This project provides a curated set of 21 writing rules designed to improve the output of AI coding and writing agents, making their prose more professional and less prone to common AI "tells." It targets developers, researchers, and power users integrating AI agents into their workflows, offering a mechanism to enhance the quality and credibility of generated technical documentation, code explanations, and research papers. The primary benefit is a significant reduction in stylistic artifacts, leading to more human-like and trustworthy AI-generated text.

How It Works

The core of agent-style is a set of 21 rules, divided into 12 canonical rules derived from classic writing authorities (Strunk & White, Orwell, Pinker, Gopen & Swan) and 9 field-observed rules identified from AI output patterns between 2022-2026. These rules are integrated directly into the AI agent's generation process, acting as guardrails during text creation rather than a post-hoc linter. This approach aims to prevent stylistic errors from occurring in the first place, leading to more consistent and higher-quality output across various AI models and platforms.

Quick Start & Requirements

Installation is straightforward via package managers: pip install agent-style for Python users or npm install -g agent-style for Node.js users. The primary requirement is an AI agent capable of integrating external commands or configuration files, such as Claude Code, Codex, GitHub Copilot, Cursor, or Aider. Setup involves enabling the style rules for the specific agent using commands like agent-style enable <adapter_name>, which modifies project configuration files. An optional style-review skill provides a second-pass audit and revision capability.

Highlighted Details

• Demonstrated significant reduction in AI-generated "tells" across multiple flagship models (e.g., 82% reduction on Gemini 3 Flash).
• Combines established writing principles with AI-specific failure mode observations for comprehensive style enforcement.
• Offers seamless integration with a wide array of popular AI coding and writing assistants via dedicated adapters.
• Includes an opt-in style-review skill for deterministic auditing and optional revision of generated text.

Maintenance & Community

The project is maintained by Yue Zhao. Contributions are welcome, particularly for canonical rules that must cite established writing authorities or empirical agent-instruction research. Users can track development progress and release history in the CHANGELOG.md file.

Licensing & Compatibility

The project employs a dual-license strategy: content files (like RULES.md, SOURCES.md) are licensed under CC BY 4.0, requiring attribution. Code files and scripts are licensed under MIT. This dual licensing allows for broad compatibility while ensuring proper attribution for content.

Limitations & Caveats

While "soft enforcement" at generation time significantly reduces stylistic issues, some AI "tells" may still appear in initial drafts, especially for longer prose, necessitating the use of the style-review skill for a polished output. Support for several AI agents is planned for v1.1, indicating that current adapter coverage may not be exhaustive.