This project provides two LLM-agnostic, static skills—humanize and ai-check—designed to make AI-generated text sound more human and to detect AI writing signals. Grounded in over 50 peer-reviewed sources up to April 2026, these skills offer a research-backed approach for users of various LLM agents, aiming to bridge the gap between artificial and natural prose without requiring external services or model training.

How It Works

The core of this project lies in two rule-based "skills" packaged as single files, designed to be integrated directly into LLM agents. The humanize skill applies nine distinct "levers" derived from AI detection literature—such as perplexity injection, burstiness enforcement, and hedge surgery—to rewrite AI text, making it more natural and relatable. The ai-check skill performs forensic analysis, scoring text across nine signal categories, citing evidence, and providing a verdict and confidence level. This approach avoids runtime dependencies, model training, and API calls, leveraging the host LLM's capabilities to interpret and apply linguistic rules.

Quick Start & Requirements

• Primary install / run command:

  git clone https://github.com/harshaneel/humanize.git
  cd humanize && ./install.sh all
  

  This command installs both skills into the default skill directories for Claude Code, Codex CLI, and ChatGPT desktop.
• Non-default prerequisites and dependencies: Requires a compatible LLM agent (e.g., Claude Code, Codex CLI, ChatGPT, Gemini, Cursor, Aider). No other specific software or hardware dependencies are listed.
• Links: GitHub Repository

Highlighted Details

• Dual Skills: Offers both humanize for rewriting AI text and ai-check for detecting AI writing characteristics.
• LLM Agnostic: Designed to function within any LLM agent that supports skill integration, including Claude Code, Codex CLI, ChatGPT, Gemini, and others.
• Voice Matching: The humanize skill can analyze a provided sample of the user's writing to personalize the output, mimicking their specific style.
• Benchmark Performance: Tested on 25 diverse inputs, the humanize skill produced outputs consistently scored as "Human" or "Likely Human" by both the internal ai-check detector and an external cross-perplexity detector (Binoculars).

Maintenance & Community

Updates are managed via standard Git commands (git pull). Contributions are welcomed, particularly for identifying new AI tells, developing novel humanization patterns, or adding test scenarios. Community interaction is primarily through GitHub issues and pull requests.

Licensing & Compatibility

• License type: MIT License.
• Compatibility notes: The MIT license permits commercial use and integration into closed-source projects. The skills run locally within the user's LLM agent, requiring no external API calls or data transmission to third-party servers.

Limitations & Caveats

This rule-based approach does not guarantee zero AI scores on commercial, machine-learning-based detectors (like GPTZero or Grammarly), which often target RLHF artifacts not fully addressed by surface rewriting. While it humanizes the form of writing, it does not inherently humanize the substance, meaning hallucinations or lack of personal experience may persist. The effectiveness of the levers may vary across different domains, such as creative fiction or highly technical legal documents, compared to general professional and technical English.