240
images from books 1 to 5 on contact sheets
My AI employee
Arthur has been working on my books since September 23, 2026. I published book 1 of my children's series “Mio and the Visitors” myself in August 2026; Arthur is now preparing books 2 to 5 and “100 Dinge” (in German) for publication. He isn't available as a package yet.
Arthur is my editor. He handles the work between a finished manuscript and the upload: four editing passes, checking images against the text and the characters, building the print files and, once I approve, creating the title on Amazon KDP as a draft. Every change he makes has a location, a reason and a status. I decide what holds across a series: titles, characters and style.
Evidence
every change with location and reason
Canon
titles and characters are my call
Approval
nothing is published without my word
Editor and book author
I had noticed that a lot didn't match in books 2 to 5 of my children's series, above all Mio's bedroom. Book 1 is published, so it is the reference. Arthur's assignment on October 11, 2026: check everything against book 1 once. He checked first, before a single image was made again.
240
images from books 1 to 5 on contact sheets
16
bedroom images, none like book 1
4
separate AI reviewers, one per book
Arthur put every image of the five books on contact sheets and had four separate AI reviewers each read one book against book 1. The result was clear: 16 bedroom images, not one of them like book 1. Window and headboard switched sides, Mio lay in bed in his day clothes with his hat on, and his dad wore five different outfits on a single day at the river. The cause was in the report: Mio and the inner world had fixed, word-for-word image descriptions, but the bedroom, the parents, the front door and the car did not. The image model invented them anew every time.
So Arthur didn't regenerate right away. First I decided the missing definitions, derived from book 1 and, where book 1 shows nothing, from the majority of the images: the bedroom, Mio in his pajamas, the parents' clothes. Only then did the regeneration run, with its scope and budget named in advance. Afterwards four separate reviewers checked again. The new images were largely in canon, but no book was ready for sign-off yet, because old images now stood out next to the new ones. Since then Arthur checks the whole book after every partial redo, and for single errors he edits the existing image instead of generating a new one.
Arthur for your business
Arthur doesn't work by his own taste but by the book's rulebook: who appears, how the text sounds, what must never change. You write that rulebook down once, and Arthur then checks every chapter and every image against it.
How I use him
Arthur edits the books of “Mio and the Visitors”, checks the images against characters and rooms, prepares “100 Dinge” for print and builds the print files for both projects.
Adapted to you
You define what applies in the book: characters, tone, audience, language and format. Arthur checks every change against it and brings open questions to you instead of deciding them himself.
What you change
Other possible uses
The process stays: Every change with location, reason and status, separate reviewers before every sign-off, facts such as ISBN and page count only from the source, and no upload without approval.
Arthur works directly in the project folders of my books, with the rulebook of each series:
Sound and read-aloud flow, language, consistency across all books, and continuity and logic. A sentence that works on paper but stumbles when read aloud is a finding for him.
Every image against the fixed descriptions of characters and rooms and against the text on the page. If an image doesn't fit, he edits it individually and keeps the rejected versions.
Several AI translators work in parallel from a shared glossary, then separate AI reviewers check accuracy and sound, and finally a scripted check runs over every chapter.
Interior with bleed, a cover that fits the page count, an e-book. Once I approve, he creates the title on Amazon KDP as a draft. I do the login myself.
Uploading and publishing on Amazon KDP are two separate steps, and both only happen after my explicit approval for that specific book.
Titles, characters, fixed lines and the style of a series are my call. Arthur proposes changes with a reason and never changes the rulebook himself.
He never touches account, tax or bank details on KDP. He doesn't keep credentials and doesn't type a password.
He may regenerate single images within an assignment. Larger image runs only happen after my approval with scope and cost stated up front, and no “until it fits” loops without a status update.
Facts such as ISBN, imprint and page count come from the source, never from memory. Every revision is checked by a separate reviewing role before I sign it off.
Standard
Start with Learn Claude, plus the articles on hiring and creating an AI employee. What's special about Arthur: he works against a written rulebook instead of taste, he has separate reviewers check his work instead of checking himself, and he backs every change with a location and a reason.
Premium
Arthur runs in my own business and works with the rulebooks of my books. Every book needs its own rulebook, which is why he isn't available as a ready-made template yet.
In parts, yes. For me, Arthur edits in four separate passes: sound, language, consistency and logic. Every change gets a location, a reason and a status, so I can follow it. What he doesn't take on: matters of taste and anything the series defines. That stays with me.
With fixed, word-for-word descriptions for everything that comes back: characters, clothes, rooms, the front door. Without such a definition, the image model invents it anew each time. And when a scene changes only slightly, the new image is made from an existing base image instead of from scratch. It still doesn't come out perfectly uniform: an elaborate setting like a climbing frame looked slightly different every time, even with a fixed description. So Arthur shows something like that in full only rarely, and then builds it from one single reference image.
He reports it with location and reason. One example from one of my book projects: the text says the father stands chest-deep in the water, the image showed him dry on the dock. Images like that are regenerated or edited before anything is uploaded.
He creates the title on KDP as a draft and uploads the print files once I approve it for that specific book. I do the login myself. Nothing is published until I explicitly say so; uploading and publishing are two separate steps.
Yes. KDP requires a disclosure when text, images or translations were generated by AI, including cover and interior images. If AI only corrected your own text, that counts as AI-assisted and doesn't have to be disclosed. The details with sources are in the article on publishing with KDP.
There's no template for Arthur's exact setup, but the learning path he grew out of is in the community.
Join the communityNovember 4, 2026 · 17 min read
How to publish a book on Amazon KDP: from manuscript to upload, trim sizes, ISBN, printing costs, royalties, AI disclosure and what to finish first.
Read the article: How to Publish a Book on Amazon KDPOctober 27, 2026 · 11 min read
Who owns AI images and AI texts, whether you may sell them and what applies to training. Explained under German and EU law. Not legal advice.
Read the article: AI and Copyright: Who Owns AI Images?November 1, 2026 · 25 min read
AI translators compared: DeepL, Google, Claude, ChatGPT and a local Qwen model tested with the same client email, plus privacy and pricing.
Read the article: AI Translators Compared: DeepL to LocalAlso:Hiring an AI Employee: The Process·Learn Claude: The Path in Five Stages·Building an AI Team: My Org Chart