Make your AI stop writing like an AI

What this actually is

Wikipedia editors have been cleaning up AI-written articles for a while now, and along the way they built a field guide of every pattern that gives the writing away. Em dashes. Curly quotes. Puffery. Fake sourcing like "experts argue". The rule of three. The bolded label at the start of every bullet. It runs to about 15,000 words and it is the most detailed list of AI writing tells that exists anywhere.

The page is called Signs of AI writing and it was built by their AI Cleanup project.

The move is simple. Take that guide, flip it, and hand it to your AI as a standing instruction. Instead of a list for catching AI writing, you get a list of things your AI is never allowed to do. I turned it into a skill called Humanize and it now runs in the background on everything I write.

Option 1: make it a skill in Claude

A skill is a file Claude loads on its own when it is relevant, so you set it up once and never think about it again. You need a paid plan for this, and code execution turned on.

1. Download the zip I sent you

   Leave it zipped. Claude wants the zip file, not the unzipped folder.

2. Open your settings

   On claude.ai, click your profile picture in the bottom left, then Settings. On Team or Enterprise, an owner has to switch on Skills and code execution in the organization settings before this shows up.

3. Go to Customize, then Skills

   You will see the built in ones sitting there already, things like Excel and PowerPoint.

4. Upload the zip

   Click upload skill, pick the file, and give it a few seconds. It lands in the list with a toggle next to it. Leave the toggle on.

5. Write something

   Claude picks it up on its own when you ask for an email, a caption, a post, anything human facing. If you want to force it, say "use the humanize skill" in the message.

Or build it yourself, which is what I did

Open the Wikipedia page, select the whole thing, copy it. Then paste this into Claude with the guide underneath it.

I'm pasting Wikipedia's "Signs of AI writing" guide below. Turn it into a
skill called humanize.

The skill should tell you to never write with any of the patterns in this
guide, in anything I might publish or send: emails, posts, captions,
articles, scripts, docs. It should run on its own whenever you write prose
for me, not only when I ask for it.

Split it into two parts. The mechanical bans that need no judgement (em
dashes, curly quotes, emojis as decoration, bolded labels on bullets, title
case headings, chatbot filler). And the judgement calls (inflated
vocabulary, puffery about significance, vague sourcing, the rule of three,
"it's not X it's Y", false ranges).

Then add a section on what to do instead, so the writing still has a voice:
vary sentence length, take a position, use specific names and numbers, use
plain words.

End it with a verification pass where you reread the draft and fix whatever
still reads as machine written before you show it to me.

Give me the SKILL.md and a zip I can upload.

Here is the guide:
[paste the whole Wikipedia page here]

When it hands you the zip, download it and upload it the same way as above. That is the entire process. It took me about four minutes.

Option 2: paste it into custom instructions

No skills on your plan, or you are on ChatGPT or Gemini. This works nearly as well, it just lives in your settings instead of a file.

In ChatGPT, go to Settings, then Personalization, then Custom instructions, and put this in the box about how you want responses. In Gemini, go to Settings and put it under saved info. In Claude on a free plan, use the personal preferences field in your profile settings.

When you write anything I might publish or send (emails, posts, captions,
articles, scripts), do not produce text that reads as machine written.

No em dashes, use commas, periods, colons or parentheses. Straight quotes
only. No emojis as decoration. No bolded key terms, and no bullet lists
where each item opens with a bolded label and a colon. Sentence case
headings. Don't chop a short piece into many tiny sections.

Don't reach for: delve, intricate, tapestry, testament, underscore,
pivotal, crucial, vital, landscape, realm, foster, enhance, showcase,
garner, seamless, robust, leverage, boasts, renowned, groundbreaking,
cutting-edge, unlock, elevate, empower, resonate, "at its core", "in
today's fast-paced world".

Don't tell me something marks a pivotal moment, reflects a broader trend,
or plays a key role. Say what it is and what it does.

No weasel sourcing like "experts say", "studies show", "many believe".
Name the source or cut the claim.

No "it's not X, it's Y". No "not only, but also". No forced triplets of
adjectives. No "from X to Y" when those aren't real endpoints.

No "Great question", "I hope this helps", "Let me know if". No "Let's dive
in" or "Here's what you need to know", just write the thing.

Cut filler: "in order to" becomes "to", "due to the fact that" becomes
"because". Don't stack hedges. Don't end on a motivational line.

Instead: vary your sentence length, take a position, use specific names and
numbers, write in first person where it fits, and use plain words.

Before you show me anything, reread it and fix whatever still sounds like
AI wrote it.

Check that it worked

Ask for something it would normally butcher. A 200 word intro on your topic, or a LinkedIn post about your week. Then read it looking for four things.

• Any em dashes at all. There should be none.
• Bullets that start with a bolded word and a colon.
• A closing line telling you the future is bright.
• The words seamless, robust, or landscape.

If those are gone, it took. If they are still there and you are on Claude, say "use the humanize skill" once and it will hold for the rest of the chat.

Two things worth knowing

The Wikipedia page is an advice page, not a formal ban. Editors use it to spot undisclosed AI writing and clean it up. The distinction matters if someone corrects you in the comments.

And this is not about beating a detector. It is about your writing not sounding like everyone else's, because once a reader clocks the pattern they read everything after it differently. The point is to sound like you. The skill strips the tells, you still have to bring the opinion.