The Standards Document
or: How to Keep AI From Writing the Floor
This morning, my wife read an article I published and stopped halfway through a sentence.
Not a grammar call. A voice call. She knows my writing well enough to recognize a pattern I would never produce, and she flagged it on sight. The same word, twice in the same sentence. A tell.
She didn’t say “this reads like AI.” She said it before she finished the sentence. She was right.
That afternoon, I added a gate to the standards document to catch that exact pattern in the future. By the time this post goes live, the gate will have been in force for less than twenty-four hours.
That’s the argument. The document grew from a catch. That’s how it’s supposed to work.
The Document Is Your Best Reader, Scaled
My wife is a first reader. She knows the voice. She reads slowly enough to feel when something is off and fast enough to name it before the paragraph ends. She is not always in the room. The document is.
That’s the point of it. The document is the first reader at scale: someone who knows your voice, never sleeps, never lets the small tell slide, and applies the same standard to every word the AI produces. The reader catches one failure on a Tuesday morning. The document catches that failure forever, along with every other failure that has ever been caught, plus the declared positions you took before anyone had a chance to fail at all.
Grammarly wouldn’t have caught what my wife caught. ProWritingAid wouldn’t have either. That’s not a grammar error. That’s a tell.
Grammar checkers catch errors. The standards document catches tells. Different problem. Different tool.
Where the Document Came From
A year ago, when I started the novel project, I fed Claude 3.5 two sets of writing samples. One stack from two decades back, rough drafts that had never been anywhere near a reader. Another stack from recent work. I told Claude to read before it said anything.
When it was done reading, I asked it to name the writer. What does this person always do. What do they never do. What would give them away across multiple samples. What are the first five rules already present in the writing that this person hasn’t written down yet.
The document it produced was not a checklist. It was a mirror. Claude found the patterns that were already there and reflected them back as principles. The standards were excavated, not invented. Claude just read what was on the page and named what it saw.
When the Substack launched, I did the same thing again. Handed Claude several non-fiction pieces I had already written, and the Substack Writing Standards grew from that reading. Same method, different medium, different document. The mirror move is repeatable. It works because the standards were always in the writing. Claude just reads what’s there.
Today’s catch is the latest entry in a document that has been iterating for a year. Every gate came from a failure somebody caught or a position I was willing to take in writing.
It’s Not That It’s Bad. It’s That It’s Overblown.
Before we go further, the reframe the document is doing, because otherwise the gates look like they’re compensating for incompetence. They’re not.
AI was trained on an enormous corpus of published prose. Literary fiction, long-form journalism, the entire universe of serious writing. It learned the moves from the best writers who ever published. The em dash isn’t in every sentence because the AI is bad at punctuation. It’s there because the AI learned that sophisticated writers use em dashes, and without a governor it applies that lesson at full throttle.
It’s not incompetent. It’s overconfident. It’s a student who read everything and learned all the moves and now deploys them all simultaneously, in every sentence, because it cannot feel when the prose is full.
The standards document is the saddle. It’s not teaching the AI to write. It’s reining in something that already knows how to write but has no governor on its own enthusiasm. The saddle doesn’t make the horse slower. It makes it rideable.
Gates Come From Two Places
The first is a caught failure. A reader flags something, you recognize it as wrong the instant you see it, you add the gate the same day. The version history is a rejection log. Today’s entry is the most recent.
The second is a declared position. You sit down and decide what you will not accept, before a single draft ships.
The em dash is the clearest example. Not caught in a draft. Not flagged by a reader. A position taken in advance, from conviction: the em dash key doesn’t exist on a standard keyboard, it wasn’t taught in English class, and anything an em dash does a period or a comma does better, with punctuation that actually exists.
That’s not a style preference. That’s an argument. I wrote up the full version as a Note if you want to see what a declared position looks like in public.
The positions can be completely idiosyncratic. Maybe you hate similes and love metaphors, and the document bans similes outright and requires metaphors instead. The opposite position is equally defensible. Plenty of strong writers lean on similes. Neither is wrong. Both are positions.
The document doesn’t care what the consensus says. It only cares what you say, because its job is to make the AI write like you, not like the average of everyone who ever wrote well.
This is the distinction between a standards document and a style guide. AP and Strunk and White try to govern all writers. Your document governs one. The idiosyncrasy isn’t a bug. It is the entire point.
Negative Models Belong in the Document. Positive Models Do Not.
Name the author you must never sound like. The negative model is useful because the failure mode is specific, recognizable, and visceral. The AI knows their work. “If this reads like [author], start over” is more enforceable than “avoid purple prose,” because the AI cannot measure purple but it can measure whether the prose explains emotions instead of earning them, whether the protagonist is passive in their own story, whether the internal monologue is doing the work the scene should be doing.
Do not name who you want to sound like. Positive models are a trap, and not just for craft reasons.
You aim for a specific literary voice and the AI gives you the mannerisms without the control, the surface without the thing underneath that makes the surface work. A human writer absorbs influences across decades and they come out transformed. The AI pattern-matches and reproduces. Those are not the same process.
There is also the legal layer. Style itself isn’t copyrightable, but the line between influenced-by and derivative-of gets uncomfortable fast when the influencing is done at scale by a system that doesn’t know it’s doing an impression.
Your voice is already in the document, excavated from your own writing. That’s the positive model. It doesn’t have a name attached to it. Keep it that way.
Here is a gate that lives in the fiction document by itself. If my female characters ever get Bella’d (Twilight, Stephenie Meyer), I will rewrite everything. One word, entire failure pattern compressed.
A Bella’d character waits for the plot to happen to her, defines herself through her relationship to a more interesting person, has no interiority that isn’t about him, and makes choices that are actually non-choices because she has no agency to exercise. The AI knows what it means. No further definition required.
That’s what a negative model is supposed to do: compress a complex failure into a term you can apply on sight.
One Document Per Voice
The reader is already asking whether they need one of these for every type of writing they do. The answer is yes. Several is not excessive. Several is accurate.
I write in three registers: literary fiction, scholarly non-fiction where the citations have to hold and the reader checks sources, and personal craft essays on Substack. Same writer. Same AI. Three documents. They share a floor, the em dash ban, AI-tell detection, no throat-clearing openers, no weak closers, and above that floor they diverge.
The scholarly document carries an Extraction Test: mentally remove an entire category of supporting evidence and ask whether the central argument still holds on its own terms. That gate would be meaningless in fiction and corrosive in a personal essay.
The personal-voice document requires dark humor in every post, observed with the dry recognition of someone who watched the thing go wrong. That would be bizarre in a scholarly register where the evidence carries the argument.
The fiction document has an Author Fingerprint Fidelity gate that doesn’t measure what’s wrong, it measures whether what’s right is present: dark humor in tense scenes, tonal whiplash, specificity over generic darkness. A chapter without the fingerprint isn’t mediocre. It’s a different writer’s chapter.
One document trying to govern all three would destroy all three voices. The documents are not variations on a theme. They are different contracts written for different purposes, enforced by the same AI.
Keep the Document Alive
The document is living, which in practice looks like two modes.
Ad hoc is what happened this morning. Something got caught, the gate went in the same day, no ceremony. The catch happened, the document reflected it before the next draft. That is most of the work.
Periodic is the formal review pass. Look across recent drafts for patterns. Which gates are not pulling their weight. Which failure modes have appeared more than once without a gate to catch them. Which positions turned out to be wrong and need retiring.
My documents live in plain markdown files near the work they govern. The AI loads the relevant one at the start of every session. My infrastructure auto-commits versions to a private repo with a message describing what changed and why, so the version history maintains itself. You don’t need my infrastructure. You need a file somewhere you’ll find it, a workflow that loads it every session, and the discipline to update it when something crops up.
Series Fiction: One Document, Whole Project
If you’re writing a standalone novel, a looser approach survives. If the voice shifts slightly between drafts, the reader never knows what it was supposed to sound like.
A series reader has already internalized the voice across three, four, five books. They know the rhythm. They know what the narrator does and doesn’t do. Any drift between books isn’t a style choice. It’s a continuity error.
Voice drift is harder to name but easier to feel than plot inconsistency. The reader just knows something is off.
One document prevents that. Book 5 loads the same standards as Book 1. The voice doesn’t wander because the contract doesn’t change. Versioned across the life of the series. Never forked. Never per-book.
For a series, the document is more important than it is for standalone work, not less.
Do the Homework First
Before the next section, a practical warning. The document cannot precede the work that makes it possible.
You cannot build a fingerprint document before you have a fingerprint. I wrote roughly 500,000 words of exploratory material before the first publishable novel. That’s my number, not the requirement. The requirement is enough material that the patterns become visible.
At AI-assisted scale, 100,000 to 150,000 words is achievable in a few weeks of serious sessions. You are not typing those words. You are directing. You tell the AI what happens, the AI writes it, you read what comes back and edit until it’s right.
The authorial labor is vision and judgment. Authorial not the fingers.
Most writers try to shortcut this. They want the system before they’ve done the work that makes the system possible. The document can’t tell you what your voice is if you haven’t written enough to have one yet.
Two questions worth distinguishing cleanly. Running the fingerprint test on your own writing takes a few pages. Rough work counts. Old work counts. The test works on whatever you have, and the fingerprint is often clearest before craft training starts smoothing it out.
Publishing a series novel takes much more. The homework builds the system. The system governs the work. Never reversed.
Find Your Fingerprint
Here is the prompt. Paste it into your AI of choice. Bring a few pages of something that feels like you, even something rough, even something old. The low bar is the point.
I'm going to give you some of my writing. It doesn't have to be good.
Read it before you say anything.
[paste a few pages - a chapter, a scene, an essay draft, anything
that feels like you]
Now answer these questions:
1. If you encountered this without my name on it, would you recognize
it as mine across multiple samples? What gives it away?
2. What does this writer never do? Constructions, rhythms, or patterns
that are conspicuously absent.
3. If I handed you a paragraph and asked whether I wrote it or an AI
wrote it, what would you look for?
4. Based only on what you just read, write my first five writing
standards - the rules already present in my work that I haven't
written down yet.
Don't invent. Don't advise. Read what's there and reflect it back.
Questions 1 through 3 are the diagnostic. They tell you whether your voice is detectable at all. Question 4 is the first draft of the document, excavated from what already exists, not imposed from outside. If the AI can’t recognize you across samples, that’s information too. The voice isn’t consistent enough yet to codify. Fix that first.
Some of what I fed Claude was written over two decades ago and was very rough. The fingerprint was there anyway.
What Claude found, specifically: I’m a dialogue writer. I hear dialogue in a scene before I see anything else. Scenes start with two people talking, minimal stage direction, minimal description outside the exchange itself.
That’s not a technique I learned. It’s how my brain processes a scene. The fingerprint test surfaced it as a pattern before I had articulated it as a standard.
The roughness didn’t hide the fingerprint. The roughness may have made it clearer.
One Last Thing Before the Close
Standards documents are model-agnostic. Prompts are procedural: they tell the AI how to run a workflow, and a more capable model may need different instructions to follow the same process. Standards don’t work that way.
A gate that bans em dashes doesn’t care what model is reading it. The output either contains the banned pattern or it doesn’t. If anything, a more capable model makes the standards document more effective, because it’s better at following precise instructions consistently.
When you upgrade your AI, update your workflows. Your standards probably do not need to change. They are yours, not the model’s.
The Close
Your AI has no taste. It has no standards of its own. It will meet whatever bar you set, and if you set no bar, it will meet the floor. The document is the bar. Without it, you aren’t co-writing with an author. You’re co-writing with a very fast typist.
The standard is yours to define and refine as you go. It should be a living document, because things crop up that you didn’t know you hated until you saw them on the page. You don’t discover your standards by thinking about them. You discover them by writing, reading what came back, and feeling the wrongness before you can name it. The document is how you name it. And how you make sure you never have to feel it again.
The version history isn’t administrative overhead. It’s the autobiography of your taste, written one caught failure at a time.
The document is your best reader, scaled. They’re not always in the room. The document is.
You may also like: