AdVenture Media
Why it works

The research behind the rules

Readers got good at spotting generated text in about eighteen months. The tells are not subtle anymore, they are a shape you recognize before you finish the first paragraph. This page lists them, says where each one came from, and explains why the fix is a file and not a cleverer prompt.

Three findings that shaped the product

Samples beat self-description. Every, Lago and the Adpharm playbook tested this separately and landed in the same place: when you describe your voice you describe the writer you would like to be. When you paste two hundred words you wrote on a Tuesday, the model sees the writer you are. So the interview here never asks for an adjective. It asks for three samples and derives the rest.

The gap between drafted and sent is the voice. Lago's engineers noticed that the most useful data they had was every AI draft they had rewritten before sending, next to what they sent. The difference between those two is a precise, testable description of what a person changes and why. Ten pairs teach a model more than a page of rules. VOICE.md has a section waiting for yours.

Prevention at the system level beats correction at the prompt level. A rule in a file is read every session. A rule in a prompt is typed when you remember it. The Adpharm's estimate is that installing the guardrails once removes about eighty percent of the obvious tells before you write a word. That is why the rules below appear inside the generated CLAUDE.md rather than on a poster.

The sixteen patterns

These are what the checker looks for. Hard means a single instance marks the passage. Soft means it is a matter of density and judgment.

hard

Em dashes

One is punctuation. Three in a paragraph is a signature, and every reader's detector fires on it now.

Fix: Use a comma, a period, a colon or parentheses. In ranges, write 'to'.

hard

"It's not X, it's Y"

The construction performs insight without delivering any. It appeared four times more often in corporate writing between 2023 and 2025, and readers have learned the shape.

Fix: Delete the rejected half and state the claim. 'It's not about features, it's about outcomes' becomes 'Buyers pay for outcomes.'

soft

Forced lists of three

Triples feel complete, so the model reaches for three adjectives, three benefits, three takeaways whether or not there are three.

Fix: Use however many items there actually are. Two is fine. Five is fine.

hard

Staccato fragments

'Not the tool. Not the training. The person.' Three clipped fragments in a row read as a machine doing drama.

Fix: Combine into one flowing sentence with real connective tissue.

soft

Metronome rhythm

Every sentence the same length. Several short ones in a row is the single most obvious tell; a wall of 25-word sentences is the second.

Fix: Vary on purpose. Short, then medium, then occasionally a long one that builds. Read it aloud.

hard

Scaffolding sentences

'Here's the thing:', 'It's worth noting that', 'Let me explain', 'A couple things:'. The sentence announces a point instead of making one. The reader learns nothing from it.

Fix: Cut it and start with the content it was introducing.

hard

Brochure vocabulary

Delve, leverage, seamless, robust, unlock, empower. Nobody says these across a table. They mark text as generated even when the ideas are good.

Fix: Say the plain word you would use out loud. 'Use' for leverage. 'Look at' for delve into.

hard

Throat-clearing openers

'In today's fast-paced world', 'Great question', 'I hope this finds you well'. The reader has to wade through a warm-up to reach the first fact.

Fix: Start where the substance starts.

hard

Summary closers

'In conclusion', 'Ultimately', 'At the end of the day'. A restatement of what was just said, plus a bow on top.

Fix: Stop when the substance ends. Real messages just end, or end on the next concrete step.

hard

Honesty preambles and hedges

'To be honest', 'I'll be honest', 'Honestly', 'If I'm being honest'. A credibility hedge before a claim, which implies everything else was not honest.

Fix: Delete the hedge, keep the claim.

soft

Significance inflation

'Stands as a testament', 'plays a vital role', 'marks a turning point', 'underscores the importance'. The writing tells you something matters instead of showing why.

Fix: Replace with the specific fact that made it matter.

soft

Avoiding the word 'is'

'Serves as', 'functions as', 'represents', 'boasts'. The model dodges a plain copula because it has been trained that variety sounds smart.

Fix: Write 'is' or 'has'.

soft

Abstraction where a fact belongs

'Significant challenges' instead of 'missed payroll twice'. Vagueness is what a model produces when it was not given the specific.

Fix: Put the number, name or date in. If you do not have it, mark it [confirm] rather than papering over it.

soft

Bolded label bullets

'**Speed:** We ship fast.' A list of bold labels with colons is a slide, not a message.

Fix: Plain bullets or, better, a sentence.

hard

Question-then-answer rhythm

'The result? Revenue doubled.' 'Why does this matter? Because...' A fake drumroll before every reveal.

Fix: State the result.

soft

Talking to the reader about the reader

'You've got this.' 'Great point.' A pep line or a compliment where a colleague would have written nothing.

Fix: Delete it.

What the good writers actually do

The writers Every interviewed disagree on almost everything except this: none of them lets the model decide what the piece is about. Kevin Roose used AI for a year of research, transcription and fact-checking on a 448-page book and drew a bright line at generated prose, then ran every chapter past a council of reviewer agents and threw out ninety-five percent of what came back. Dan Shipper does let the model write sentences, but only as something to react against: he says what is wrong, it tries again, and the final paragraph is the residue of that argument. Maggie Appleton hands the model her own imperfect sentence and asks for clearer versions, because a suggestion anchored to her words still sounds like her.

Mike Taylor's rule of thumb ties it together: the prompt should be larger than the output. If you give a model forty words and ask for four hundred, the extra three hundred and sixty are its defaults, and its defaults are the tells on this page. If you give it your business, your vocabulary, your proof and three hundred words of your own writing, the output is mostly yours with the typing done.

Why the checker uses almost no AI

Fifteen of the sixteen patterns are regular enough that a few lines of code catch them, quote them and explain them, instantly and for free. The model is reserved for the one thing code cannot judge, which is whether a draft sounds like a specific person. The same principle runs through the kit builder: three of the four files are assembled from your answers with no model involved, and the fourth uses one because deriving rules from prose is a language task. The best AI tools use surprisingly little AI.

Sources

Build your kit