In September Claude handed me a Friday post and said it had been checked. The first word was "Most".
That word is on a list of seventy-three I keep, and it is on the list because almost every AI-written post on LinkedIn starts with one of them. I caught it before the post went out. The tool that said the draft was clean had not run the check. It had gone from memory, which is exactly what the check exists to prevent.
I use AI when I write, and I want to be exact about how. The idea is mine. The rough draft is mine, typed fast and badly. Claude polishes it, and then the draft goes through four layers of checks. This article is those layers in the order they run, and the two things they cannot catch.
The list
The openers that give AI posts away went on the list: "Here's the thing", "In today's world", "Let's be honest", "Imagine", "Most", "Today", and about seventy more. No post and no paragraph may start with one of them.
Paragraph level matters more than it sounds. A draft that opens well usually relapses by paragraph four, once the model settles into its defaults. These openers are not wrong because they are old. They are wrong because they are sentence-shaped nothing, a way of clearing the throat before saying something, and readers have learned to see them from across the screen.
Punctuation is a signature
No em-dashes, anywhere. The em-dash is not even on my keyboard, so I keep asking myself how I could have written one. A comma or a full stop does the same job without announcing itself.
Full stops and commas go outside quotation marks, the way I learned it at school. The models put them inside, every time, and that alone is enough to give a draft away. And no contractions in the formal register. It slows the sentence down, and a slow sentence shows its seams, which is the point: I want the reader to feel a person choosing each word.
Three checks that never get tired
A checker comes first, a short Python script that splits a draft on blank lines, tests every paragraph against the list, counts the dashes, and flags contractions. It runs on every draft before I read it as a reader. Claude skipped it that Friday and checked the list from memory, which is how "Most" got as far as my screen.
Next comes a reviewer. The content calendar I built has a review endpoint that reads a draft against my written ruleset and returns a verdict, what fits, and what to fix. The post I wrote about this system went through six rounds before it passed. The reviewer never tires and never gets attached to a sentence, which is the quality I lose by round four.
Last is a general-purpose editing skill that strips the phrasing patterns no list can hold: the hedges, the balanced pairs, the summary sentence that restates the paragraph above it.
None of this makes a post good. It only removes what makes a post obviously not mine.
What the checks cannot see
Rhythm is the first gap. After the surface was clean, I noticed something I had no rule for: every paragraph came out the same length. Even, tidy, and instantly readable as a machine. The checker now prints the length of each paragraph, and a column of near-identical numbers is a flag.
The second gap is worse, because it is not about style at all. A model fills gaps with the most plausible version of your life. In October it told the story of my agency as if I had built it alone, when I started it with a business partner. Another time it invented a confession for a post about client testimonials, a line where I admitted I had been nervous to ask. I was not nervous. Both clients had replied within days without a draft, which was the actual story and a better one.
No list catches that. The reviewer checks my rules, not my life, and only I can read a sentence about me as a claim I am signing. So the rule I added is not about words at all: the machine does not supply episodes. Every dated, concrete thing in a post comes from me.
The lists make a post sound like me. Whether it is true is still a job I do by hand, and the day I stop doing it by hand is the day my posts read like AI again, whatever is on the list.