New AI models seem to launch every few weeks now, so this course won't tell you which one to use this week. It teaches you how to choose for editing, marketing, cover art and drafting, how to weigh cost against the time you actually get back, and how to protect your rights along the way.
Start with these three, and come back for the rest later:
When an author says "I use ChatGPT" or "I use Claude," they're squashing three separate choices into one word. Pulling them apart is the first step to choosing well. The names on each layer will keep changing. The layers themselves won't.
These layers don't always travel together. Some author writing apps let you pick which company's model runs underneath. And the same model can feel different on two platforms, because each one wraps it in its own instructions and features.
Use AI only where you want it
Nothing here assumes you want AI anywhere near your prose. Plenty of authors use it only for editing checks, marketing or admin, and keep every sentence of the book their own. The method works the same whichever jobs you choose to hand over, including none of the creative ones.
Some authors avoid AI altogether over how these models were trained, often on published books used without permission, and the lawsuits that followed. That's a legitimate reason to opt out. If you do use AI, how a company sources and licenses its training data can be one of your platform criteria (Module 8).
Almost every AI company sells its models in three rough sizes. The product names change constantly, but the three roles have held steady for years:
| Size | What it's like | Good author jobs for it |
|---|---|---|
| Flagship | The smartest, slowest and most expensive. It holds nuance, long style guides and complicated reasoning. | Big-picture edits, plot and continuity problems, your blurb, and drafting if you do it. |
| Workhorse | Close to flagship quality, much cheaper and faster. It's the right default most of the time. | Line edits, newsletters, ad copy, research summaries, brainstorming. |
| Fast | Small, cheap and nearly instant. Good at clear, simple instructions, and poor at taste. | Typo sweeps, formatting, dozens of ad-hook variations, sorting reader email. |
A fourth group sits outside the ladder. Specialist models are built for one medium, like images, voice or video. They aren't bigger or smaller than a text model. They're a different tool.
The sizes are a starting guess, not a rule
One company's workhorse sometimes beats another company's flagship at a specific job, and "thinking" modes (which make a model reason step by step before answering) blur the lines further. Use the sizes to decide what to test first. Your own test decides the rest (Module 9).
Run any task through these five questions. They narrow the field faster than any leaderboard, and they'll still work after the next launch.
A social caption can be merely good. Your blurb sells the book, so it needs to be excellent. The higher the bar, the bigger the model.
A one-off job can afford a flagship. Something you'll run hundreds of times should use the cheapest size that still clears your bar.
A blurb is tiny. A 100,000-word manuscript plus a series bible is not. For whole-book work, you need a model that handles long documents well: one that still remembers chapter 12 when it reaches chapter 30, not one that merely accepts the upload.
The medium often decides the kind of tool before size even comes into it.
Can the platform train on your unpublished manuscript? Can you sell what it makes? Will you need to disclose AI use when you publish? This one is big enough to get its own module (Module 6).
Check your brief before you blame the model
The instructions you give often matter as much as the model you pick. A clear brief, your style sheet, a voice sample and one or two examples of what "good" looks like can lift a workhorse above a flagship that got a one-line request. Before you switch tools over a disappointing result, fix the brief and try again.
Each part of an author's workflow asks for something different. Here's what to prioritize and how to test for it, no matter which tools are on top this year.
"Editing" is really three jobs, and each suits a different size. Match the size to the pass, not to the manuscript.
Marketing copy is short and formulaic, and you need a lot of it. A workhorse with a sharp brief beats a flagship with a vague one. Save the flagship for the one piece everything else is built on: your blurb and positioning.
A text model can write a great image prompt, but an image model makes the art. Image tools differ more in capabilities than in raw quality, so choose by what the project needs.
This is where models differ most, and where a poor choice costs the most rewriting. Brainstorming and outlining run fine on a workhorse. Prose you'd keep deserves the strongest model you're willing to pay for, and it carries the heaviest disclosure obligations.
| Also in an author business | Size | Look for |
|---|---|---|
| Research & comps | Workhorse + search | Live web search with clickable citations. Without search, it's guessing from memory. |
| Audio | Specialist | Voice quality, pronunciation control for names, licensing, and retailer rules on AI narration. |
| Reader email & admin | Fast | Speed, low cost, and working inside the tools you already use. |
| Your website | Flagship to build | A coding tool that can test what it builds. A workhorse is enough for small tweaks. |
This runs the five questions for you. Pick a task, answer a few quick questions, and it suggests which size to test first and what to check on the platform. It's a starting point for your own test, not a verdict.
Choose one option in each row.
Answer every row to see a suggestion.
For authors, this can matter more than which model is cheapest. Four questions come with every tool you pick.
Not legal advice · Last reviewed September 2026
These rules are still changing and differ by country and retailer. This module shows you what to check, and it reflects the rules as of the date above. For anything with money riding on it, like a contract, a rights sale or a dispute, check the current rules or talk to a publishing lawyer.
Some retailers ask at upload whether your book contains AI content, and that covers text as well as covers. At the time of writing, Amazon KDP separates AI-generated content (text, images or translations an AI produced, even if you then edited them heavily), which you must disclose, from AI-assisted work (you created it and used AI to brainstorm, edit or check it), which you don't. Other stores, contests, awards and anthology calls set their own rules, and some ban AI content outright. Check each one's current wording before you publish.
These questions are about what's in the book: the text, the cover and interior images. The social graphics and ads you make to promote it aren't part of the retailer's AI question, though ad platforms and your readers may have views of their own.
In the US, copyright protects human authorship. Text or images generated entirely by AI generally can't be protected on their own, and writing a detailed prompt doesn't change that. Your own writing, and your meaningful changes, selection and arrangement of AI material, can be. In practice:
Outside the US, the rules differ. The UK, for example, currently has a separate provision for "computer-generated works," though the government has consulted on changing it, and other countries are still deciding their approach. Check where you live and where you sell.
Legal isn't the same as welcome. In some genres, romance and romantasy especially, readers and reviewers push back hard on AI covers and AI-written prose, and a visible misstep can follow a book around. Many readers care much less about AI help with editing or admin. Know where your own readership stands, and decide what you're comfortable being open about before anyone asks.
The platform scorecard in Module 8 treats training and commercial rights as dealbreakers, because for most authors they are.
The cheapest model isn't cheap if you spend three hours fixing what it wrote. A pricier plan isn't a luxury if it gives you back a day a month. What counts is what's left after the bill and the cleanup time.
A subscription or pay-per-use charge. It's the easy one to see, and often the smallest.
Time spent rewriting, fact-checking and fixing. A weaker model often costs less money and more time.
Hitting limits mid-edit, re-uploading your series bible, juggling tools. Small each time, but it adds up over a book.
Subscriptions (a flat monthly fee) suit hands-on work you do yourself, and one plan usually covers several jobs. Higher plans mostly buy more usage, not a smarter model. Pay-per-use (billed per "token," roughly ¾ of a word) suits bulk jobs and automations, and it's where the price gap between sizes gets wide.
Use real numbers. Time a few pieces by hand, then time the same pieces with each option, including your cleanup. Your test kit (Module 9) gives you these numbers.
Rules of thumb that hold up
Spend where the reader will notice: blurb, cover, and prose if you use AI for it. Save on jobs with clear rules, like formatting, typo sweeps and variations. Upgrade a plan when its limits keep breaking your focus, not before.
The model is the brain, but the platform is where you spend your working hours. Today, platforms fall into a few groups: big general assistants, research tools built around web search, author-specific writing apps (many let you choose the model underneath), image and voice specialists, and open models you can run privately on your own computer. Those groups will outlive any product name. Judge any of them on:
Or lets you switch that off clearly and easily.
You can sell what you make on your plan.
How it did in your blind test, not in someone else's review.
Projects or memory that hold your style sheet and series bible between sessions.
It takes a full manuscript and actually reasons over all of it.
What you pay, how soon you hit the cap, and what happens when you do.
Clean import and export with Word, Docs, your writing app and your design tools.
Web search, file upload, images, voice, or a choice of models.
How the company sourced and licenses the material its models learned from, if that matters to you. Look for published statements, licensing deals and how it has handled authors' claims.
Apps built for fiction writers usually run other companies' models underneath, with author-friendly features on top: story bibles, scene cards, genre-aware tools. They can be a great fit, but check three things:
Name up to three platforms. Mark the two dealbreakers Pass or Fail first, because a Fail rules a platform out whatever its score. Then score the rest from 1 to 5 and set how much each matters to you. Saved in this browser.
Leaderboards mostly measure coding and math. They can't tell you which model line-edits your genre without flattening your voice. A blind test on your own material can. Setting one up takes an hour or two. After that, each rerun is quick.
Save it as a test kit so you can rerun it whenever something new comes out:
After your blind read, you can ask a flagship model for a second opinion. Treat it as input, not the verdict:
With launches coming every few weeks, you can't test them all, and you don't need to. Only run your kit when step 2 says a launch actually touches your work. Separately, rerun your kits every three months anyway, because platforms quietly swap models, change limits and update their terms.
A sample setup
Kate here is a hypothetical author, not a description of the course author's own workflow. Kate's choices and numbers are illustrative. Your mix will differ, and it should.
Take Kate, a romantasy writer releasing two books a year, with a weekly newsletter and a small ad budget. In this sample, Kate writes every word of the books and uses AI only for editing, marketing and promo graphics.
| Process | Size | How Kate chose | Hours back / month |
|---|---|---|---|
| Drafting | No AI | A personal choice, and it keeps copyright and disclosure simple. | — |
| Big-picture edit | Flagship | Tested two long-context platforms by asking about mid-book details. One kept mixing up side characters. | ~2 |
| Line edit | Workhorse | Test kit on three assistants. The winner had the fewest changes Kate reverted. | ~6 |
| Blurb | Flagship | Same subscription as the edits. Used it for 20 variations, and kept and reworked 2. | ~1 |
| Newsletter & ads | Workhorse / Fast | Workhorse drafts the core, fast model makes bulk ad variations at a usable rate of about 3 in 10. | ~4 |
| Promo graphics | Specialist | Chosen on character consistency and commercial rights. These are promo images only, not in the book, so the retailer's AI question doesn't apply. The cover is by a human designer. | ~3 |
| Monthly total | One assistant subscription + one image plan + a few dollars of pay-per-use ≈ $40–60 | ~16 |
Kate doesn't use the "best" model anywhere. Kate uses the cheapest option that passed each test, spends on the two jobs readers see first (blurb and graphics), and keeps the book itself entirely human.
A flagship everywhere burns money and limits. The cheapest everywhere burns hours on cleanup.
A vague request gets a vague result from any model. Fix the instructions before switching tools.
"Tops the leaderboard" usually means coding or math, which says little about your voice or your genre.
Without web access, a model invents plausible comps and stale trends. Turn search on and click the sources.
Training settings, commercial rights and retailer disclosure. Check them once at sign-up and again when the terms change.
A different model's habits can leave a visible seam. Finish the book, then switch.
Every option is something a reasonable author might try. Pick the best one.
Click each item as you finish it. Your progress is saved in this browser.