This project addresses the challenge of identifying and mitigating patterns characteristic of AI-generated text. It offers a practical skill for AI assistants, enabling users to audit and rewrite content to sound more human. The primary benefit is the production of natural, nuanced prose by systematically removing common AI writing tells, making it valuable for content creators, developers, and researchers working with AI-generated text.

How It Works

The skill operates in two distinct modes: "Rewrite" (default) and "Detect." The rewrite mode identifies AI patterns, rewrites the text to eliminate them, and employs a second pass to catch any residual AI-isms. The detect mode flags patterns without altering the original text, suitable for auditing or when content modification is undesirable. Its approach is distinguished by a structured audit process, providing specific flagged text, rewrites, change summaries, and a second-pass analysis. This is supported by a 109-entry, tiered word replacement table and analysis across 36 distinct pattern categories, moving beyond subjective "vibes" to a systematic, data-informed methodology.

Quick Start & Requirements

Installation for Claude Code involves cloning the repository into ~/.claude/skills/ or copying SKILL.md. For OpenClaw, users can install via clawhub install avoid-ai-writing or clone into ~/.openclaw/skills/. The skill can be triggered by instructing the assistant to "Remove AI-isms from this post" or similar commands. A web application is available at https://avoid-ai-writing-app.vercel.app, which requires burning a community-created $avoid token. The skill requires an agentskills.io-compatible agent environment.

Highlighted Details

• Features a 109-entry word replacement table across three tiers, designed to reduce false positives while effectively flagging AI vocabulary.
• Analyzes content across 36 pattern categories, encompassing language, structure, rhythm, and communication styles.
• Employs a two-pass detection system to ensure comprehensive identification and correction of AI writing patterns.
• Offers a "Detect" mode that flags issues without rewriting, categorizing them by severity (P0/P1/P2) for nuanced auditing.

Maintenance & Community

The project is authored by Conor Bronsdon, with pattern research informed by contributions from Pangram Labs, Wikipedia, and other humanizer skill developers. A community has formed around the project, evidenced by the creation of a Solana token ($avoid) and a dedicated web app for auditing. Community resources include a Telegram channel (t.me/avoidaiwriting) and an X community.

Licensing & Compatibility

The project is released under the MIT license, which permits broad use, modification, and distribution, including for commercial purposes and integration into closed-source applications. It is designed for compatibility with agents conforming to the agentskills.io standard.

Limitations & Caveats

The provided README does not explicitly detail limitations, alpha status, or known bugs. The effectiveness of the skill is dependent on the comprehensiveness of its defined patterns and word lists, which may not encompass all future or highly sophisticated AI writing styles.