We are building Maxwell, an AI developmental and line editor, for a world that is both excited about AI and increasingly wary of it. Every week I write down what we have learned trying to navigate that, our failures included, because a writer deciding whether to trust us with a manuscript deserves to see how the work is actually done. One principle sits under all of it. Maxwell is an editor, not a writer. It will tell you what a sentence has to accomplish. It will not write the sentence for you. Not One Word.
Maxwell does not have a single quality bar. It has eight families of writing: fiction, poetry, writing about the author’s own life, writing about real events outside it, student coursework, peer-reviewed scholarship, writing that argues a case to a particular reader, and instructional writing that teaches a reader how to do something. Each family has its own editorial standard, and forty-four genres are filed under those eight. Just like your local bookstore (I hope you still have one). Maxwell reads your manuscript and proposes a genre and a form (novel, short story, novella, collection of poems, etc.). You confirm it or pick your own. Maxwell then builds the standard for that exact pairing of genre and form. So a free verse poetry collection and a police procedural are never measured against the same list of questions.
We got there by being wrong three times. A review we commissioned caught the first. My cofounder caught the second. Real writers walked into the third before we did.
Two poetry genres were being judged as fiction.
On the second of August, we found that Literary Poetry and Free Verse Poetry were being read against a fiction standard. Not something adjacent to fiction. The fiction standard itself: 527 lines, including sections on dialogue, scene function, and narrative authority. The poetry standard those poems should have had is 138 lines and asks about image, line structure, and sound.
The cause was something any librarian, researcher, or bookstore owner would see in a heartbeat. Every genre belongs to a family, and the family determines the backbone the standard gets built on. The list of families the app would accept was typed out by hand in two separate places in the codebase, and they didn’t match. So a genre correctly identified as poetry ended up in a family with no poetry backbone and was quietly rewritten as fiction.
The families/genres structure is now driven by the standards themselves, so they can never be out of sync. A genre whose family we cannot place is held for review instead of filed under fiction.
A textbook is not an essay without an argument.
My cofounder Jeff uploaded a draft of his international arbitration textbook on the first of August. Maxwell called it an academic essay. It was not being careless. An academic essay was the closest thing we had, because we had no shelf for a textbook at all.
The academic standard asks whether a book defends one arguable claim. A textbook is not trying to defend a claim. It is trying to set out established knowledge in an order a learner can follow, and our standard would have marked it down for doing exactly its job.
Two days later we shipped instructional writing as its own family, judged on whether the book actually teaches. Does it serve one consistent learner rather than drifting between a beginner and an expert? Is every idea available at the moment it is first relied on? Do the worked examples show what they claim?
Before we read the book, we drafted a set of checks based on what we assumed a textbook contains. One of the checks was wrong. We had planned to require that every exercise trace back to its own stated learning objective. Jeff’s book states only two objectives in 93,000 words and teaches well, so that check would have marked down a book that works. We dropped it and asked instead whether the book covers what it tells the reader it will cover.
Then we ran the new editorial standard over that same book, and it found a real defect. Chapter four leans on a legal doctrine that the rest of the chapter depends on, and the text says it is “discussed below.” The definition never arrives — the reader is told the rule matters and is never told what the rule is. The old academic standard would have missed that, because one of its checks was whether a thesis had been defended, and a missing definition is not a failure to defend one.
Even then, we nearly shipped it broken. The instruction that decides whether Maxwell can judge a manuscript at all used a textbook as its written example of something we cannot judge. This was an artifact of Maxwell’s original focus on novels. The day the textbook family went live, that sentence would have taught Maxwell to turn away the exact manuscript the family was built for.
We were reading the word “business,” not the standard behind it.
The same type of mistake, in a different place, twice more. In Maxwell, the family that argues a case was defined by business language. So the writer of a book of strategic guidance for church leaders was told no standard existed for that manuscript. A 36,000-word practitioner guide for working professionals was turned away too, with high confidence. Maxwell had already agreed that the book was instructional, but “instructional” at the time meant textbooks and students.
Both books were rejected because of the shelf they sit on, not what they were trying to do. We now describe each family by what it judges, not by what it tends to be about. Business writing is one genre in that family, not the whole of it. The same is true for instructional works.
The edge cases of what a tool will accept reflect real product decisions, whether you make those decisions deliberately or not. Ours were made by accident three times.
Next week in Not One Word: why Maxwell now refuses manuscripts it used to accept.