Stop your AI agent from writing like every other AI.

A free skill I prepared for AI agents. Your agent keeps its own voice, or yours. MIT licence, one-line install, no setup.

View on GitHub

Definition

write-without-ai-slop is an open-source skill for AI agents: a SKILL.md file plus reference pages that name the writing patterns to avoid. It checks a draft and lists the problems, rewrites a draft with the smallest edit that removes them, or guides an agent while it writes. It teaches no style and adds no voice. MIT licence. Version 0.5.0, first public on 2026-11-17.

What write-without-ai-slop is

AI text looks, feels and reads the same, and readers notice. The setup line before the point, the contrast that opens “not just X”, the group of three, the closing question that asks for a comment. Any one of them is ordinary. Together they give a text away.

The skill is a list of what not to do. Your agent still writes in its own voice, or in yours if you hand it a voice file or a short sample.

02Why I made it

People have their own voice, and people use AI to write content, which makes content creation easy. But each piece of content feels the same, because AI writes mostly the same way, with minor differences from model to model. Still, when a user reads it, it feels AI-generated.

I have identified the symptoms and put them into words, so that when AI is writing something, it considers how not to write. Once it knows how not to write, it can choose other, less common options. So this skill will keep evolving, and people will sound less and less like the general norm, giving users more space to find their own voices.

03See it work

Before

Here’s the thing: shipping fast isn’t just about speed, it’s about trust.

Last month our team cut the checkout bug backlog from 42 to 9 in two weeks. Not just a win for engineering, but a win for customers, for the business, and for the future.

Agree? What’s your biggest lesson?

After

Shipping fast builds trust.

Last month our team cut the checkout bug backlog from 42 to 9 in two weeks.

The numbers stayed. The setup line, the “not just X” contrast, the group of three and the bait question went. Both texts are invented, as are all the examples in the repo.

04Install

For agents that support the skills installer, one line:

npx skills add nocodeanish/write-without-ai-slop

Or copy the folder into your agent’s skills folder. In Claude Code:

cp -r write-without-ai-slop ~/.claude/skills/

That is the whole setup. The skill does not ask questions mid-task. If your agent has no skills support, paste the SKILL.md file at the start of a chat and tell the model to follow it. I have not tested that route in other apps.

The code, the changelog and the tests are in the GitHub repository.

05How it works

You do not pick a mode. The skill reads the request.

Check
Ask it to check a draft. You get a list of problems, each with the quoted line, the pattern and a short fix. Nothing is rewritten.
Rewrite
Give it a draft. You get the smallest edit that removes the problems, then a short list of what changed. If you say the text will be pasted as it is, you get only the text.
Write
An agent that is about to write follows the rules while it writes, then checks its own draft before it hands it over.

It has short pages for LinkedIn, X, email, website copy, docs and Slack, because the same pattern means different things in a Slack message and in a README.

06What it catches, and what it does not do

It catches invented facts, numbers, quotes and names. It catches chatbot leftovers such as “Certainly!”, “I hope this helps” and unfilled placeholders. It catches sameness, such as five posts that open or end the same way. And it catches patterns that matter in combination: “not X, but Y”, groups of three, setups before the point, fake meaning tacked onto the end of a sentence, stacked fragments, piled-up hedges and vague sources.

It does not teach a style
It lists what not to do and adds no voice of its own.
It does not detect AI
Detectors are unreliable. In one study ([Liang et al., 2023](https://arxiv.org/abs/2304.02819)), seven of them flagged 61% of essays by non-native English writers as AI-written, though people wrote every one. The skill never says who wrote a text.
One pattern proves nothing
It acts on combinations. A specific text with a single tell comes back unchanged.
English only
For other languages it checks only for invented facts, padding and sameness.

07Does it work?

The repository holds 57 invented test cases. Each has a request, an input text, and what a good answer must and must not do. Several are cases where the right answer is to change nothing. I ran all 57 on three models from one provider, once each. A second model then judged the answers without knowing which model wrote which.

Results on version 0.5.0
Test cases
57Four sets, all invented
Exact checks passed
171 of 171Numbers, links, code, quotes, dashes, leftovers
Cases passed, of 57
54 to 56Judged by a second model, per model

The five misses were small: two models kept one or two hashtags, one kept a lead-in label, one check finding quoted no line, and one kept a heading in a Slack message meant to be pasted as it is.

Read the numbers as a guide. The rules were shaped by earlier failures on two of the four sets, so those sets now only check for regressions. The judge is one model from the same provider as the models it judged, and I have not tested models from other providers. A gap of one or two cases between models is noise at this size. The tests and results are in the repository, with a script to run them on your own model, and where each rule comes from is listed with its sources.

08Report a miss, or add a case

If the skill changes something it should have left alone, or lets a pattern through, open an issue on GitHub with the request, the text before, the text after and what went wrong. A real miss is worth more to this project than a star. The contributing notes say how to add a test case.

I write about shipping AI-built software that holds up under real users on the blog. If you prompt Claude models directly, the Opus 5 field manual covers effort levels and migration changes.

09FAQ

Common questions.

Does it work with ChatGPT, Gemini or other models?

It is a SKILL.md file, written for agents that support skills. In other apps you can paste the file at the start of a chat and tell the model to follow it. I tested it on three models from one provider. I have not tested models from other providers, and I have not tested the paste route in other apps.

Will it get my text past AI detectors?

No, and it does not try. Detectors are unreliable, and the skill never says who wrote a text. It points at patterns a reader can see. The study behind that is under what it does not do.

Will it change my voice?

It adds no voice of its own. Give your agent a voice file or a short sample with your request, and the skill keeps your phrasing.

What does it cost, and what is the licence?

It is free. The licence is MIT.

Does it work on languages other than English?

The pattern rules are for English. For other languages it checks only for invented facts, padding and sameness.

Where do I report a bad result?

Open an issue in the GitHub repository with the request, the text before, the text after and what went wrong.