The three prompts
Run them in order in the same chat. Prompt one sets the intention, prompt two gives it the actual rules, prompt three makes it stick so you never do this dance again. Works the same in ChatGPT and in Claude.
Tell it what you want
The obvious one. On its own it gets you maybe halfway — the model sounds a bit more relaxed for two paragraphs and then drifts straight back into press-release voice. Do it anyway; it sets up prompt two.
Give it the actual list
Here's the step nearly everybody misses. Wikipedia's editors have spent years cataloguing exactly how AI writing gives itself away — because they have to spot it every day. That catalogue is public. You hand it over and say "don't do any of this".
Make it permanent
Right now this only lives in one chat. Next conversation, you're back to square one. Turn it into a skill and it loads itself every time.
Bonus: audit something you already wrote
Got a draft that feels off but you can't say why? This one marks it up rather than rewriting it, so you learn the tells yourself.
The list nobody reads
The page is Wikipedia:Signs of AI writing, maintained by the volunteers of WikiProject AI Cleanup — people who wade through undisclosed AI submissions daily and have got very good at spotting them. It's not a policy document and it's not trying to sell you anything. It's a field guide, and it's the most useful thing on this subject anywhere on the internet.
A taste of what's on it, so you know what you're handing over:
Negative parallelism
"It's not just X, it's Y." Once you've seen it you'll see it eleven times a day. Models adore this construction and humans rarely reach for it twice in one paragraph.
The rule of three
Three neat items where two or four would be honest. "Practical, powerful and proven." Real speech isn't that tidy — the third item is usually there for rhythm, not meaning.
Inflated significance
Everything is described as marking a shift, reflecting a broader trend, or underscoring the importance of something. A human just says what happened.
Vague attribution
"Experts say", "many argue", "it is widely regarded as". No named source, because there isn't one. This is the tell that actually damages your credibility.
The wrap-up nobody asked for
The closing paragraph about challenges and future prospects. You didn't ask for a conclusion. It wrote one anyway, because it's been trained on essays.
Stock phrases
"Stands as a testament to". "Plays a vital role in". "Underscores the importance of". Individually harmless. Stacked up in one page, they're a signature.
Why this works better than "write more casually"
"Sound more human" is a vibe, and models are hopeless with vibes. A named list of twelve specific patterns is a rule set — something a model can genuinely check itself against. You're swapping an adjective for a specification, and that's the whole trick.
Before and after
Same brief both times: a short paragraph introducing a workshop on AI for small business owners.
Before
"In an era where artificial intelligence is rapidly transforming the business landscape, small business owners face both unprecedented opportunities and significant challenges. This workshop is designed to help you chart a course through this complex terrain with confidence, clarity and competence. It's not just about learning the tools — it's about fundamentally reimagining how your business operates. Experts agree that early adopters will be best positioned to thrive in the years ahead."
After
"Most small business owners I meet have opened ChatGPT once, felt vaguely disappointed, and closed it again. Fair enough. Nobody showed them what it's actually for. This workshop is three hours of doing, not watching — you'll leave with two things running in your business that weren't running this morning. Bring the task you hate most. We'll start there."
Count what changed. The before opens with "in an era where" and never recovers. It's got a rule of three ("confidence, clarity and competence"), a negative parallelism ("it's not just about X — it's about Y"), a vague attribution ("experts agree") and a future-prospects sign-off. Four textbook tells in one paragraph, and not a single concrete detail anywhere.
The after says who it's for, what happens, how long it takes, and what to bring. It's shorter. It's also the only one of the two that would make somebody sign up.
The test that actually matters
Read it out loud. Not skim it — say it. If you'd never say that sentence to a friend in a kitchen, it isn't your writing yet. This test costs nothing and beats every detector on the market.
Make it permanent
A skill is a folder with a SKILL.md file in it that the model reads before it starts working. Set it up once and every future chat already knows the rules — no re-prompting, no forgetting halfway through a long draft.
Get the file
Run prompt three above and it writes the SKILL.md for you, tailored to the checklist it extracted.
Put it somewhere the model looks
Claude Code or Claude Desktop: save it as ~/.claude/skills/humaniser/SKILL.md and restart. On claude.ai or Cowork: Customize → Skills → Add → Upload a skill. In ChatGPT, the same rules go into your custom instructions or a Project's instructions.
Now build the second one
The humaniser stops you sounding like a robot. It doesn't make you sound like you. That's a separate skill, and it's the one that actually matters.
Two skills, one after the other
Humaniser removes the tells. Brand voice adds you. Run only the first and you get clean, competent writing that could belong to anyone — which is better than AI slop, but it still isn't yours.
The honest catches
This genuinely works. Here's what I'd want you to know before you go and tell someone your writing scores 0% AI.
Detector scores prove almost nothing
I've had text come back at 0% AI-generated, and that's a nice feeling — but AI detectors are unreliable in both directions. They flag entirely human writing as machine-made (it happens to non-native English speakers constantly, and to anyone with a tidy writing style), and they wave through AI text that's been lightly edited. Wikipedia's own guide opens by warning people off relying on them. Treat a score as a smell test, never as evidence, and never send one to a client as proof of anything.
Some of that Wikipedia page isn't about you
It's written for spotting AI in Wikipedia articles, so a chunk of it is house-style specific — title case in headings, table formatting, citation conventions. Prompt two tells the model to skip those, but glance at the output checklist. If it's started refusing to use bold text anywhere, it's over-applied a Wikipedia rule to your newsletter.
Humanised is not the same as true
All of this changes how the writing sounds. None of it changes whether the writing is right. A confidently wrong sentence in a warm conversational voice is more dangerous than one in stiff corporate prose, because you're less likely to squint at it. Fact-check exactly as hard as you did before.
Don't strip out your own style
Em dashes are on the list, and I use them constantly — as you may have noticed. The tell isn't the em dash, it's five of them in one paragraph doing the same job. Same with lists of three: they're fine, they're just not fine four times a page. Aim for sounding like yourself, not for scoring zero on a checklist.
It drifts on long jobs
Halfway through a 2,000-word draft the model slips back into old habits without telling you. That's the whole argument for making it a skill rather than a prompt — but even then, on anything long, drop in "check that against the humaniser rules and fix what's slipped" before you accept the final draft.
Ready to go deeper?
Here's where to go if you want more of this.
Wright Mode Membership
Join a community of women entrepreneurs implementing AI and automation in their businesses. Live calls, templates, and ongoing support.
Claude Masterclass
Learn how to use Claude like a pro. From prompting fundamentals to building real workflows that save hours every week.
Claude Code Masterclass
Go beyond the chat interface. Build automations, process data, and create tools with Claude Code — no developer experience needed.