Most AI humanisers are worse than the AI.
They swap synonyms until the prose reads like it has a thesaurus concussion. Or they add a few contractions, paste in some em dashes, and call it done. Voice isn’t a surface finish. If your draft sounds like a machine wrote it, one pass of word-level substitution isn’t going to fix that.
I use Claude for first drafts. Blog posts, documentation, post-meeting notes, the occasional cover letter when I’m too tired to start from scratch. The drafts are always obviously AI. Every paragraph closes on an -ing clause pretending to add depth. Things stand as testaments to other things.
So I built something to fix that instead.
What /humaniser Does
/humaniser reads your draft, runs it against twenty-five patterns that show up in AI-generated text, rewrites the problem sections, then audits its own rewrite for anything it missed. You get a final version after the audit, not a single first-pass attempt.
The patterns come from Wikipedia’s Signs of AI writing guide, maintained by the volunteers at WikiProject AI Cleanup. They spend their days reverting AI-generated articles from the site. They know what to look for better than anyone.
The twenty-five patterns sort into five groups: content (promotional language, vague attributions, superficial -ing analyses), language and grammar (copula avoidance, rule of three, false ranges, AI vocabulary words like “landscape” and “testament”), style (em dashes, boldface spam, emojis, curly quotes), communication artefacts (“I hope this helps”, knowledge cutoff disclaimers), and filler (excessive hedging, hyphenated pair overuse, generic upbeat conclusions).
Full catalogue lives in the SKILL.md. The article would be longer than the skill if I walked through all of them.
Why the Self-Audit Pass
Most humaniser tools produce one pass and stop. That’s the part that doesn’t work. Claude is good at clean prose, and it’s equally good at clean prose that still reads as AI. Removing the obvious tells is easier than removing the subtle ones.
So the skill prompts itself after the first rewrite: “What makes the below so obviously AI generated?” It answers honestly. The tidy rhythm is still there. The named citations might be plausible placeholders. The closer leans a touch slogan-y. Then it does a second pass on what it caught.
That second pass is the difference between output you can publish and output that still smells machine-made. One small bonus: the audit notes are legible to you, not just to Claude. Read a few of them and you start spotting your own patterns faster.
Why Soul Matters More Than Patterns
This is the bit I care about most, though I’ll admit the skill only half-solves it.
Removing AI tells is half the job. Clean prose with no opinions, no rhythm variation, no first person, no admitted uncertainty still reads like a press release. A Wikipedia article has no AI tells and doesn’t read like a person either. No AI tells doesn’t mean your writing has voice. Those are different problems. Actually, scrap that. They’re overlapping problems. Removing AI tells clears space; voice is what you put in the space. If you don’t have voice to put there, the skill will give you prose that’s technically clean and still reads dead. That’s an honest limit of what this kind of tool can do.
The skill has a “Personality and Soul” section that pushes for the stuff the patterns alone don’t give you. Short sentences next to longer ones. Stated opinions. Admitted uncertainty. First person when it fits. The occasional tangent. Specific emotional detail over generic descriptors.
Before:
The agents generated 3 million lines of code. The implications remain unclear.
After:
3 million lines of code, generated while the humans presumably slept. I keep thinking about those agents working through the night.
Same information. Different universe.
The Actual Problem This Solves
Most people who use AI for drafts do one of two things. They publish the obvious output, and their readers clock it and trust them a bit less. Or they edit manually, which is slower than writing from scratch and never catches everything, because the patterns are easy to miss when you’re inside your own draft.
The skill gives you a third option. A checklist run by something that doesn’t share your blind spots. Twenty-five specific patterns, applied consistently, then critiqued by the same system that produced the rewrite. It’s what editing an AI draft should have been in the first place.
I run this on every Claude-drafted thing I publish now. This article included.
How to Use It
If you’re running Cowork, /humaniser is one command. Point it at a file or paste text. A few seconds later you've got the draft rewrite, the audit notes, and the final version.
If you’re not, the skill file is plain text. Paste it into a system prompt, or fork the repo and adjust the “Personality and Soul” section to match your own register. The patterns stay the same. The voice you’re pointing at is up to you.
Built using the Cowork skill system. The /humaniser skill is available **her**e if you want to install it yourself.
Pairs well with [/handof](https://github.com/YKehinde/handoff)f if you're trying to work with Claude at scale without the work sounding like Claude did it.
Comments
Loading comments…