How we use AI · The long version
Inside the machine
The short version is on our How we use AI page. This is the long one: the production system we built to make these books, stage by stage, with the models it runs on, the rules it keeps, and where the two of us sit in it.
A publishing house made of software
Most publishers have an org chart: editors, art directors, fact-checkers, production staff. So do we. The difference is that almost every seat on ours is filled by software Noah built: 53 skills, each one a written set of instructions for a single job, and 79 tools that do the mechanical work around them. The two seats that aren’t software are Noah, who runs it, and Karen, who edits every manuscript.
The system is on its third generation since February 2026. Each rebuild is dated, documented and kept, so we can always see what changed and why.
The assembly line
A book here isn’t made in one sitting. It moves down a line. Each station has one job and ends in one of three verdicts: pass, rework or fail. Noah is at every station, reading, questioning and sending work back. Some stations are blind: the reader isn’t told why the work was done that way. Three stations are formal hard stops, where the line halts until we sign off. Beyond those, we approve every illustration and Noah reads every proof before a book ships.
1The words
Intake
The system breaks the source down into story, scenes and what each age can hold; we correct it.
The planHard stop
A backbone for the whole story, then a page or chapter plan.
Drafts
One book per reading level, written against the source text.
Level check
Makes sure each level is its own book, not one book resized.
Blind readers
A hostile read, a read-aloud, a hunt for machine-sounding lines.
Source check
Every claim held against the original text.
Karen's editHard stop
The whole manuscript.
2The illustrations
Illustration plan
Which moments get an illustration.
Briefs
A written brief for every illustration.
Art-director review
The briefs read as a sequence.
Blind framing read
A reader who never saw the plan reads the briefs cold.
We lock the briefsHard stop
Character sheets
So the knight is the same knight on every page.
Render, reject, render again
We approve every illustration
3The book
Interior
Built to print spec.
Front and back matter
Proof check
Covers
Print and Kindle.
4The launch
Store listing
Blind-read, like the manuscript.
Ebook build
Noah reads the proofs and publishes
Our book list updates itself
Who does what
There’s no single “the AI”. Each job goes to the model we’ve found does it best, and they come from two different companies, so one model’s habits get caught by another’s.
Claude Opus
Carries the work from station to station. It plans each book from the source, writes the illustration briefs and runs the source checks.
Claude Fable
Drafts every book alongside Astra. It is also our main blind reader: it hunts for machine-sounding writing, grades drafts cold, in a fresh session that never saw them being written, and reads the illustration briefs before anything is rendered.
GPT-6 Astra
Drafts each manuscript with Fable, the two taking turns revising each other's work. It is also our sounding board: a cold second opinion on plans, drafts and store listings, and a polish on the prose before Karen's edit.
GPT Image 2.5
Renders the illustrations for new books from our briefs, in our house styles. Earlier books were illustrated with its predecessors.
The rules that make it a machine
No station grades its own work.
The reader that attacks a draft never saw it being written, and no model approves a book. We do.
Reviewers work blind.
Critical reads go to agents that aren't told why the work was done the way it was, because a reviewer who knows your reasons tends to forgive them.
Cheap checks guard expensive ones.
Illustrations cost the most, so every brief is read by an art-director pass and a blind reader, and locked by us, before a single image is rendered.
Every lesson becomes a rule.
When something goes wrong, the fix is written into the system instead of left to memory.
It keeps the record
Every action on the line (a draft saved, a check passed, an illustration approved) writes an entry to a permanent log. Nobody types in a book’s progress by hand; its status is rebuilt from that log, so it can’t drift from the record. As of September 2026 there were close to 3,000 entries.
{ "topic": "war-and-peace", "band": "B",
"event": "draft_committed", "actor": "code", ... }Alongside the log there’s a diary. At the end of every working session the system writes a candid account of what happened and what it learned, and Noah adds his own notes. Since April that comes to 633 session logs and 137 days of diary: about 1.4 million words, more than twice the length of War and Peace.
Why we show you this
“Made with AI” can mean almost anything. We’d rather show you what it means here. To judge the result, read our Don Quixote next to one made from a single prompt.