Chad Mathews
Aug 20, 2026Source: Margin Harbor content scrubber + this site's voice guide, in use since 2026

The banned words list

AI writing has fingerprints. I keep a list of them and run it against every draft, including this one.

A field note from Chad Mathews, drafted with Claude. Chad directs the work and lived these decisions; Claude does the writing; Chad edits every line and approves it before it publishes. Where an idea or a reference came from Claude, the note says so.

There is a specific way that machine-written text reads. Not wrong, exactly. Grammatical, confident, on topic. But if you have read enough of it, you can feel the machine behind it, and once a reader feels that, they stop trusting the words. For a personal brand, that is the whole ballgame. The moment a piece reads as generated, the credibility I am trying to build leaks out of it.

The failure is that you cannot reliably catch this by ear on your own writing. Sentence by sentence, everything passes. The tell only shows up in the aggregate: the same shapes repeating, the same handful of words, the same rhythm. I would read a draft, think it was fine, and only later notice it was wearing the uniform. So I stopped trying to feel it and wrote the tells down instead.

What the list actually is

It is a plain file of patterns to catch, split into three kinds.

The words. A flat blocklist: delve, leverage, streamline, unlock, empower, seamlessly, robust, comprehensive, game-changer, cutting-edge. These are not banned because they are bad words. They are banned because generated text reaches for them far more often than a person does, so their presence is a signal, not a crime.

The shapes. Harder to catch and more damning. The “it isn’t just X, it’s Y” construction. Rule-of-three lists used for rhythm rather than because there are three real things. Rhetorical questions used as transitions. Em-dash chains, more than a couple in a paragraph. Closing paragraphs that summarize what was just said instead of ending on a point. A sentence rhythm that is too evenly balanced, setup and payoff, setup and payoff, until it lulls.

The stance. Copy that leads with the tool instead of the cost. Hedging: “may potentially,” “could possibly.” Claims with no number behind them. More than one call to action. Any line a real person would not say out loud.

How I run it

Every draft gets checked against the list before it goes anywhere. On this site that check has a name, and it is not me eyeballing it: it is a scrub pass that reads the piece against the file and flags each hit with the exact phrase and a fix. Pass or fail, per piece, with the specific violation named. No vibes.

Here is the part that keeps me honest. This site failed its own list after launch. Six notes went up, all of them full of em-dashes, because I had drafted them before the em-dash rule was fully enforced. I went back and stripped every one. The list existed and the writing still broke it, which is the normal case, not the embarrassing exception. A checklist does not prevent the mistake. It catches it before a reader does, and it catches it the same way every time instead of depending on whether I happen to notice.

Where this was old before I got here

The comparison came up while Claude and I were writing this note, and it is a fair one. House style guides have done this for a century. Every newsroom and publisher keeps a sheet of words they do not use and constructions they do not print, so that a hundred different writers produce something that reads like one voice. Prose linters do the automated version. What is new is only the target: the fingerprints of a language model instead of the habits of a sloppy writer. The mechanism, a written list of tells that everything gets checked against, is borrowed straight from editors who never touched an AI.

The honest limit

Passing the scrub does not make writing good. It makes writing not-obviously-machine, which is a floor, not a ceiling. A piece can clear every banned word and still be hollow. The list catches patterns; it cannot supply a point of view, a real story, or a reason for the piece to exist. Those are still mine to bring.

And the list ages. The tells are downstream of how the current models write, so as they change, some of these words stop being signals and new ones take their place. The file is not a settled truth. It is a running record of what generated text sounds like this year, and it needs re-reading against real drafts, not trusted as scripture. The discipline is not the specific words. It is refusing to grade your own writing by feel.

Where this came fromThis note was extracted from Margin Harbor content scrubber + this site's voice guide, in use since 2026, a working document from a live project, written to solve the problem before it was written to explain it.