t ask it to do something easy slightly better. You ask it to do something it should fail at — and watch how it fails.

That’s the whole reason I didn’t run another chapter demo this time. I asked Opus 4.8 to write a horror musical — story, songs, meter, sustained dread, a separate voice for every character, two genres that openly fight each other — all at once. It’s the hardest creative test I know how to run, and I break down exactly why in the video.

A test like that separates “they all look the same” models in seconds. Easy mode hides the gap. The edge exposes it.

What this actually means for you

This isn’t really about my video. It’s a tool you can use every single time a new model launches and the whole timeline starts arguing about whether it’s “actually better.”

Don’t judge it on the easy stuff. Judge it on the one thing your writing genuinely needs and that AI usually botches — your hardest 5%. For most novelists, that’s one of three: holding a distinct character voice across a whole book, sustaining tension without it going flat, or matching a specific prose style instead of defaulting to beige.

Run that test. If the new model is genuinely better, that’s where you’ll feel it. If it isn’t, you just saved yourself an upgrade you didn’t need.

So…was the new Claude actually different, or was that commenter right? I’d rather show you than tell you. The full test, warts and all, is here:

Watch it here →

Watch where it cracked and where it genuinely surprised me, and decide for yourself.

— Christopher

I asked Claude to write me a horror musical.

Not a horror story. Not a couple of creepy lyrics. A musical. Acts, numbers, a villain who gets his own song — something that’s supposed to scare you and then send you home humming.

On paper, this is an unhinged thing to ask a language model to do. Which is exactly why I did it. Opus 4.8 just landed — the newest, most capable Claude you can actually get your hands on — and instead of running another tidy “write me a chapter” demo, I wanted to throw one of the single hardest creative tests I know at it.

Here’s why a horror musical is that test.

A musical is the hardest writing test there is

Most “can AI write?” demos are easy mode. Premise, outline, one chapter. One skill at a time, graded gently.

A musical makes the model do all of it at once:

Story. Real acts, real escalation, a turn you don’t see coming.

Lyrics that actually scan. Rhyme and meter — lines a human could sing out loud without tripping over them.

Tone, held for the whole runtime. And horror is the cruelest tone there is, because dread evaporates the second the writing gets lazy.

A separate voice for every character. The ingénue and the monster cannot sing in the same register.

Two genres that actively fight each other. Tip too campy and the horror dies. Tip too grim and the whole “wait, why are they singing?” problem swallows the show. Sweeney Todd and Little Shop of Horrors are masterpieces precisely because that balance is so hard to land.

Now here’s the part that matters to you — even if you never write a single song in your life.

Every one of those skills is a skill your novel needs

Watch what each demand actually maps to:

Meter is prose rhythm with the training wheels off. If a model can hold a beat across a sixteen-line number, it can hold cadence across your chapters. If it can’t, you’ll hear it limp in your dialogue too.

Horror in a lyric is harder than horror on a page. You’ve got rhyme and line length fighting you the whole way, with a fraction of the words to land the scare. So if it can unsettle you in a verse, it can absolutely sustain dread across a manuscript. The tighter the constraint, the better the stress test.

Distinct character songs are distinct character voices. The exact thing that sinks most AI fiction — everyone sounding like the same mildly helpful narrator — has nowhere to hide in a musical. Either the monster sounds like the monster or he doesn’t.

The genre-fusion balance is tonal control, full stop. If you write anything that blends — horror-comedy, cozy mystery, literary thriller — this is the muscle you lean on, and a horror musical shows you in about ten minutes whether the model has it.

That’s the whole reason I run tests like this. I don’t care about the spectacle. I care about what the spectacle reveals — and a horror musical tells you more about a model’s real creative range than a dozen chapter demos ever will.

Where it cracked

I’m not going to pretend it was flawless, because it wasn’t. That’s the honest part — and it’s exactly where you’d still need to step in as the writer.

And then there was the moment that actually got me.

The one number I keep replaying

[DROP IN THE STANDOUT — the song or line that genuinely surprised you. Tease it, don’t spoil it. One or two sentences is plenty.]

I’d rather you hear it than take my word for it. In the video, I react to the whole thing in real time — the premise, outline, writing quality — so you can decide for yourself whether the new Claude earned the noise.

Watch the full experiment here →

If you write fiction and you’ve been wondering what Opus 4.8 actually changes for you, this is the fastest way to find out: watch it attempt the hardest thing on the menu.

— Christopher

The book is real. Cover. Word count. Twenty-two chapters. I’m holding it on the desk next to me right now.

The full walkthrough is on YouTube — link at the bottom. This is the field manual that goes with it.

Read this first if you want the prompts in copy-paste form. The video shows them on screen; this newsletter lets you steal them.

The six moves that wrote the book

There’s no one prompt that writes a (good) novel. Not yet, anyway. There are six, in order, and the order is what makes it work.

I broke the workflow into six steps because every other AI writing tutorial I’ve seen treats it like a single command. Write me a novel about a fired tech worker. That gives you four pages and a stall. The work is in the sequence — context, then pressure, then structure, then voice, then chapters, then revision. Skip a step and the book collapses two-thirds of the way through.

The most important step is the one nobody does — Step 1.

The step nobody does

Before I wrote a single chapter, I set up a project inside Le Chat and pasted in a one-page rule set that governs every output for the rest of the workflow. Tone. POV. Banned words. Voice rules. Names to avoid. Six writing rules that apply to every chapter.

This is the difference between an AI chapter that sounds like every other AI chapter and one that sounds like a specific book.

The full project instructions block is on the infographic above. Paste it into your Le Chat project once. Adjust the character names and the book title. Then never think about it again — Mistral holds the rules across the entire workflow because projects keep memory across sessions.

The prompts (steal these)

Five prompts, in order. Run them in the same Le Chat project.

1 — Premise stress test:

I have a premise. Stress test it before we build anything else.

PREMISE: [your premise]

DO THE FOLLOWING:
1. THE DRAMATIC QUESTION — Name the single dramatic question this novel answers. One sentence. No hedging.
2. THE THREE FAILURE MODES — Name the three most likely ways this premise collapses in the middle. For each, one-sentence fix.
3. THE COMP TITLES — Three published novels from the last six years with similar DNA. For each, one sentence on what mine has that the comp doesn't.
4. THE LOGLINE — Rewrite the premise as a 30-word logline ready for a query letter.
5. THE VERDICT — Should I write this book? Yes or no. One sentence why.

Be useful. Not generous.

2 — Outline: the full prompt is on page 4 of the workflow PDF below. The key move: hard constraints. Daniel’s tools must get less broken over time, but his life must get more complicated. That single line shapes 22 chapters.

3 — Character bible: five characters, three voice ticks each, one physical object per character. The banned-words list for the antagonist is what makes her sound like a real CEO and not a stock corporate villain.

4 — Chapter generation: one prompt per chapter. Roughly ten minutes start to finish — generation plus a quick read and one revision request. Twenty-two chapters at ten minutes = about four focused hours.

5 — Full-draft revision pass: this is the move you can’t do with most other models. Mistral Medium 3.5 has a 256,000-token context window — about 190,000 words. You paste the entire manuscript back in. One prompt. One structural diagnosis.

What I didn’t say in the video

Two things I cut for runtime that matter if you actually do this:

One — pre-test every prompt before you start writing chapters. If a prompt produces a weak result twice in a private chat, rewrite it before you commit to running it 22 times. The cost of a bad prompt compounds chapter by chapter.

Two — Mistral’s instinct on dialogue is better than its instinct on description. I ended up rewriting three of the 22 chapters by hand because the descriptive passages went flat, but I almost never touched the dialogue. If you’re spot-checking output and you have to pick one thing to read carefully, read the description.

The book

It’s called Severance Package. Daniel Park, 43, fired by his own company’s AI, decides to fight back with the rusty Python from a bootcamp he barely finished. He’s bad at it. The book is about what happens anyway.

I’ll share more about the book itself in a future post — including the cover and the first chapter for anyone who wants to read it.

If you want the full ecosystem of workflows

Author OS is the system I use for every book I write with AI. It’s not a prompt pack. Prompt packs break the second a new model drops. It’s the operating system underneath the prompts.

If you write books with AI, that’s the thing worth having.

▶ Watch the full video here:

▶ Author OS on Ko-Fi: https://ko-fi.com/s/bfc2d2a978

Talk soon.

— Christopher

Last week, Anthropic shipped a feature called Auto Dream.

The pitch is exactly as weird as it sounds. Their AI agents now go “to sleep” between sessions. While they sleep, they review their own memory — every conversation, every decision, every correction the user made — and consolidate it. Resolve contradictions. Prune what’s stale. Re-index what matters.

Modeled, almost on the nose, on what your brain does at night.

The numbers Anthropic published are striking. Big jump in task success rates. Up to six times improvement on long-running projects. All from housekeeping. Same model, cleaner memory, dramatically better output.

I read the announcement and immediately thought: this is the bug that has been quietly killing every long manuscript I have worked on with an AI.

The Bug That's Quietly Killing Your Long Manuscripts

You know the bug. You have hit it.

You are six weeks into a novel. You have made dozens of decisions. Your protagonist’s eyes are hazel, not gray. You cut the sister subplot in chapter four. You moved the inciting incident from spring to autumn. You decided your magic system has a hard cost.

Then you ask Claude for help with chapter twelve. And Claude — confidently, cheerfully, completely — gives you advice based on the previous version of your book. The one you have moved past.

This is not the AI being stupid. It is memory bloat. Your context window has filled up with old, half-baked, contradictory notes from a dozen sessions. The model is doing its best with junk data.

Anthropic’s fix is to let the agent dream. Review. Clean. Re-index.

The catch: their version runs at the agent’s memory layer, on their managed platform, in research preview. Most of you reading this will never touch it.

So I built one we can use. Today. On any Claude session, any manuscript, any book.

Meet the Dreaming Skill for Authors

I am calling it the Dreaming Skill for Authors.

It is a single skill file plus a small launcher. You drop it into your manuscript folder. You can run it manually inside Claude.ai with zero installation, or set it to run automatically every night while you sleep.

Either way, the skill reads your manuscript, your notes, and a memory file it has been building over time. It pulls out everything durable — your character decisions, your style preferences, your corrections, the rules of your world. It merges what is new with what it already knew. It writes a clean canonical memory file that the AI loads at the start of every future session.

You wake up. Your AI knows your book.

That is the gift. That is what every long-form writer needs and almost nobody has.

The skill runs four phases. Inventory — what files are here? Signal extraction — what is worth remembering? Consolidation — merging the new stuff with what it already knew, resolving contradictions, and pruning what is stale. And writing the outputs — a memory file the AI will load in every future session, plus a dream log entry that tells you what changed.

The dream log is the thing I love most. It is dated, append-only, written in plain English. Last night, I learned X. I reinforced Y. I noticed you have been ending chapters on single-line beats — promoting that to a working pattern. Open question for you: Is Tomas seventy-eight or eighty-two? I have evidence for both.

It reads like a journal of what your AI is learning about your book.

Read This Before You Download It

This is not a port of Anthropic’s Auto Dream. Theirs runs at Claude’s memory layer. Mine runs at your project’s memory layer. Theirs is automatic for the AI. Mine is automatic for your book. They are not competing — they are complementary. If Anthropic ever opens up Auto Dream for general use, and they probably will, you should use both.

But until then, this is what we have.

How to Grab It (No Strings)

Free. Today. No email gate. No upsell.

It is in my free Ko-fi bundle alongside everything else I have given away this year. The bundle has the skill file, double-click launchers for Mac and Windows, setup guides for nightly automation, integration guides for AuthorClaw and other agents, a visual dashboard, and an example project so you can test it before trusting it on your own book.

Get the free bundle on Ko-fi →

I also recorded a full walkthrough video showing it running on a real manuscript — what the dream log looks like the first morning, how the launcher works, and how the nightly automation gets set up. About six minutes.

Watch the full demo on YouTube →

Why free? Two reasons. Anthropic’s version will (likely) land for everyone eventually. While it is still useful, I want it in your hands. And if you have been on this newsletter a while, you know I do not gate small useful things. Pay for the deeper workflows on Ko-fi if they earn it. The little tools are gifts.

Even Without the Skill, Try This

One tactic you can use right now, even without the skill:

Open your current Claude project. At the end of a long working session, type:

Summarize every decision, preference, and correction we made in this conversation that you should remember for future sessions. Format as a brief bulleted list, suitable for a project memory file.

Claude will produce a clean summary. Copy it. Save it somewhere with your manuscript. Next session, paste it in at the top.

That is a manual one-paragraph version of what the full skill does. It will not catch as much. But it will catch the worst stuff. Try it today.

Cheers,

Christopher
Writing Secrets

This week’s video puts DeepSeek V4 through five tests — premise, outline, chapter, revision, and one for fun. It’s already up, and I’ll link it at the bottom.

But there was a sixth test. One I started running while I was editing the video, and one I couldn’t fit because the result took a turn I didn’t expect, and the analysis would have blown the runtime past 20 minutes.

So I’m telling you about it here. This is the version of the story that only the newsletter gets.

I asked AI to write a complete 77,000-word thriller. In one run. No human input. No mid-stream prompts. Just hit go and walk away.

Here’s what happened, what it taught me, and why this is the most useful thing I’ve learned about AI for authors all year.

The Setup

After finishing the scorecard tests for the video, I had everything ready to keep going — the premise, the twenty-chapter outline, a thriller called Kiss of the Weld about a saturation diver trapped in a sabotaged underwater habitat at 300 meters. Halomate (the platform sponsoring this week’s video) has an auto mode feature that’s supposed to let the AI continue working on its own without you prompting it.

I turned it on. Walked away. Came back fifteen minutes later expecting either nothing or a couple of chapters.

What I came back to was a complete manuscript. Twenty chapters. An epilogue. A character appendix. Internally consistent. Plot threads tracked. Hook lines at every chapter break. Voice held across the whole arc.

And I’ll tell you — the opening is genuinely good. Not “good for AI.” Good.

The weld pool was running too hot. I saw it before the alarm could tell me — that faint orange shimmer under the surface of the molten steel, a colour that belonged in a forge, not in a saturation diver’s prep box at three hundred metres.

That’s chapter one. Real opening line, real sensory grounding, real character voice. I wouldn’t have flagged this in a slush pile.

So why am I telling you it failed?

Look at the Chapter Lengths

Chapter 1 is a real chapter. Nice. Chapter 11 is a beat sheet pretending to be a chapter. Not so nice. Chapter 16 — the climax-adjacent chapter where the protagonist is racing against a countdown — is 251 words long.

The total came out to roughly 18,000 words. Novella length. About a quarter of what I asked for, written in about fifteen minutes.

Here’s What I Think Is Happening

As context fills up, the model starts compressing. It stops writing scenes and starts writing summaries of scenes. It stops dramatizing and starts narrating. The middle chapters of Kiss of the Weld aren’t chapters. They’re plot notes the model wrote to keep itself oriented while it raced toward the ending.

You can feel it in the prose. Here’s chapter 1, where the protagonist counts seconds to manage panic:

I have been counting in fours since I was fourteen. I do not think about why. I will tell you that on a wet road in November my brother Thomas was alive in the passenger seat for three minutes and forty seconds after the car went into the canal, and I was on the bank, and I could not work the door, and my mother was screaming, and the screaming did not unbuckle anything.

And here’s chapter 9, where the same protagonist is supposed to be making a high-stakes decision about a sabotaged valve:

Sam had mentioned it that morning — a throwaway line, a remember that time story about the refit crew getting it wrong and nobody noticing for three months. I had filed that information away. And now I was going to use it.

That’s not the same writer. Chapter 1 trusts the reader. Chapter 9 explains itself like a Wikipedia entry. Same model, same session, same voice prompt. The difference is fatigue.

And here’s the part that fascinates me: when the model hit chapter 18, auto mode crashed. I had to manually prompt it to continue. When I did, it didn’t finish chapter 18 — it skipped ahead to the epilogue and wrote 2,082 words of clean, controlled, thematically grounded prose.

The model wasn’t out of its capabilities. It was out of attention. Given the choice between finishing a middle chapter and protecting the ending, it protected the ending.

That’s not a bug. That’s a model making the same triage decision a tired writer would make at 2 AM with a deadline at dawn.

What This Actually Means for Authors

The lesson isn’t “AI can’t write novels.” The lesson is more useful than that.

One. AI can produce novel-quality prose at the start and end of a long generation. The frame holds up. If you’re using AI for opening chapters or for closing scenes, you’re working with the model’s strongest output.

Two. AI cannot sustain quality across the middle of a long generation without human intervention. The middle of every novel is where the work is. It’s where the model gets tired, and tired models compress.

Three. Auto mode is not “set it and forget it.” It’s “set it and stay in the room.” If I’d been checking in every two chapters, I could have caught the compression at chapter 4 — when chapters dropped from 2,000 words to 800 — and re-prompted with “expand chapter 4 to full novel scale before continuing.” The model wasn’t going to do that on its own.

Four. And this is the one that surprised me: the model’s instinct to skip chapter 18 and protect the epilogue is exactly what a good human editor would have advised. If you can only finish one of these, finish the ending. The AI made the right structural call. It just couldn’t tell me it was making it.

Why Halomate Mattered for This Experiment

This experiment couldn’t have happened in a regular ChatGPT or Claude window. The context window would have blown out somewhere around chapter 6, and the model would have lost the premise, the outline, and the voice.

Halomate’s persistent memory — what they call “Mates” — is what kept the model oriented across all twenty chapters. The Mate held my premise, my outline, my voice samples, and the running manuscript as a continuous reference. The model could check itself against the source material without me re-feeding it.

That’s the actual unlock. Not “AI writing a novel.” It’s “AI writing a novel while still remembering what novel it’s writing.” Most platforms can’t do that.

If you want to try it: Halomate is here. They’re sponsoring this week’s video and they’ve been good to work with. Free tier exists. Auto mode lives behind the paid tier and burns credits fast — fair warning — but the persistent memory feature is the part I’d actually keep using.

Watch the Full Test

The newsletter is the honest version of this story. The video is the entertaining version — five DeepSeek tests including the prompts I used (which you can steal). They’re meant to be read together, and you’ll get a different angle from each.

👉 Watch the full DeepSeek V4 test here

It’s the most fun I’ve had testing a model in months.

The Takeaway

If someone tells you AI can write a complete novel autonomously in 2026, they are either lying to you or they have not actually tried it. AI can write the first quarter of a novel beautifully. AI can write the ending beautifully. The middle is where the human still has to live.

That’s not a disappointing finding. It’s a useful one. It tells you exactly where to spend your time, and exactly where to let the AI run.

I’ll be back next week with something else weird I tried.

— Christopher

Anthropic released a new tool last week called Claude Design, and the entire internet is talking about it wrong.

Every breakdown I’ve read is from a product designer or a startup person treating it like a Figma replacement. Which is fine, I guess, if you build dashboards for a living. But none of them are asking the question that matters for people like us: does this actually do anything for authors?

So I spent a week putting it through four real tests — a book cover, a pre-order landing page, ad creatives, and a book launch media kit. The results genuinely surprised me.

Let me show you what happened.

First, the limitation

Claude Design is not a book cover generator.

I want to lead with this because I’ve already seen people on Twitter implying it is, and that’s going to set authors up for disappointment. It does not generate photorealistic figures. It does not do the moody-thriller-figure-in-a-trench-coat thing. It is not Midjourney. It is not Ideogram.

It is a layout and design system tool. Think of it like a very fast, very competent designer who can build the structure of anything visual you describe — but who hands the final image generation off to someone else.

Once you accept that, this thing becomes really useful. Because covers are a layout problem too, and every other author marketing asset is purely a layout problem.

The test that changed my mind

The pre-order landing page.

I gave Claude Design a single prompt — hero section, two-line tagline, pre-order button, book description, three early reader blurb cards, author bio, email signup — and in about thirty seconds it built me a page that would have cost me four hundred dollars from a freelancer.

Hierarchy was right. Mobile responsive. The blurb cards even had placeholder content that read like real reader quotes, not Lorem Ipsum. I clicked one of them, used the inline comment feature, and just typed “make this feel more literary, less marketing-y.” It rewrote the card in place. While I watched.

And then I found the sliders.

This is the part nobody is covering. Claude Design generates these custom adjustment sliders for whatever you’re working on — for the landing page it gave me sliders for “atmospheric intensity,” “text density,” and “color saturation.” I dragged the atmospheric one up and the entire page got darker, the typography got moodier, the hero section gained a gradient. No re-prompting. No regeneration. Just live tuning.

That’s the difference between a tool you use once and a tool you keep open in a tab.

The cover workaround

Back to the limitation, because there’s a real workflow here.

Claude Design has a one-click export to Canva. The export isn’t a flat image — it lands in Canva as a fully editable file. So the workflow that ended up working for me is:

1. Generate a cover image with ChatGPT images 2 (or your favorite image generator, but this is the one I like right now).

2. Drop the image into the Claude Design chat prompt box.

3. Build the cover layout in Claude Design — typography, hierarchy, composition, using the image you attached.

4. Watch Claude Design build something beautiful for you.

About eight minutes from blank canvas to a working cover concept.

The launch media kit test

The fourth test was the one I think most authors will get the most use out of: a book launch media kit. Not a sponsorship kit — a press kit. The kind of thing you send to podcast hosts when you’re pitching for an interview, to bookstores when you’re asking for an event, to bloggers when you’re requesting a review.

Most authors I know either don’t have one of these, or they’re using a Word doc from 2019 that nobody wants to open.

What Claude Design built for me, in about ninety seconds, is genuinely better than the kit I sent out for my last launch. Clean hierarchy. A “discussion topics for podcast hosts” section that’s the thing that actually gets you booked on shows.

Total time, end to end: under fifteen minutes. PDF-ready.

What it’s best at

In rough order of how impressed I was:

— Pre-order landing pages (genuinely excellent) — Book launch media kits (the dark horse winner) — Ad creative variations (three coherent variants from one prompt, ninety seconds) — Lead magnet covers and one-pagers — Series design systems (the long-term play)

What it’s not for

Photorealistic book covers (use the Canva workaround). Anything requiring AI-generated figures. Highly stylized illustration.

How to try it

If you’re on a Claude Pro or Max plan, you already have it. Click the palette icon in the sidebar at claude.ai, or go straight to claude.ai/design.

The full breakdown is up on YouTube — I walk through all four tests live, including the slider thing, the Canva handoff for covers, and the launch media kit demo.

[Watch here →]

There’s an open-source Claude Code “clone” you can try for free, but it takes some setup and technical know-how. Check it out on GitHub: https://github.com/OpenCoworkAI/open-codesign

One more thing

Everything I built in those tests — the prompt library, the Canva handoff workflow, the SCRIPT framework — I packaged into the Claude Author Bundle on Ko-fi. It’s the full design guide plus three bonus guides: a publishing checklist, a 60+ prompt library for authors, and a pocket reference deck.

[Get the bundle →] https://ko-fi.com/s/c38d7ba38d

Talk soon,

Christopher

ChatGPT 5.5 — Sam Altman is calling it “Spud” internally, which is either charming or unhinged depending on your tolerance for tech CEOs naming things — dropped Wednesday. By Thursday afternoon I had it open in three browser tabs, fed it a 20,000-word manuscript, made it cry uncle on a voice match test, and learned more about what it can actually do for authors than I have from any model launch in the last eight months.

The full demo went up on YouTube last night. You can watch the head-to-head against Claude Opus 4.7 here:

https://youtu.be/8p8HL5gTQK8?si=nCQwHH4abO-cuIfh

But this newsletter is the thing the video doesn’t have room for: the actual prompts. Copy them. Paste them. Use the ones that fit your workflow and ignore the ones that don’t.

A few honest notes before we start: I’m not paid by OpenAI. I’m not paid by Anthropic. I tested Spud against Opus on four real author tasks and I’m going to tell you exactly where each one wins and loses. I have a Ko-fi store. If something I make is useful, you can grab it there. If not, no hard feelings. We’re all just trying to write the next book.

Let’s go.

What Spud is actually good at (the short version)

Long context. Plot holes. Continuity. Memory.

You can hand it a 20,000-word manuscript plus a 3,500-word series bible — that’s about 23,000 words of context, more than three short stories’ worth — and it will hold all of it at once. It will check Chapter 11 against the bible. It will catch the moment your left-handed character reaches for a cup with her right hand. It will flag the brother you mentioned in Chapter 9 before he was supposed to exist.

That is not a small thing. That is the workhorse capability I have been waiting for since I started using AI for fiction work two years ago.

What it is not as good at — and I’m going to spend a section on this so nobody buys the wrong thing — is voice. Specifically, distinctive first-person voice. If you write in a particular register, with particular tics, the model can write something competent in your direction. It cannot, yet, write something indistinguishable from you. Opus is still better at voice. I’ll show you the prompt I used to test it below.

Now: the prompts.

Prompt 1 — The continuity check (the one I use the most now)

If you write a series, or even just a long novel with a lot of moving parts, this is the prompt that earns Spud’s keep.

You are my developmental editor. I’m going to give you two documents: my series bible and my working manuscript. Read them both carefully. Then flag every continuity error, timeline break, and plot hole you find. Check manuscript details against the series bible where relevant. Don’t critique prose. Don’t suggest line edits. Only logic. Format your output as a numbered list of issues; for each, the chapter or page where it occurs, the specific contradiction, and the bible reference if applicable.

Then upload your bible first. Then your manuscript. Send the prompt.

I planted three errors in my test manuscript on purpose. Spud caught all three, including a sneaky one where I’d referenced a character’s brother eleven chapters before the letter that was supposed to introduce him. It even noted the page number where the bible established the timeline rule.

This is the prompt that turned my “I should clean up continuity sometime” Saturday into a forty-five minute task.

Prompt 2 — The scene punch-up (where instructions matter more than vibes)

You wrote a flat scene. We’ve all done it. Here’s the prompt that actually gets a usable rewrite instead of a generic AI gloss-up.

Below is a scene from my novel that needs revision. It currently reads flat. Here’s what I need you to do: Don’t change the POV. Don’t add new plot events. Keep the dialogue count roughly the same. Add physical tells — what hands are doing, where eyes land, what bodies betray. Build subtext: the character who knows the secret should sit on it; don’t have them volunteer information. The voice should feel like [insert author / register here — e.g., “Joe Abercrombie working-class fantasy, not Tolkien”]. Then return the rewrite as a single clean draft. After the draft, give me a one-paragraph note explaining what you changed and why. Here is the scene: [paste scene]

The reason this prompt works: it’s specific about what not to do. Most authors paste a flat scene and ask for “punch-up.” That gets you AI slop with extra adjectives. The list of constraints is the point. The model will follow them.

Tip: the comparison author at the end (in the brackets) matters more than you’d think. “Make it grittier” is vague. “Joe Abercrombie working-class fantasy” gives the model a target.

Prompt 3 — The series bible builder (this one’s free if you use it)

You don’t have a series bible. I know. Most authors don’t until they’re three books in and drowning. Here’s how to build one in an evening.

You are helping me build a series bible for an existing manuscript. I’m going to upload the manuscript. Read it carefully and extract a structured series bible covering: main characters (name, role, distinctive traits, key relationships, current arc); supporting characters (name, function, where they appear); world rules (magic system, technology, social structures, political bodies); locations (name, description, who is associated with each); timeline (events in chronological order with chapter references); and established facts (anything stated as canon — eye color, dates, family relationships, sworn oaths, etc.). Format as a clean reference document. Use headings. Include a “loose ends” section at the end for anything that seems unresolved or contradictory.

The first time you run this, the output will be 80% right and 20% missing or off. That’s fine. The point is you now have a draft bible instead of a blank page. Spend an hour cleaning it up. Save it. Re-use it forever.

Now run Prompt 1 against the bible you just made. Watch what surfaces.

Prompt 4 — The voice match (where Spud actually struggles)

I want to be honest about this because it’s the test that matters most for fiction. Voice. The thing that makes you you.

I gave Spud 4,000 words of a deliberately distinctive narrator — first person, sardonic, em-dash heavy, specific recurring vocabulary — and asked it to continue. It produced something competent. Readable. Not bad.

But not her. The em-dashes were there. The cadence wasn’t. It hit the surface tics and missed the rhythm underneath. Opus did better. Opus didn’t nail it either, but it got closer.

If you’re going to use Spud for voice work anyway, here’s the prompt that gets you the most out of it:

I’m going to paste a long passage in a very specific first-person voice. Read it carefully — twice — before you write anything. Study: sentence rhythm and length variance; recurring vocabulary and phrasings; how the narrator handles asides (parens vs em-dashes vs fragments); where the narrator lands jokes and where she pulls back; specific tics that show up more than once. Then write the next 500 words of the next chapter, picking up from where the passage ends. Match the voice exactly. Do not summarize. Do not explain what you’re doing. Do not shift register. Just write. Here is the passage: [paste 3,000-5,000 words minimum]

Two things to know:

Sample size matters. 1,000 words isn’t enough. The model needs to see the voice repeat itself before it can mimic anything but the most surface features. Give it 3,000 minimum. 5,000 is better.

Run it twice. The first output is usually too generic. Tell it specifically what it missed — “you dropped all the em-dash asides,” “the second paragraph sounds like a competent thriller writer, not the narrator” — and ask for a revision. The second pass is almost always better.

If you need this nailed cold, use Opus. If you’re doing voice-adjacent work where “close enough” is fine — synopsis writing, pitch letters, marketing copy in your voice — Spud is fine.

Prompt 5 — The website builder (the trick I learned this week)

This is the one I’m most excited about and it’s the prompt I’d never have figured out without the live demo.

If you ask ChatGPT to build you an author website with code, it’ll generate something. It’ll be functional. It’ll also look like a 2014 portfolio template.

Here’s the trick. Don’t start with code. Start with images.

I’m building an author website for [pen name / series]. Genre is [genre]. The vibe should be [3-5 mood words — e.g., “atmospheric, literary, slightly gothic, paper-and-ink, cold light”]. Before we write any code, I want you to design the site visually. Generate the following images: a homepage hero image (describe what should be in it); three book cover concepts for the first book in the series; an “about the author” image (mood/atmosphere, not a portrait — I’ll add my photo separately); a background texture for the site; an icon set (home, books, about, contact). Generate them now. After I approve them, we’ll build the HTML and CSS using these images as the design system.

The reason this works: most AI-generated websites look bad because the model has no design intent. It defaults to safe-and-generic. When you force it to commit to a visual world first, the code that comes after has something to live up to.

You can use this for more than websites. Book covers, social media kits, newsletter banners, Substack header images, Ko-fi store thumbnails. Generate the design system as images. Then build the assets that need to match it. The whole brand stays coherent because it was designed coherent from the start.

This was the prompt that genuinely surprised me during testing. I’d been using ChatGPT for three years for code-first work. I’d never thought to flip the order.

So which one should you use? Spud or Opus?

The honest answer:

Continuity, plot holes, long-context analysis, series bible work, structural editing: Spud. Use it.

Voice work, distinctive prose, anything where the writing has to sound like a specific human: Opus.

Both for less than $40/month if you’re serious: Why not. They’re not the same tool. They both cost less than one editor pass.

You don’t have to pick one team. I use both. Most working authors I know use both. The trick is knowing which one to open for which task.

Keep writing.

— Christopher

Every author who’s used AI for writing has had the same experience.

You spend 20 minutes crafting the perfect prompt. You hit enter. The AI generates 800 words of competent, clean, professionally structured prose that sounds exactly like nobody you’ve ever met. It’s writing that could have been produced by a moderately talented MFA graduate who’s never had a strong opinion about anything.

Then you spend the next two hours rewriting every sentence to make it sound like you.

This is the unsolved problem in AI-assisted writing. Not capability. Not speed. Voice.

I figured out a fix this week. I want to walk you through it because it works on every model I’ve tested — paid, free, local, cloud — and it takes about 90 seconds to set up.

I’m calling it the Voice Profile trick, and the moment I realized it actually worked was a little unsettling.

What Most Authors Do Wrong

The standard advice for getting AI to match your voice is to feed it samples and say “write like this.” This barely works. The AI mimics surface features — sentence length, maybe a couple of word choices — but the deeper patterns that make your writing yours don’t transfer.

Your voice isn’t your vocabulary. Your voice is the architecture underneath: which senses you default to, whether you name emotions or show them through action, whether you rush at moments of tension or slow down, what kinds of contradictions appear in your character interiority, the specific verbal tics you don’t notice until someone points them out.

Generic “match my style” prompts can’t extract that. They produce vibes, not architecture.

The fix is to make the AI do the analysis first, then use its own analysis as a permanent style guide.

The Trick

You feed the AI three paragraphs of your own writing — fiction, blog, essay, anything in your natural voice — and ask it to reverse-engineer your style across specific dimensions. Not “describe my voice.” Specific dimensions.

Here’s the prompt to use. Copy it.

Here are three paragraphs from my writing:

[PASTE 3 PARAGRAPHS]

Analyze my voice across these dimensions:

1. SENTENCE RHYTHM — average length, variation
   pattern, tendency toward fragments or
   complex constructions

2. SENSORY DEFAULT — which senses I lean on,
   which I neglect

3. EMOTIONAL APPROACH — do I name emotions
   directly or show them through physical action

4. DISTINCTIVE PATTERNS — verbal tics, favorite
   constructions, signature moves I might not
   notice myself

Then create a compact VOICE PROFILE I can paste
into any future writing prompt as a persistent
style guide. Format it as instructions another
AI could follow to mimic me.

The output will be a 200-300 word style guide. Save it as a text file. Paste it at the top of every future writing prompt.

That’s the entire trick.

Why It Works

The reason this works is counterintuitive: you’re not asking the AI to be creative. You’re asking it to be analytical, which AI is much better at. Then you’re using that analysis as a constraint for the creative work.

When the AI generates prose with your Voice Profile pasted at the top, it’s not trying to “sound like you” — a vague instruction. It’s following specific rules: short sentences in tense moments, taste-and-touch as default senses, emotions shown through hand movements not internal monologue, no adverbs after dialogue tags. Concrete instructions produce concrete output.

The compounding effect is what makes this powerful. You do the 90-second voice analysis once. You paste it into every prompt forever. Every piece of AI-assisted writing you produce inherits your architecture from the first word.

Note: The longer and more natural the writing example(s) you provide, the better the analysis.

The Part That Was Unsettling

I tested this with Gemma 4 — Google’s brand new free model that runs locally on your computer — for a video this week. I fed it three paragraphs from a blog post I wrote two years ago.

The Voice Profile it generated identified two patterns in my writing that I had never consciously noticed and that no human reader has ever pointed out:

I default to sound as my primary sensory channel before any other sense.

I tend to use single-sentence paragraphs at moments of emotional intensity, but only when the emotion is something I’m personally familiar with. Joy, grief, embarrassment — yes. Anger I haven’t experienced — no.

That second one is the kind of pattern only a writing professor (or therapist) would catch after months of close reading. Gemma 4 caught it from three paragraphs in 11 seconds.

When I tested the trick by feeding the Voice Profile back into a fresh prompt and asking for new prose, the output didn’t sound exactly like me. But it sounded like me much more than any AI output I’ve ever seen. The fingerprint was unmistakable.

That’s when I realized this was the missing piece. Not a better model. Not better prompts. Better self-knowledge about your own voice, extracted by the AI and fed back as a constraint.

Three Models Where This Works (One Free)

I’ve tested the Voice Profile trick across three models. Here’s what works:

Gemma 4 (free, local) — Free local model from Google. Surprisingly precise at voice analysis. Best option if you want privacy and zero cost. Install through Ollama (ollama run gemma4) and you’re running in five minutes. Full setup in my video below.

Claude (paid, cloud) — The most sophisticated voice analysis of the three. Catches subtler patterns. If you’re paying for one AI, this is where the Voice Profile trick produces the most useful output.

ChatGPT (paid, cloud) — Solid voice analysis, slightly less nuanced than Claude but excellent at producing the structured “voice profile as instructions” format.

For free use, Gemma 4 is genuinely competitive. For paid use, run the analysis on Claude once, save the profile, then use it everywhere.

What to Do Today

If you write with AI, even occasionally, do this:

Open whatever AI tool you use most. Find three paragraphs of your own writing — anything in your natural voice. Run the prompt above. Save the output as a text file called my_voice_profile.txt.

That file is now the most valuable writing asset you have. Paste it at the top of every prompt for the rest of your career.

Watch the Full Demo

I ran the Voice Profile trick on Gemma 4 along with four other tests in my new video. The voice cloning result is in there. So is the part where Gemma 4 listens to a 30-second audio clip of me doing a character voice and analyzes the subtext from the recording alone.

Watch the video:

The hardest problem in AI-assisted writing isn’t getting the AI to write. It’s getting the AI to write like you.

The Voice Profile trick is the closest I’ve come to solving it.

Happy writing.

Most “write faster with AI” advice is garbage.

It’s either generic prompt tips you’ve seen a hundred times, or it’s people pretending they wrote a novel in 24 hours when really they generated 80,000 words of unreadable AI sludge that nobody will ever read.

I just published a video that breaks this down honestly. There are actual speed levels — 5X, 10X, 20X, 50X, 100X — and each one requires a different approach. Not a different prompt. A different system.

Here’s the framework in brief:

5X (Prompt Bibles) is where most authors stop. You build a library of high-quality, tested prompts and stop reinventing the wheel every time you sit down to write. You can do this in Word, Google Docs, or anywhere. This alone gets you to 5X. Most people never get past this level.

10X (Prompt Systems) is where prompts start chaining together. Instead of one prompt at a time, you build sequences — premise to outline to scene beats to draft, all connected, with each step feeding the next. Output multiplies because you stop wasting time on context-switching.

20X (Automation) is where the work starts running without you typing. Tools fire prompts in sequence automatically, save outputs to folders, and notify you when tasks are done. You become a director, not a typist. You can automate with tools, python scripts, or platforms like Make, Zapier, N8N, etc.

50X (Agents) is where specialized AI agents handle research, structure, and drafting in parallel inside the same workspace. One agent researches while another outlines while another drafts. The bottleneck stops being you. Almost every major AI platform has an agent now. Open source agents exist, too. Think: OpenClaw.

Another option is Skywork AI (sponsored): https://skywork.ai/p/8Wo0lk
Code: writingsecrets

100X (Looping Pipelines) is where the entire workflow runs end-to-end with quality gates and stop conditions. You queue up an entire book project and walk away. When you come back, you have drafts, audits, and notifications waiting.

The catch — and this is what most “write a novel in a day” videos won’t tell you — is that speed without quality is worthless. Every level above 10X requires more upfront infrastructure and more honest editorial judgment on the back end. The AI doesn’t replace the author. It replaces the busywork.

I demo all five levels in the video, with prompt examples for each, plus a section on how to know which level is right for where you are right now. Most authors should start at 5X and earn their way up.

Watch the video:

Happy writing.

Christopher

Most authors are paying $20–$200/month for AI writing tools (gulp, that’s me).

Yet, I just built a complete workflow — premise development, outlining, chapter drafting, revision, and reader prediction — that costs exactly nothing. Not a free trial. Not a freemium tier. Actually free.

Here’s every free AI option authors should know about right now, and exactly where each one fits in your workflow.

The Layered Approach: Why One AI Isn’t Enough

The mistake most authors make is picking one AI and using it for everything. That’s like using a chef’s knife to open a can. The smarter approach is layering: use the right model for the right task.

Here’s how I think about it:

Creative work (premise, voice, prose quality): This is where paid models like Claude or ChatGPT still have an edge. They’re better at nuance, subtext, and maintaining a consistent voice across long outputs. If you’re paying for one AI, pay for the one you use for actual writing.

Structural work (outlines, scene beats, plot logic): Free models handle this almost as well as paid ones. Structural tasks have clearer right/wrong answers, which plays to the strengths of smaller models. This is where free AI earns its place.

Grunt work (continuity checking, formatting, brainstorming variations, research): Free models are perfect here. You don’t need GPT-4 intelligence to check whether your character’s eye color changed between chapters 3 and 7. Save your paid tokens for creative work.

Uncensored work (action scenes, intimacy, dark content): Free local models are the only option that works without fighting content filters. Cloud models — paid or free — will resist writing your villain’s internal monologue or your thriller’s graphic fight scene. Local models just write it.

The Free Models Worth Your Time (April 2026)

In my opinion, here are some free AI models you should try.

Llama 4 Scout (via Ollama — free, local)

Meta’s latest. 109 billion total parameters, but it only activates 17 billion at a time, so it actually runs on consumer hardware. I just put it through my full author scorecard — premise, outline, chapter writing, revision — and it scored an S on outlining and a solid B on prose. For a model that costs nothing and runs on your laptop, that’s remarkable.

Best for: Outlining, structural analysis, revision diagnostics, and brainstorming.

I did a full demo:

Gemma 4 (brand new — dropped this week)

Google just released Gemma 4, built on the same technology as Gemini 3. Four sizes: E2B and E4B for lightweight devices, 26B mixture-of-experts, and 31B dense. The 31B model currently ranks #3 among open models globally. It runs on Ollama right now — just type ollama run gemma4 in your terminal and you’re in.

What makes Gemma 4 interesting for authors: it has a native “thinking mode” for reasoning tasks, supports 140+ languages, and handles 256K context (that’s roughly 190,000 words — more than most novels). The 26B model activates only 4 billion parameters during inference, so it runs fast on modest hardware.

I haven’t done a full scorecard test yet, but this is at the top of my list. If you beat me to it, let me know what you find.

Best for: Worth testing for everything — it’s legitimately competitive with much larger models.

Meta AI (cloud — free, no account needed)

If you don’t want to install anything, Meta AI at meta.ai runs Llama 4 in the cloud for free. Just open the website and start prompting. No API key, no account, no credit card. The interface is basic but functional.

Best for: Quick premise testing, brainstorming, and anyone who doesn’t want to touch a terminal.

Ollama + API (free local API)

This is the power move. When Ollama is running on your computer, it exposes a free API at localhost:11434. That means any automation tool — n8n, Make, Zapier — can send prompts to your local model programmatically. You can build an entire automated chapter-drafting pipeline that runs overnight and costs zero dollars.

Best for: Authors building automated workflows, batch processing, and anyone who wants to chain prompts without copy-pasting.

Setting Up Ollama in 5 Minutes

This is genuinely all you need:

1. Go to ollama.com and download the installer for your OS

2. Open your terminal (Mac/Linux) or PowerShell (Windows)

3. Type: ollama run llama4:scout (or ollama run gemma4 for the brand new Google model)

4. Wait for the download (one time only)

5. Start typing prompts

That’s it. No API key, no account, no configuration. Your first prompt will work in under 5 minutes from a cold start.

If you want a ChatGPT-style visual interface instead of the terminal, install Open WebUI with Docker and point it at your local Ollama instance. Same free models, prettier interface.

That’s similar to what my Micro LM Studio does, plus some extra stuff just for authors.

Where Free Falls Short (Honest Take)

I’m not going to pretend free models replace paid ones for everything. Here’s where I still reach for Claude or Gemini:

Long-context creative coherence. When I need a model to maintain a character’s voice and emotional arc across a 3,000-word scene with complex subtext, paid models are still noticeably better. Free models tend to drift or flatten.

Revision sophistication. Paid models catch subtler craft problems — not just filter words and show-don’t-tell violations, but structural pacing issues and thematic inconsistencies. Free models catch the obvious stuff.

Instruction following under complexity. Give a free model 4 constraints and it nails them. Give it 8, and it starts dropping instructions. Paid models handle more complex prompt architectures.

The solution isn’t free OR paid. It’s knowing which tasks justify paying and which don’t.

Three Prompts to Test Right Now (Free)

The Premise Pressure Test — paste your premise into any free model and add:

“You are a brutal developmental editor. Score this premise 1-10 on originality, then identify the three most similar published novels. If the score is below 7, tell me exactly what needs to change.”

The Continuity Scan — paste two consecutive chapters and add:

“Check these chapters for continuity errors: character descriptions, time of day, weather, character knowledge states, emotional consistency, and proper noun spelling. Flag every conflict with specific passages.”

The Radical Cut — paste any passage and add:

“Cut this by 30%. Remove throat-clearing sentences, furniture sentences (stage direction that adds nothing), and redundant attribution. Show the cut version, then list everything removed and why.”

All three work beautifully on free local models. Save your paid tokens for the creative heavy lifting.

My Toolkit of FREE

If you want pre-built prompt systems, automation tools, and author workflows designed specifically for these models:

🎬 New Video: I just tested Ollama through my full author scorecard — https://youtu.be/T0O3j3N8vk8

🔧 AuthorClaw — 100% free open source OpenClaw for authors: https://github.com/Ckokoski/authorclaw

🐝 AuthorSwarm — (FREE while in beta) This is a 500,000 reader prediction engine (also works with free APIs): https://ko-fi.com/s/60f16cbb8b

🛠️ Micro LM Studio — Not technically free, but this is my author-optimized local model setup with automated author workflows: https://ko-fi.com/s/34fd4ab3ae

The best AI writing tool is the one you actually use. When half the stack is free, the barrier isn’t cost anymore — it’s knowing how to use it.

Happy writing.

If you’ve been anywhere near the AI space in the past two weeks, you’ve probably heard about MiroFish.

A college student in China built it in 10 days. It hit #1 on GitHub. It raised $4 million in 24 hours. And it does something genuinely different from every AI tool you’ve used before.

Instead of asking one AI model for an answer, MiroFish creates thousands of AI agents — each with their own personality, memories, and opinions — and drops them into a simulated social media environment. They post, argue, form coalitions, change their minds, and create emergent behavior that no single AI prompt could predict.

It’s called swarm intelligence, and it’s one of the most interesting developments in AI this year.

But here’s the thing: MiroFish is a general-purpose tool. It was designed to predict market sentiment, public opinion shifts, and policy reactions. Nobody was building this for writers.

So I built one.

What Swarm Intelligence Actually Means

Traditional AI gives you one perspective. You prompt ChatGPT, you get one answer. Even if you ask it to “roleplay as 10 different editors,” it’s still one model generating one output stream. The responses tend to converge. They sound the same. They’re polite. They agree with each other.

Swarm intelligence is fundamentally different. You create hundreds or thousands of distinct agents, each with its own:

• Personality traits (critical level, emotional reactivity, openness to new ideas)

• Preferences (genre loves, trope hates, platform habits)

• Communication style (a BookTok reviewer sounds nothing like a Reddit poster)

• Social behavior (some pile on popular opinions, some push back against the crowd)

Then you let them loose. You don’t script their reactions. You give them a stimulus — in our case, a book description — and watch what emerges from the collective.

The results are messier than a single AI response. They’re also far more realistic.

Why This Matters for Authors

Every author has asked some version of these questions:

“Would readers actually like this concept?” “What would BookTok say about my book?” “Would this premise get 1-star reviews? For what reason?” “Is this controversial enough to go viral, or just controversial enough to tank?”

Beta readers can answer these — eventually. But they’re slow, expensive, and you can only get 5-10 of them. A focus group is even harder to organize.

What if you could simulate 1,000 or even 10,000 readers reacting to your book concept in a few hours? Not one AI pretending to be many readers, but a genuine swarm of distinct personas running independently and then interacting with each other’s reactions?

That’s what I built with AuthorSwarm.

How It Works (Simply)

1. You write a book description — title, genre, premise, tropes, comp titles, tone

2. AuthorSwarm loads 1,000-10,000 reader personas from a pre-built database

3. Each reader reacts independently — star rating, social media post in their platform’s voice, emotional reaction, DNF decision

4. The most viral/controversial posts from Round 1 are shown to other readers

5. Readers respond to each other — agreeing, arguing, piling on, changing their minds

6. A prediction report is generated with ratings, platform breakdowns, controversy analysis, and a viral potential score

The reader personas aren’t random. They’re calibrated to real reader demographics: 30% BookTok (young, emotional, trend-driven), 25% Goodreads (analytical, review-focused), 15% Reddit (skeptical, discussion-oriented), 15% Bookstagram (aesthetic, mood-based), 10% X/Twitter (hot takes), and 5% lurkers (rate but never review).

Each reader has trope loves and hates, a critical level from 1-10, a DNF threshold, and platform-specific voice patterns. A BookTok reader writes in ALL CAPS with emojis. A Goodreads reviewer writes measured paragraphs comparing your book to others in the genre. A Reddit poster starts with “Am I the only one who thinks...”

What I Learned From My Test

I ran AuthorSwarm against one of my own book ideas. I won’t spoil the full results (the video drops Saturday), but here’s what surprised me:

The swarm didn’t just rate the book. It predicted specific controversy points I hadn’t considered. It identified which trope would divide readers. It showed me that BookTok would love the concept, but Reddit would tear apart the premise. And the simulated social feed — actual posts these AI readers would write — was eerily realistic.

The harshest 1-star review was genuinely funny. And genuinely useful.

The Bigger Picture

Swarm intelligence isn’t just a novelty. It represents a shift in how we use AI: from asking one model for the answer to simulating many agents and observing what emerges.

For authors, this means:

• Test book concepts before spending months writing

• A/B test different blurbs to see which resonates more

• Identify controversy points and potential backlash before publishing

• Understand platform-specific reactions (what works on BookTok vs. Goodreads)

• Get the brutal 1-star perspective without waiting for real negative reviews

This isn’t a replacement for real beta readers or real market data. It’s a new tool in the toolkit. One that runs in hours instead of weeks and gives you a perspective no single AI prompt can match.

Try It Yourself

AuthorSwarm is available now on my Ko-fi. It runs on your computer, supports free local AI through Ollama, and includes a premium prediction report.

Try the tool: https://ko-fi.com/s/60f16cbb8b

Full demo video on my YouTube channel:

P.S. If you’re curious about swarm intelligence beyond the author use case, the original MiroFish project is open source on GitHub. Someone plugged it into a prediction market bot and reportedly made $4,266 in profit over 338 trades. The technology is real and evolving fast.

You know that folder on your computer?

The one with fifty files. Chapter drafts named “ch3_FINAL_v2_actually_final_this_time.docx.” Random notes you wrote at midnight that say things like “what if the cat is a spy” with zero context. An outline from three months ago that contradicts the outline from last month. Research you’ll definitely use someday.

That folder.

I pointed an AI at mine and told it to fix everything.

Not ChatGPT. Not the regular Claude chat window. A mode called Cowork that most authors haven’t discovered yet — and after a week with it, I think it’s the most practical AI update for writers since context windows got big enough to hold a full manuscript.

I ran it through six real tests. Here’s what actually happened.

What Claude Cowork Actually Is

Quick version: Cowork is a separate mode inside the Claude desktop app. Instead of chatting back and forth, you point it at a folder on your computer and tell it what to do. It reads your files, creates new ones, organizes things, and delivers finished work — actual documents, not chat responses you have to copy and paste.

Think of regular Claude as the smart friend who gives great advice over coffee. Cowork is the friend who shows up with a truck and actually helps you move.

You need the desktop app (free download, Mac and Windows) and a Pro subscription ($20/month).

Test 1: Organize the Chaos

I gave Cowork a realistic author mess — chapter drafts, scattered notes, research docs, contradicting outlines, a file literally named “what if the cat is a spy.txt” — and told it to organize everything into a clean project structure.

It didn’t just alphabetize by filename. It read the contents of every file to figure out what each one actually was. The spy cat note? Filed under character concepts. The two contradicting outlines? Flagged and grouped together so I could reconcile them.

It generated a README summarizing the entire project — like a table of contents for my creative disaster.

The takeaway: If your project folder gives you anxiety every time you open it, this is a 5-minute fix.

Test 2: Auto-Generate a Book Bible

A book bible tracks every character, location, timeline event, world rule, and relationship in your story. It’s the thing that stops you from giving your protagonist blue eyes in chapter one and brown eyes in chapter twelve.

The problem: building one manually is incredibly tedious. You have to extract every detail from your entire manuscript. Nobody does it until it’s too late.

I pointed Cowork at my chapters and asked for the full package — character profiles, location guide, timeline, world rules, relationship map, and open plot threads.

The magic: Cowork reads your entire folder simultaneously. When a character’s backstory gets mentioned in chapter two but doesn’t become relevant until chapter eight, it catches that connection. A detail buried in a random notes file? Cross-referenced automatically.

Going from zero to a solid first-draft book bible in minutes instead of days isn’t a small thing. It’s the difference between having one and not having one at all.

Test 3: Continuity Checker

This one surprised me most.

I asked Cowork to scan all chapters for contradictions — physical descriptions that change, timeline impossibilities, characters acting on information they shouldn’t have, world rules that get broken.

It caught things I missed across three manual read-throughs.

Not just surface-level stuff like inconsistent hair color. It found logical contradictions — moments where two scenes don’t make sense side by side even though neither one breaks a specific rule on its own. It caught a timeline error where a character gets assigned to a case the day before the victim dies.

This isn’t replacing a human editor. But as a first pass before you send your manuscript out? It could save you an entire round of expensive revisions.

Test 4: Developmental Editor

A professional developmental edit — the kind where someone evaluates your pacing, structure, character arcs, and thematic coherence — costs $2,000 to $5,000 and takes 4 to 8 weeks.

Most authors skip it entirely.

I asked Cowork for a full report: where the story drags, which chapters could be cut or reordered, where character arcs stall, where tension drops, and what’s actually working that I should do more of.

Honest take: this will not replace a great human developmental editor. The instinct, the industry knowledge, the gut-level story sense — AI doesn’t have that yet.

What it does give you is a diagnostic. You walk into the professional edit already knowing where the problems are. That makes the whole process faster, more focused, and worth every dollar.

Test 5: Story Visualization

This one was pure fun.

I asked Cowork to generate charts: pacing tension across chapters, word count distribution, dialogue-to-narrative ratio, character appearance frequency, and an emotional arc graph.

Being able to see your story as data is weirdly powerful. You can literally spot where the middle sags in a line graph. You can see which chapters are too dialogue-heavy. You can identify the exact point where your protagonist drops off the page for too long.

Could you build these charts manually in a spreadsheet? Sure. Would you ever actually do it? We both know the answer.

Test 6: Research + Worldbuilding

The wildcard. Cowork can connect to the Claude Chrome extension, which means it can browse the web and work with your local files in the same task.

I asked it to research a historical setting online, then combine what it found with my existing worldbuilding notes into a structured setting guide.

The result: a comprehensive document organized by category, cross-referenced with my notes, with flags where my existing worldbuilding contradicted historical reality. For historical fiction, sci-fi, or anything research-heavy — this combo is legitimately useful.

8 Cowork Prompts for Authors (Steal These)

These go beyond the basics. Copy them, adapt them, make them yours.

1. The Manuscript Diagnostic (Quick)

Read every file in this folder. Give me a brutally honest 
diagnostic of this manuscript:

1. Where does the story's energy peak and flatline?
2. Which character feels most alive? Which feels like furniture?
3. What's the single biggest structural problem?
4. What am I doing well that I should do MORE of?

Be specific. Reference chapters and passages.

2. The Ghost Reader

Read every chapter in order as a first-time reader with no 
context. After each chapter, write one sentence capturing 
your honest emotional reaction as a reader.

At the end, answer:
- Where would a casual reader stop reading? Why?
- What scene would make someone recommend this book?
- What question were you most desperate to have answered?

3. Voice Extraction Skill

Read every chapter. Analyze my writing voice and create a 
SKILL.md file capturing: sentence rhythm, dialogue style, 
overused words, distinctly "me" phrases, description approach, 
what I do at my best vs. when I'm coasting.

Format as a Claude skill file with YAML frontmatter.

4. The “Canon Cop”

Read the book-bible.md and all chapter files. Flag every 
place where chapters contradict established canon. 
Distinguish between errors (unintentional) and evolution 
(changes I should update the bible to reflect).

5. The Pacing X-Ray

Read the full manuscript. For each chapter, rate tension 
on a 1-10 scale. Then identify: the longest stretch without 
a turning point, any chapter where tension drops more than 
3 points from the previous chapter, and the exact moment 
where the midpoint should hit. Save as pacing-report.md.

6. Character Relationship Tracker

Read all chapters. Map every meaningful interaction between 
characters. Track how each relationship evolves chapter by 
chapter. Flag any relationship that stalls for more than 
3 chapters without development. Flag characters who exist 
in the same story but never interact — is that intentional?

7. The Comp Title Researcher

Using Chrome, research the top 10 bestselling books in 
[YOUR GENRE] from the past 2 years. For each, find: title, 
author, premise, what readers loved, what they criticized. 
Compare these patterns to my manuscript. Where does my book 
align with market trends? Where does it diverge?

8. The “What’s Missing” Audit

Read every file in this folder. Based on the story so far, 
identify: scenes that are referenced but never written, 
character introductions that are promised but missing, 
subplots that start but go nowhere, and any setup in early 
chapters that lacks a payoff. Present as a checklist I can 
work through.

The Video

I tested all six use cases on camera — live, with real screen recordings. If you want to see what Cowork actually looks like in action (the file reorganization alone is satisfying to watch), here’s the full breakdown:

👉 Claude Cowork is a CHEAT CODE for Authors — Watch it here

Under 15 minutes. No filler.

One More Thing

I put together a Cowork Edition of my Claude AI Book Blueprint with all six use cases, folder structures, prompt templates, and workflows packaged up and ready to go. If you want the whole system instead of piecing it together yourself, it’s on my Ko-fi store: [Claude AI Book Blueprint]

OpenAI just dropped ChatGPT 5.3 Instant. The headline feature? Less “cringe.” Fewer refusals. More direct answers.

I tested it. For authors, the results are...complicated.

I’ll get to my full demo video in a minute. But first, I want to talk about a problem that 5.3 Instant claims to fix—and doesn’t. Not really. Not where it matters for us.

The Problem That Won’t Die

You’re writing a thriller. Your protagonist gets stabbed in an alley. You need the scene to feel visceral, desperate, real. You ask ChatGPT for help.

What you get: “The blade found its mark. Pain blossomed across his torso.”

Pain blossomed. Like a flower. In a stabbing scene.

Or you’re writing a romance. Two characters who’ve been circling each other for 200 pages finally give in. You need heat. Tension. Physicality.

What you get: “Their connection deepened as the evening progressed.”

Thanks. Very steamy.

ChatGPT 5.3 Instant is genuinely better at not lecturing you before it answers a question. It reduced hallucinations by almost 27%. The conversational flow is noticeably improved. These are real upgrades.

But when you need AI to help you write the scenes that actually make readers feel something—violence, intimacy, fear, grief at full intensity—the guardrails haven’t meaningfully moved.

It’s not just an OpenAI problem:

And I get it. These companies are building products for hundreds of millions of people. They can’t fine-tune safety filters for every use case. A thriller author asking about stab wound mechanics looks identical to something genuinely concerning if you’re a pattern-matching algorithm with no context.

But here’s the thing: that’s a solvable problem.

The Case for Author Validation

Imagine if ChatGPT, Claude, or Gemini had an “Author Mode” or a “Creative Professional” context layer. Not a jailbreak. Not removing safety entirely. Just a system that understands:

This person is writing chapter 14 of a crime novel. They have 80,000 words of manuscript context. They’re asking about blood spatter patterns because their detective character needs to analyze a crime scene. This is research, not intent.

We already have context windows big enough to hold entire novels. We already have models smart enough to understand narrative structure, character arcs, genre conventions. The technology to distinguish “author researching for a story” from “person with harmful intent” exists right now.

The same logic applies to academic researchers. A history professor studying wartime atrocities, a criminology student analyzing case studies, a medical writer describing surgical procedures—these are all legitimate professionals whose work requires engaging with difficult subject matter. The models are smart enough to make these distinctions. The companies just haven’t prioritized building the validation layer.

What would this look like in practice?

Something like: you upload your manuscript or project outline. The model reads your genre, your characters, your plot. Now when you ask, “How would this character realistically experience a knife wound,” the AI has context. It knows you’re building a scene, not planning one. It can give you the gritty, specific, useful answer a professional author actually needs.

Until that exists, authors are stuck choosing between two bad options: wrestle with safety filters that treat your murder mystery like an actual murder plot, or go find a less powerful model that won’t judge you.

Most authors choose option two.

The Local Model Solution (Right Now, Today)

While we wait for the big AI companies to build author-aware context systems, local uncensored models are the practical answer. And they’ve gotten dramatically easier to use.

LM Studio is the tool I recommend. It’s free, it runs on your computer, and it gives you a ChatGPT-style interface for local models—no code, no terminal commands, no technical headaches.

Here’s what makes it work for authors:

Zero filters. Ask it to write a graphic fight scene. It writes a graphic fight scene. Ask it to describe a character’s injuries in clinical detail for your medical thriller. It just does it. No lectures, no disclaimers, no watered-down output.

Complete privacy. Your manuscript never leaves your machine. No training data concerns, no server logs, no “we may use your conversations to improve our models” fine print. For authors working on unpublished material, this matters.

Surprisingly capable models. The latest uncensored local models—Dolphin 3.0, Nous Hermes 3, the abliterated Qwen and Llama variants—are legitimately good at creative writing. They’re not ChatGPT or Claude quality for complex reasoning, but for raw prose generation with no content restrictions? They punch way above their weight.

The tradeoff is real, though. Local models are less intelligent. They have smaller context windows. They hallucinate more. They can’t do the deep structural analysis that a model like Claude or ChatGPT 5.3 excels at. You’re trading raw power for creative freedom.

Which is exactly why the ideal setup is both: use the big models for the heavy lifting—premise development, structural analysis, revision feedback, plot hole detection—and use your local uncensored model for the scenes where you need the AI to stop flinching.

This is what it might look like in production:

My Micro LM Studio Setup (Built for Authors)

I built a lightweight, author-focused LM Studio configuration that I use alongside my main AI workflows. I call it the Micro LM Studio setup, and it’s designed specifically for writers who need:

• A dead-simple local model running for scene work, dialogue, and anything the big models refuse to touch

• Pre-configured settings optimized for creative writing (temperature, context length, system prompts tuned for fiction)

• Integration with my Story Streams workflow so you can bounce between your main AI and your local model without breaking your creative flow

It’s available on my Ko-fi store. No subscriptions, no recurring fees. Download it, set it up once, and you’ve got an uncensored writing partner sitting right on your desktop whenever you need it.

ChatGPT 5.3 Instant: The Actual Test

Now—the video. I put ChatGPT 5.3 Instant through my standard author gauntlet: premise generation, outlining, chapter writing, and revision for both fiction and nonfiction.

The verdict? It’s a genuine improvement in some areas and still frustrating in others.

I break down every test, show you the actual outputs, and give you my honest take on where 5.3 Instant earns its place in your workflow—and where it still falls short.

Watch: Testing ChatGPT 5.3 Instant for Authors (Full Demo)

No fluff. No filler. Just the tests, the results, and what they mean for your writing.

The Bottom Line

ChatGPT 5.3 Instant is less annoying. That’s real progress. But for authors who write in genres that require intensity—thriller, horror, romance, literary fiction with teeth—the core censorship problem hasn’t been solved.

The solution right now is a two-tool approach: big models for brains, local models for bravery. And until OpenAI, Anthropic, or Google builds genuine author context awareness into their safety systems, that’s the setup that actually works.

The technology to fix this already exists. Someone just needs to build it.

Happy writing.

P.S. — The Micro LM Studio setup is on my Ko-fi store.

OpenAI dropped ChatGPT 5.4 Thinking last night, and I immediately did what any reasonable person would do: I asked it to do everything at once.

One prompt. No follow-ups. No hand-holding.

I told it to generate a complex, commercially viable literary novel premise — then expand that into a full outline with chapter and scene breakdowns, write the five most pivotal pillar scenes, create a complete book bible, build a promotional website, and draft a 30-day marketing plan to sell a million copies.

Unfair? Absolutely. But that’s kind of the point. You learn more about a model by pushing it past its comfort zone than by lobbing softballs at it.

Here’s what happened — and what it means for your writing.

Watch the video instead:

The Premise

ChatGPT 5.4 generated a literary gothic thriller called The Hollow Inheritance — an obituary writer discovers that strangers are dying exactly as she imagined them in an unpublished novel she wrote and abandoned twenty years earlier. A killer using an abandoned manuscript as a murder map.

It positioned the book at the intersection of literary prestige and book club commercial appeal, which tells me something: this model is thinking about marketability, not just creativity. That’s a shift.

The Outline & Scenes

The full outline came with dual POV structure (first person present + archival voice past timeline), chapter-level breakdowns, a central secret driving the plot, and a scene-by-scene emotional arc. That emotional arc detail is genuinely useful — it’s the kind of structural planning that separates a draft from a manuscript.

The five pillar scenes were competent. Not breathtaking prose, but solid scaffolding. Exactly the kind of thing you’d rewrite in your own voice, but be glad you didn’t have to invent from scratch.

The Book Bible

Symbols, motifs, setting bible, character secrets, possible chapter titles — all present, though each section could’ve gone deeper. This is where asking for everything at once costs some depth. One task at a time, this would’ve been significantly more fleshed out.

The Website (The Surprise)

This is what caught me off guard. ChatGPT 5.4 built a fully functional promotional website — hero section, synopsis, reading excerpts, book club discussion guide, newsletter signup, retail links, navigation that actually works. It thought fast, built fast, and the result was actually usable right out of the box.

The Verdict

ChatGPT 5.4 Thinking is strong for authors — particularly around planning, revision, structural analysis, and marketing assets. The thinking model adds a layer of deliberation that shows up in how it organizes complex requests. It’s not a magic novel-writing machine. But as a creative partner for the architecture of storytelling? It punches above its weight.

Watch the full demo.

Advanced Prompts to Try With ChatGPT 5.4 Thinking

These go way beyond “write me a story about a detective.” These are designed to leverage the thinking model’s deliberation and push it into territory where most people never take it.

1. The Premise Pressure Test

Generate a novel premise that meets ALL of the following criteria simultaneously:

- Commercially viable for both literary fiction readers and book club audiences
- Contains a central mystery or question that cannot be resolved through a single revelation
- Features a protagonist whose profession directly creates ironic tension with the plot
- Includes at least one structural innovation (non-linear timeline, unreliable narrator, dual POV, epistolary elements, etc.)
- Has a "positioning sentence" — one line that could appear on a book jacket and make a reader pick it up immediately

After generating the premise, critique it yourself. Identify the weakest element and offer an alternative that strengthens it.

Why this works: You’re forcing the model to satisfy multiple competing constraints at once — commercial appeal vs. literary ambition, structural complexity vs. accessibility. The self-critique step exploits the thinking model’s ability to evaluate its own output before you even have to.

2. The Emotional Architecture Mapper

I'm writing a [GENRE] novel with the following premise: [YOUR PREMISE]

Create a scene-by-scene emotional arc for the full novel. For each major scene, provide:

- The dominant emotion the reader should feel
- The secondary emotion creating tension beneath the surface
- The specific narrative technique driving that emotional response (dramatic irony, withholding, reversal, sensory immersion, pacing shift, etc.)
- A one-sentence "emotional objective" for the scene — what the reader should feel differently about by the end of it

Map these across a three-act structure and flag any emotional dead zones where reader engagement is likely to drop.

Why this works: Most people ask AI for outlines. This asks for the emotional engineering underneath the outline — the invisible architecture that determines whether a reader keeps turning pages. The thinking model excels here because it has to reason about cause and effect across the full narrative.

3. The Manuscript Autopsy

I'm going to paste [a chapter / the first 5,000 words / a pivotal scene] from my manuscript. Analyze it across these dimensions simultaneously:

VOICE: Is the narrative voice consistent? Where does it slip into generic phrasing vs. where does it feel distinctive? Identify the 3 strongest voice moments and 3 weakest.

PACING: Map the pacing rhythm. Where does it accelerate, decelerate, or stall? Are there sections where the prose is doing work the reader doesn't need?

SUBTEXT: Where is the dialogue carrying subtext effectively vs. where are characters saying exactly what they mean? Identify missed opportunities for tension between what's said and what's meant.

SENSORY WEIGHT: Rate each scene on sensory immersion (1-10). Where am I telling when I could be grounding the reader in physical experience?

Deliver this as a diagnostic report with specific line-level examples, not general advice.

Why this works: You’re turning the AI into an analytical reader, not a writing partner. The multi-dimensional analysis forces the thinking model to hold several lenses simultaneously, and the demand for specific line-level examples prevents it from retreating into vague praise.

4. The Reverse-Engineered Bestseller

Analyze the narrative structure, prose style, and commercial positioning of [SUCCESSFUL BOOK IN YOUR GENRE]. Now use that analysis as a structural template — NOT to copy the story, but to reverse-engineer the principles that made it work.

Apply those principles to my premise: [YOUR PREMISE]

Generate a full outline that uses the same structural rhythms (pacing of reveals, character arc timing, subplot weaving pattern) but with entirely original content. Explain what you borrowed structurally and why.

Why this works: This is how professional authors actually study craft — not by copying, but by extracting transferable structural principles. The thinking model’s deliberation helps it separate what a book did from how it did it, giving you a framework instead of a knockoff.

5. The Character Contradiction Engine

Create a character profile for [CHARACTER NAME/ROLE] that is built entirely around contradictions:

- A public persona that directly conflicts with a private truth
- A core belief that their own actions consistently undermine
- A skill or strength that is simultaneously their greatest vulnerability
- A relationship where they are both the protector and the threat
- A secret that, if revealed, would recontextualize every previous scene they appear in

For each contradiction, explain how it creates narrative tension and at what point in a three-act structure it should be revealed or exploited for maximum impact.

Why this works: Flat characters come from single-axis descriptions. This prompt forces multi-dimensional construction where every trait contains its own tension. The thinking model has to reason through how contradictions interact across a full narrative arc, not just list attributes.

6. The “Kill Your Darlings” Editor

I'm going to paste a chapter from my novel. Your job is to cut it by 30% without losing any essential plot information, character development, or emotional beats.

Rules:
- Every cut must be justified with a specific reason (redundancy, telling-not-showing, pacing drag, information the reader can infer, etc.)
- Identify sentences that are "furniture" — they exist to move characters around a room but add nothing to the scene
- Flag any passage where I'm explaining an emotion the reader already felt from the action
- Preserve my voice — cuts should make the writing sound MORE like me, not less

Show the edited version, then show a summary of what was cut and why.

Why this works: Revision is where most AI-assisted writing falls apart because people ask for “improvements” instead of giving specific editorial constraints. The 30% cut target and the explicit rules force the model into the kind of ruthless editing that actually makes prose stronger.

These prompts leverage a thinking model to handle the heavy analytical and structural lifting that used to take weeks of revision and beta-reader feedback. Use them as starting points, modify them for your project, and always bring your own voice to whatever comes back.

If you want the full ChatGPT 5.4 Power Guide for Authors — with workflows, revision systems, packaging tactics, prompt formulas, and more — grab it for FREE on my Ko-fi store. If you already purchased the ChatGPT Book Blueprint, just re-download through your original purchase link for the updated version.