humanizer-ru

This repository is a set of instructions, written as a plain text file, that tells an AI language model how to rewrite AI-generated text so it reads more like something a human wrote. The primary target is Russian-language text, where common AI tells are specific: the frequent long dash, officialese phrasing, and marketing buzzwords. An English version of the same rules is also included.

There is no code to install or run. The main file is a markdown document containing a detailed prompt. You either paste its contents into a chat with any LLM (ChatGPT, Claude, Gemini, and so on) as your first message, or you give the file directly to an AI agent that can read files. Then you paste the text you want rewritten. The rules instruct the model to look for specific patterns that reveal machine-generated writing: filler phrases like "it's worth noting", empty hedging that avoids taking a position, the mechanical rule-of-three structure, walls of unbroken text, and servile sign-offs. After a first rewrite, it runs a second check to catch anything that still reads like an AI wrote it.

For people using Claude Code specifically, the repo can be installed as a skill, letting you call a short command to trigger the humanizer on any text in your current session. The README notes that skill packaging is still coming in the first release, so that path is partially complete.

Optional voice calibration lets you paste one or two samples of your own writing, and the model adjusts the output to match your style. The examples in the README show a before-and-after where the AI-written version is rewritten to drop vague superlatives and filler structure, leaving something shorter and more direct.

The project is MIT licensed and is maintained by someone who publishes content about AI tools and vibe-coding on Telegram and YouTube.