How to Clean Text from AI-Generated Clichés Using avoid-ai-writing

You can probably recognize text generated by a language model within the first two sentences. This characteristic syntax is hard to mistake for anything else. It constantly features fake emotions, admissions of being wrong, "important milestones," and endless constructions like "unlocking potential." It seems like AI is trying to sound as polite and professional as possible, but in reality it produces a bland corporate summary.

If you use LLMs for writing documentation, blog posts, or repository responses, this problem becomes a constant headache. Usually you have to manually rewrite half the response or compose cumbersome prompts. Developer Conor Bronsdon decided to solve this problem systematically and created the open-source project avoid-ai-writing.

What avoid-ai-writing Is

The project is a ready-to-use tool and instruction set made according to the agentskills.io specification. It integrates with Claude Code, Cursor, OpenClaw, Hermes, and other popular agents. Unlike typical prompts along the lines of "write in a more human tone," this skill works as a full-fledged editor.

The repository is based on research from Pangram Labs, Wikipedia editors' documentation on signs of AI-generated text, and statistical data on vocabulary diversity. The result is a tool that forces an agent to analyze text against specific checklists rather than just rephrasing it to taste.

How the Filter Works

The skill relies on several text-checking mechanics.

First, it includes a catalog of 53 pattern categories. These include fake introductory phrases, uniform sentence lengths, bullet lists made up entirely of nouns, emojis in headings, and characteristic phrases like "let's dive into the topic" or "ultimately."

Second, the project includes a replacement table with 111 entries. Words are divided into three strictness levels:

• Level 1 words are always replaced with simpler synonyms. For example, leverage becomes use.
• Level 2 words are subject to replacement if they appear too frequently within a short stretch of text.
• Level 3 words and phrases are tracked by density. If phrases like "decentralized computing" or "solution integration" accumulate in one paragraph, the tool suggests rewriting the sentence.

Third, the system uses a two-pass audit. On the first pass, the model rewrites the text and removes obvious AI-generated fluff. Then the second pass kicks in: the agent re-reads its own version and looks for what slipped through accidentally. Usually these are repeating conjunctions, smoothed-out rhythm, or introductory words that migrated from the original.

Three Operating Modes

You can use the tool in different scenarios:

1. Rewrite mode. This is the default. A draft goes in, and the output includes an analysis of found issues, a clean version of the text, a list of changes made, and the results of the re-audit.

2. Detect mode. Suitable for cases where you cannot or do not need to change the source text. The agent highlights found markers and ranks them by criticality, from P0 to P2, indicating where there is an obvious error versus just a stylistic nuance.

3. Edit mode. Used for working with files directly. The agent changes only problematic paragraphs through the Edit tool, leaving alone fragments that are already written in normal, living language.

Standalone JS Engine for CI/CD

The detector directory contains a self-contained module in pure JavaScript with no external dependencies. You only need Node.js version 18 or higher to run it. The engine can execute directly in the browser or be embedded in automated document-checking scripts.

The code analyzes text structure, entropy of function words, punctuation distribution, and calculates a final score from 0 to 100. The lower the score, the less AI-generated writing style the text contains. You can run tests with the command:

npm test

Or import the module into your JS code:

const AIDetector = require("./detector/patterns.js");
const { score, label, issues } = AIDetector.analyzeText("Ваш текст для проверки...");
console.log(`Итоговая оценка: ${score}`);

How to Install the Skill in Your Tools

Integration depends on your environment.

For Claude Code, simply clone the repository into the skills folder:

git clone https://github.com/conorbronsdon/avoid-ai-writing ~/.claude/skills/avoid-ai-writing

If you prefer using Cowork plugins, installation can be done through console commands:

/plugin marketplace add conorbronsdon/avoid-ai-writing
/plugin install avoid-ai-writing@conorbronsdon-skills
/reload-plugins

Developers working in Cursor need to download the MDC rules file into their project folder:

mkdir -p .cursor/rules
curl -o .cursor/rules/avoid-ai-writing.mdc \
  https://raw.githubusercontent.com/conorbronsdon/avoid-ai-writing/main/cursor-rules/avoid-ai-writing.mdc

Similar deployment options exist for OpenClaw, Hermes, Codex, Windsurf, Cline, and GitHub Copilot. All use the same instruction file SKILL.md.

Text Cleaning Example

To see the difference, just look at how the skill handles a typical AI-written press release.

Original text:

Certainly! Acme Analytics, a vibrant startup nestled in the heart of Boulder's thriving tech ecosystem, has secured $40M in Series B funding — marking a watershed moment for the observability landscape. The platform serves as a unified hub, featuring real-time dashboards, boasting sub-second queries, and presenting a seamless integration layer. Moreover, experts believe Acme is poised to disrupt the market. In conclusion, the future looks bright!

Edited version:

Acme Analytics raised a $40M Series B led by Sequoia. The Boulder-based startup makes an observability platform that runs queries in under a second and plugs into existing monitoring stacks without custom integration work.

The tool removed the chatbot greeting, evaluative epithets ("vibrant," "thriving"), clichés ("watershed moment"), unsourced generalizations ("experts believe"), and meaningless fillers. Only concrete facts, numbers, and the point remained.

Who Will Find the Repository Useful

The project is useful for anyone who regularly uses AI for content generation and wants to maintain a normal human style. The templates and rules from SKILL.md clearly demonstrate how to format instructions for agents so they produce quality output without verbal clutter.

The project is distributed under the MIT license. It has over 25,000 stars on GitHub and an active community that adapts the rules for different languages.