AI Text Humanizer — Claude Skill to Remove AI Writing Patterns

Humanizer is a free AI text humanizer skill for Claude Code and Claude Desktop. It reads a piece of writing, finds the patterns that make text obviously AI-generated, and rewrites them — without flattening the meaning or the voice. Install it once and it triggers automatically whenever you ask Claude to edit or review writing.

Version 2.5.1, MIT licensed, updated August 2026.

What it fixes

The skill checks against 27 documented patterns. The ones that catch most writing:

Pattern	What it looks like
Em dash overuse	Three em dashes in a paragraph where a comma or full stop would do
Rule of three	Every list padded to exactly three items, whether or not there are three
Inflated symbolism	“stands as a testament to”, “serves as a reminder that”
Promotional language	“rich cultural heritage”, “breathtaking”, “must-visit”
Superficial -ing analysis	“…, highlighting the importance of”, “…, showcasing its versatility”
Vague attribution	“experts say”, “many believe”, “it is widely regarded”
Negative parallelism	“It’s not just X, it’s Y”
Filler transitions	“Moreover”, “Furthermore”, “In conclusion”
Hedging vocabulary	“could potentially”, “may possibly”, “it could be argued”

How to install

1. Download the zip below.
2. Unzip it into your skills directory — ~/.claude/skills/ for Claude Code.
3. Restart your session.

The skill declares its own trigger conditions, so you don’t invoke it by name. Ask Claude to edit, review or clean up a piece of writing and it loads on its own.

How to use it

The simplest use is to paste text and ask for a humanizing pass. Two things make the output substantially better.

Give it a voice sample. The skill supports voice calibration: hand it something you wrote previously and it analyses your sentence rhythm, vocabulary and habits before rewriting. Without a sample it defaults to “generically natural”, which is better than AI-flat but is not you.

Ask it to explain the changes. The skill runs a final pass that asks, in effect, “what still makes this obviously AI-generated?” and revises again. Asking it to list what it changed and why turns a one-off edit into something you learn from — the second time you write that paragraph, you write it properly.

Why AI writing has a tell

The patterns are not random. Models are trained toward text that reads as balanced, complete and inoffensive, which produces specific artefacts: hedging before a claim, symmetry where none exists, and a compulsion to summarise what was just said.

The rule of three is the clearest example. Three items reads as complete and rhythmic, so a model reaches for three whether the world contains three or not. Once you notice it, you cannot stop seeing it — and neither can an editor reading your work.

The em dash case is subtler and worth understanding before you strip them out. Em dashes are legitimate punctuation; plenty of good writers use them. The tell is density and uniformity — several per paragraph, all doing the same parenthetical job. This skill targets the pattern, not the punctuation mark, which is why it is more useful than a find-and-replace.

Where it falls short

This is an editing tool, not a detector-evasion tool. It makes writing better by removing patterns that genuinely weaken it — but nobody should market it as a way to beat AI-detection software, and the results in that direction are unreliable at best. Detection tools are noisy in both directions, and text edited for quality can still be flagged, while genuinely human writing gets flagged all the time.

It also cannot add substance. If the underlying draft says nothing, this returns a more natural-sounding version of nothing. The patterns it removes are usually symptoms of thin thinking, and a good pass often makes the emptiness more obvious rather than less. That is useful information, but it is not what most people want from a “humanizer”.

Run it on writing that has something to say, and it earns its place. Run it on filler and it just gives you honest filler.