How to Make AI Writing Sound Human (Free Claude Skill)

How to Make AI Writing Sound Human, With a Free Skill

We asked Claude for a LinkedIn post about the first year of running a small bakery. The draft opened like this: "In January, I unlocked the door of my bakery for the first time. Ten months later, I learned more than any business course would have taught me." Six bold lesson titles followed, starting with "Consistency beats creativity."

If you spend any time on LinkedIn, you have read that post this week, probably more than once.

Then we sent one line into the same chat. It loaded a free skill called humanize, and Claude wrote the post again. The new version opened: "I opened my bakery in January. Ten months in, a few things surprised me." Same bakery, same year, and this time it reads like someone telling you about it. This guide explains why AI writing sounds the way it does, then walks you through using the skill in Claude, Claude Code or Codex, step by step.

Short answer: to make AI writing sound human, cut the patterns that give it away, delete the closing moral, swap slogans for real moments and add details only you know. The free humanize skill runs those passes for you in Claude, Claude Code or Codex.

The patterns are well documented: contrast lines like "it's not X, it's Y", lists that always come in threes, inflated words like "pivotal", and stock words like "delve".
The skill is free and public on GitHub under an MIT license. One line loads it.
It works from Wikipedia's "Signs of AI writing" guide and StoryScope, a 2026 study of 61,608 stories by people and AI.
It never adds facts. If your draft has nothing real in it, it asks you for one detail.
It does not promise to beat AI detectors, and no tool honestly can.

Video: How to Make AI Writing Sound Human, With a Free Skill — Prefer video? Watch the short version, or open it on Instagram or YouTube.

Why AI writing sounds like AI

You can usually feel it before you can name it. The tells are specific, though, and two groups have written them down.

The first is Wikipedia's own editors. WikiProject AI Cleanup keeps an advice page called Signs of AI writing, a field guide built from real AI text they found in articles. It has entries for "undue emphasis on significance, legacy, and broader trends" (the "pivotal moment" and "testament to" habit), for negative parallelisms such as "not just X, but also Y", for the rule of three, and for a high density of "AI vocabulary" words, with "delve" on the list.

The guide is careful about what it proves. It says plainly that not all text with these signs is AI-generated, and it warns that AI detectors have "non-trivial error rates." Its entry on em dashes now sits under historical indicators, dated 2022 to September 2026, so even this list changes as the models do.

The second group is a research team. In StoryScope, Jenna Russell, Rishanth Rajendhran, Chau Minh Pham, Mohit Iyyer and John Wieting collected 61,608 stories of about 5,000 words each. People and five AI models (Claude Sonnet 4.6, GPT-5.4, Gemini 3 Flash, DeepSeek V3.2 and Kimi K2.5) wrote them from the same 10,272 prompts. The team scored 304 features of how each story was put together, and those features alone told human from AI with a 93.2% macro-F1 score.

Two of their findings show up all over business writing. AI narrators explained the theme of the story 77% of the time, against 52% for people. And AI showed feelings through the body ("my chest tightened") in 81% of stories, against 38% for people. In a LinkedIn post the first habit is the moral in the last line, the sentence that tells you what you just learned.

One caveat: StoryScope studied fiction. The paper never tested posts or emails, so applying its findings there is our judgment, and the skill's.

The bakery post, before and after

Our request to Claude was simple: write a LinkedIn post about the lessons learned in the first year of running a successful small bakery, opened at the start of this year. The first draft was the one above, with its unlocked door and its row of bold lessons.

The second draft kept the same material and changed what it did with it. Line them up and you can see the swap from slogan to moment:

First draft (lesson title)	After the humanize skill
Consistency beats creativity.	I came in wanting a menu full of experiments. My customers wanted the same thing done well every morning.
Your regulars are your real marketing team.	Word of mouth did more than any ad.
Say yes to community.	Local events, school fundraisers and teaming up with shops nearby brought in people who'd never have found us otherwise.

The pairing is our reading of the two posts. What is not in doubt is the ending. The first draft closed on its list of lessons. The second closed on a question to the reader: "If you run a small business, what caught you off guard in year one?"

One thing to keep straight. The bakery is imaginary, so both posts describe a year that never happened. The skill made the made-up year read better, and it could only use what Claude had already written. A real post gets its life from things only the writer knows, like what went wrong in week two. You have to put those in.

What the humanize skill is

Our founder, Allen, wrote humanize as a plain-English instruction file and published it in a public GitHub repo, allenanant/humanize, under an MIT license. The whole skill is one file, SKILL.md, about a thousand words long. There is no app and no account to create. Your AI reads the file and follows it.

It runs four passes over your text, in order: the words, the shape, the life and the voice. It gives you the new version first, then a short list of what it changed. It leaves code, data, quotes from real people and legal text alone, and it never tells you a text is "now human" or "undetectable".

Before you start

An AI to run it in. Claude works on any plan, including Free. Claude Code needs a paid Claude plan. Codex comes with ChatGPT plans, including Free with usage limits.
For Claude on the web: code execution switched on, which skills need. On a personal plan, that is under Settings, then Capabilities.
To keep the skill in Claude Code or Codex: Git, which the install command uses. If you would rather not touch a terminal, Claude Code can run the command for you.
A piece of your own text to test it on: a post, an email, a product page. Something you would really publish works better than a sample.
Time: loading the skill takes one message, and saving or installing it takes a few clicks or one command.

Step 1: load the skill with one line

This works in Claude, Claude Code and Codex, and you do not install anything to try it.

Copy the line below.
Open a new chat or session. A fresh chat keeps earlier instructions from mixing with the skill.
Paste the line and send it. The AI opens the file on GitHub and reads it.
Wait for the reply. The skill tells it to answer with one line when you have not given it text yet: "Ready. Paste the text you want humanized."
If it says it cannot open links, open SKILL.md on GitHub, copy all of it, paste that in and add "follow this as your humanize skill".

Read https://raw.githubusercontent.com/allenanant/humanize/main/SKILL.md and follow it as your humanize skill. If you can install skills, install it. Then ask me for the text to humanize.

Prefer a one-click copy? The same line is on our free prompt page.

Step 2: save it in Claude on the web

Loading the skill in a chat lasts for that chat. Saving it puts it in your Claude account, so every new chat can use it.

Switch on code execution. On a personal plan, open Settings, then Capabilities, and turn on code execution. On Team and Enterprise plans an admin controls this.
Send the one line in a new chat. Because it says "if you can install skills, install it", Claude can offer to save it. When we ran it, Claude opened a pop-up showing the whole skill, with a Save button.
Read the skill in the pop-up. It is the same SKILL.md as on GitHub, so you can check exactly what you are adding.
Click Save. It is stored in your account as /humanize.
Use it. Claude decides when a skill fits your request, so it can pick humanize up when you ask for writing. To make sure, type /humanize in the chat box before your request.

Your saved skills are listed under Customize, then Skills. If the pop-up never appears, Anthropic's help page on skills describes the other route: package a skill folder as a ZIP and upload it from the same Skills screen with the + button.

Step 3: install it in Claude Code or Codex

In Claude Code and Codex, a skill is a folder on your computer with a SKILL.md file inside. The install is one command that copies the repo into the right folder.

Open a terminal. On a Mac that is the Terminal app. On Windows, use Git Bash, which comes with Git for Windows.
Run the command for your tool:

Claude Code:

git clone https://github.com/allenanant/humanize ~/.claude/skills/humanize

Codex:

git clone https://github.com/allenanant/humanize ~/.codex/skills/humanize

Check it landed. The file should now be at ~/.claude/skills/humanize/SKILL.md (or the Codex path). If you prefer not to use a terminal at all, open Claude Code and say: "clone https://github.com/allenanant/humanize into my personal skills folder".
Start using it. Claude Code picks up a new skill in ~/.claude/skills without a restart; if the skills folder did not exist when your session started, run /reload-skills. Type /humanize to call it directly. In Codex, type $humanize or pick it from /skills.

A note for Codex users. OpenAI's current skills page lists $HOME/.agents/skills as the folder for personal skills. If Codex does not see humanize, clone it into ~/.agents/skills/humanize instead and restart Codex.

Step 4: humanize your first piece of text

Say humanize this and paste your text underneath, in the same message.
Read the rewrite. It comes first, on its own.
Read the change list under it. The skill reports counts per pass, in a form like "Words: cut 3 stock words and 2 not-X-but-Y lines. Shape: removed the closing moral." This tells you what it touched, so you can check it.
Answer its question if it asks one. When your draft has no real detail to work with, the skill asks for one: a name, a number, a moment. Give it a true one. This is where most of the improvement comes from.
Read the result out loud. If you would not say a line to a customer's face, cut it or ask for another pass on that line.

Two other ways to call it:

humanize --check followed by your text. It only points out what reads as AI and changes nothing. Use it on your own writing, or when you want to fix things yourself.
humanize <file> in Claude Code or Codex. It rewrites that file in place, then gives the same short list.

What the four passes do

Pass 1: the words

This pass checks 20 patterns based on Wikipedia's guide. Among them: "not X, but Y" contrasts that add drama instead of facts, one-line closers like "And that changes everything", sayings that sound deep and mean nothing, lists of three out of habit, a dash between every pair of clauses, inflated importance ("a testament to"), stock AI words, vague authority ("studies show" with no study named) and chatbot leftovers like "I hope this helps".

Pass 2: the shape

Text can pass every word check and still read as AI because of how it is built. This pass, based on StoryScope, deletes the line that explains what it all means, says feelings plainly ("I was nervous") instead of acting them out through the body, and puts back what really happened when a story resolves too neatly, such as the false start or the person who helped. It only uses what you or the text supplied.

Pass 3: the life

Clean text can still be boring. This pass opens on the most interesting true thing in the text, swaps each general claim for the concrete picture behind it, mixes long sentences with short ones and ends on something specific or a real question. If there is nothing real to work with, it asks you instead of inventing.

Pass 4: the voice

A starter set of rules that wins over the other three: no em dashes, no filler openers, short sentences, plain words, and keep the writer's own pet phrases.

Make it sound like you

Out of the box, the skill does not know your voice. Pass 4 is meant to be replaced.

Open your copy of SKILL.md. In Claude Code that is ~/.claude/skills/humanize/SKILL.md. For Claude on the web, download the file from GitHub and edit it in any text editor.
Find the section headed "Pass 4: the voice".
Replace the starter bullets with your own rules. Good ones are specific: words you never use, how you open an email, how you sign off, whether you write "we" or "I", how long your paragraphs run.
Save it. In Claude Code the change applies without a restart. For Claude on the web, paste your edited file into a chat instead of the one line, or upload it as a ZIP from the Skills screen.
Test it on something you wrote by hand. If the rewrite still does not sound like you, add a rule for the thing that is off.

Variations worth trying

Emails: "humanize this email, keep it under 120 words". Shorter limits force the skill to keep only the lines that carry something.
Landing pages and product pages: run it section by section. A whole page at once gives you a long change list that is hard to check.
Scripts for video or audio: read the result aloud before you record. Pass 3 mixes sentence lengths, which helps spoken lines.
Your own drafts: humanize --check on something you wrote without AI is a quick way to learn the patterns. Most of us use a few of them.

What it costs, and where it works

The skill costs nothing. What you pay for, if anything, is the AI you run it in.

Claude on the web: Anthropic's help page says skills are available on Free, Pro, Max, Team and Enterprise plans, with code execution switched on. If you only want to try it once, pasting the one line into a chat is enough.
Claude Code: included in every paid Claude plan. Pro is $17 a month on yearly billing or $20 month to month, according to Claude's pricing page.
Codex: OpenAI includes it in ChatGPT plans from Free upward, with usage limits on the free tier.
Other chatbots: SKILL.md is a set of written instructions, so you can paste it into another AI. We have only tested it in Claude.

What it will not do

It will not make your text "undetectable", and the skill is written to refuse that claim. Detectors are a poor target anyway. OpenAI launched its own AI text classifier in January 2023 and reported that it caught 26% of AI-written text while flagging human writing as AI 9% of the time. On 20 July 2023 it withdrew the tool, saying it was "no longer available due to its low rate of accuracy." If your goal is to get an assignment past a checker, this skill was not made for you.

It also sticks to the facts in front of it. That rule protects you, because an AI that invents a customer story to make your post livelier has just put a lie under your name. It also means a vague draft stays thin until you give it something real.

A flagged pattern is worth a second look and proves nothing on its own. Wikipedia's editors say as much. Plenty of people wrote "a pivotal moment" long before chatbots did.

Troubleshooting

The AI says it cannot open the link: paste the full SKILL.md text instead of the one line.
No Save pop-up in Claude: check that code execution is on under Settings, then Capabilities. On a Team or Enterprise plan, ask your admin whether skills are enabled.
Claude Code does not list /humanize: check that the file is at ~/.claude/skills/humanize/SKILL.md, not one folder deeper, then run /reload-skills.
Codex does not see it: move it to ~/.agents/skills/humanize and restart Codex.
The rewrite barely changed: your draft may be clean at word level and thin on detail. Give it two or three true things, such as your first price or the client who said no, and ask again.
It touched a quote: tell it the text is a real person's words. Quotes are on the skill's never-touch list, and it should put them back.

We run our own scripts and posts through a fuller version of the same skill before they go out. This guide went through it as well.

Who this is not for

Anyone who wants to publish without reading what the AI wrote. The skill reports its changes so you can check them, and that only helps if you read them.
Legal text, code and quotes from real people. It leaves them alone by design.
Writers who already draft by hand and only use AI for research. You may get more from humanize --check than from a rewrite.

Frequently asked questions

How do I make AI writing sound more human?

Cut the patterns that give it away: contrast lines like "it's not X, it's Y", lists that always come in threes, inflated phrases like "a pivotal moment" and stock words like "delve". Delete the closing moral, swap slogans for specific moments, and add true details only you know. The free humanize skill runs these passes for you in Claude, Claude Code or Codex.

Is the humanize skill free?

Yes. It is public on GitHub at allenanant/humanize under an MIT license. You send one line to Claude, Claude Code or Codex and it reads the skill from GitHub. You can also save it to your Claude account or clone it into your skills folder to keep it.

Will humanized text pass AI detectors?

The skill does not promise that, and no tool can honestly promise it. Detectors make mistakes in both directions. OpenAI withdrew its own AI text classifier in July 2023 because of its low accuracy. The skill aims at writing people want to read.

Do Claude skills work on the free plan?

Anthropic's help center says skills are available on Free, Pro, Max, Team and Enterprise plans, with code execution turned on. On a personal plan, switch on code execution under Settings, then Capabilities. Saved skills are listed under Customize, then Skills.

Can I use the humanize skill with ChatGPT or other AI tools?

The skill is a plain-English instruction file, so you can paste SKILL.md into another chatbot and ask it to follow it. Claude, Claude Code and Codex can load it straight from the one line. We have only tested it in Claude.

What are the most common signs of AI writing?

Wikipedia's "Signs of AI writing" guide lists undue emphasis on significance and legacy, negative parallelisms such as "not just X, but also Y", the rule of three, and a high density of words like "delve". The StoryScope study adds structural habits: explaining the theme outright and describing feelings through the body.

How to Make AI Writing Sound Human, With a Free Skill