avoid-ai-writing: a portable writing skill for AI agents that audits and rewrites content to remove robotic AI-isms
What it solves
It addresses the problem of "AI-isms"—predictable writing patterns, robotic phrasing, and structural tells that make AI-generated content easily identifiable and often tedious to read. It provides a systematic way to audit and rewrite text to make it sound more human and natural.
How it works
The project is implemented as a portable "skill" (a set of structured instructions in a SKILL.md file) that can be integrated into various AI agents like Claude Code, OpenClaw, and Cursor. It uses a multi-layered approach to detection and editing:
- Three Operational Modes: It can Rewrite (flag and fix), Detect (flag without changing), or Edit (targeted in-place file changes).
- Structured Audit: Instead of a simple prompt, it performs a structured analysis returning identified issues, the rewrite, a change summary, and a second-pass audit to catch lingering patterns.
- Tiered Vocabulary: It uses a 112-entry word replacement table divided into three tiers based on frequency and clarity, replacing words like "leverage" with "use."
- Pattern Catalog: It identifies 61 distinct pattern categories, including significance inflation, copula avoidance, chatbot artifacts, and structural uniformity.
- Voice Profiles: Users can apply optional profiles (e.g., casual, professional, technical) to control the tone of the output.
Who it’s for
It is designed for users of AI agents (such as Claude Code, Cursor, or Hermes) who want to produce high-quality, human-sounding content and remove the "robotic" fingerprints of LLMs from their writing.
Highlights
- Two-pass detection: Re-reads its own rewrites to ensure no AI patterns survived the first edit.
- Extensive pattern library: Covers everything from "chatbot openers" to "hashtag stuffing" and "wall-of-text replies."
- Cross-platform compatibility: Works across a wide range of tools including Claude Code, OpenClaw, Cursor, Windsurf, and ChatGPT Custom GPTs.
- Non-vibes-based: Relies on a specific replacement table and defined categories rather than generic "make it human" instructions.