ilyautov/humanizer-ru

What it solves

AI-generated Russian text often suffers from "bureaucratese" (kantselyarit), English syntax calques, and predictable patterns that make it sound artificial. This project provides a tool to remove these 64 specific AI markers—including 21 "hard bans" on common clichés—to make Russian AI text sound more natural, human, and less detectable by AI scanners.

How it works

Instead of completely rewriting text, the tool performs local editing. It identifies specific markers across 14 categories (such as content, language, style, and tone) and applies targeted corrections. It can calibrate itself to an author's specific voice by analyzing samples of their writing to maintain rhythm and vocabulary. The system includes a deterministic scanner (written in Python) that can be used for auditing text without changing it.

Who it’s for

Users of AI agents like Claude Code, Cursor, Codex, and Copilot who generate Russian content and want to remove the "robotic" feel. It is also suitable for writers and marketers who need to ensure their AI-assisted text doesn't trigger AI detectors or sound overly formal.

Highlights

• Extensive Marker Library: Tracks 64 AI generation patterns and 21 strictly forbidden clichés.
• Multi-Agent Support: Integrates as a skill for Claude.ai, Claude Code, Cursor, Codex, and other agents via the SKILL.md standard.
• Four Operating Modes: Offers full editing, audit-only (diagnostic), targeted category correction, and a silent background mode for agent-generated text.
• Voice Calibration: Ability to align replacements with the author's personal writing style and tone.
• Client-Side Scanner: Includes a web-based and Chrome extension scanner that processes text locally in the browser.