By Matty Cartwright · @themattycartwright · mattycartwright.com
8-minute read · Last updated April 2026
Why your AI copy sounds like AI copy
You can spot it in two seconds. The em dashes. The three-item lists. The "let's dive in." The sentence that sounds like it belongs on a motivational poster in a WeWork bathroom.
I write with Claude every day. LinkedIn posts for clients, video scripts, blog posts, captions. Thousands of words per week. And I spent months frustrated by the same problem: the output read like a press release from a company that doesn't exist.
The fix wasn't better prompts. It was a filter document. A file Claude reads before writing anything for me. It contains every word, phrase, sentence pattern, and structural habit I've caught Claude defaulting to. When Claude reads this file, the writing gets cleaner. Not perfect. But the floor goes way up.
I call it the Anti-Slop Writing Rules. I'm giving you the whole thing.
What the document does
The rules work in layers. Each layer catches a different type of AI-flavored writing.
Layer 1: Banned words. The obvious tells. "Delve," "tapestry," "landscape," "robust," "seamless." You know them when you see them. This list has 50+.
Layer 2: Banned phrases. The throat-clearing openers ("Here's the thing:"), the emphasis crutches ("Full stop."), the meta-commentary ("Let me walk you through this"). These are harder to spot because they feel like voice. They're not. They're filler.
Layer 3: Banned sentence patterns. "It's not about X, it's about Y." "The result? [Statement]." "X isn't just A. It's B." Claude leans on these like a crutch. Every model does. Banning the pattern forces original structure.
Layer 4: Banned structural patterns. False agency (objects doing human things), dramatic fragmentation, negative listing, rhetorical setups. These are the deepest habits and the hardest to catch without a ruleset.
Layer 5: Grammar rules and tests. Active voice with named humans. No synonym cycling. No padding with three items when you only have two real ones. And three tests you run on every piece of output: the Specificity Test, the Infomercial Test, and the Commitment Test.
How I use it
The document lives in my Claude Code project as banned-language.md. Every writing skill I've built reads this file before drafting. It's the first thing Claude sees when I ask it to write.
You don't need Claude Code to use it. Paste the document into any LLM conversation before your writing prompt. Or upload it as a file. Or copy the sections you care about into your custom instructions. The rules work the same regardless of where you put them.
The document also has a "What Good Looks Like" section at the top that's blank on purpose. You fill it in with 4-5 sentences describing what your writing sounds like when it's working. The model needs a destination, not just a list of things to avoid. Without it, the output gets clean but generic. With it, the output starts sounding like you.
A few rules that changed the most for me
No em dashes. Claude uses em dashes every third sentence. Banning them forces the model to write complete thoughts separated by periods. The rhythm of the output changes overnight.
No synonym cycling. AI has built-in repetition penalties that make it swap words for no reason. "Customers" becomes "clients" becomes "users" becomes "patrons" in the same paragraph. Banning this forces consistent language.
No Rule of Three. LLMs love groups of three. Three examples, three benefits, three takeaways. Half the time, the third item is padding. If you only have two real things, use two.
Kill hedging adverbs. "Really," "genuinely," "fundamentally," "inherently." These prop up weak verbs. If the verb needs an adverb to land, the verb is wrong.
The Infomercial Test. Read the output out loud. Does it sound like something a friend would say, or something from a late-night commercial? If you'd hear it on TV at 2 AM, rewrite it.
The full document
The Anti-Slop Writing Rules cover banned words, banned phrases, banned transitions, banned sentence patterns, banned structural patterns, grammar rules, rhythm rules, and three quality tests. The document is designed so you can hand it to any LLM and get immediate improvement.
What changes and what doesn't
The banned words and phrases are stable. I haven't added to that list in weeks. The structural patterns section grows every time I catch a new habit. Last month I added "Fake Personal Voice" after noticing Claude using "I" without committing to an actual opinion. "I think this could be interesting" is not a personal voice. "This works" is.
The document is version-controlled. I update it in place. When I catch a new pattern, I add it. When a rule stops being useful, I cut it.
If you use AI to write anything, you need a filter like this. Not because AI is bad at writing. Because AI is bad at knowing when it sounds like AI. That's your job. This document makes the job faster.