# humanizer-ru > [Русская версия: основная](README.md) · [中文](README.zh.md) Claude Code / Cowork plugin. Cleans bureaucratese and AI-style clichés out of Russian text without touching facts or the author's voice. The English [humanizer](https://github.com/blader/humanizer) won't help here. Russian AI markers are their own beast: bureaucratic noun-chains (канцелярит), English-syntax calques, missing particles like "же" and "ведь" that make Russian sound alive. [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE) [![Version](https://img.shields.io/badge/version-3.31.5-blueviolet)](https://github.com/ilyautov/humanizer-ru/blob/main/CHANGELOG.md) [![Stars](https://img.shields.io/github/stars/ilyautov/humanizer-ru?style=social)](https://github.com/ilyautov/humanizer-ru/stargazers) [![skills.sh](https://skills.sh/b/ilyautov/humanizer-ru)](https://skills.sh/ilyautov/humanizer-ru/humanizer-ru) [![npm](https://img.shields.io/npm/v/humanizer-ru?label=npm%20%C2%B7%20dsh)](https://www.npmjs.com/package/humanizer-ru) [![HOL](https://img.shields.io/endpoint?url=https%3A%2F%2Fhol.org%2Fapi%2Fregistry%2Fbadges%2Fplugin%3Fslug%3Dilya-utov%252Fhumanizer-ru%26metric%3Dstatus%26style%3Dflat)](https://hol.org/registry/plugins/ilya-utov%2Fhumanizer-ru)

humanizer-ru: fewer clichés, more meaning. AI Frontier editorial workshop

🧪 **Live demo:** [humanizer-ru.aifrontier.tech](https://humanizer-ru.aifrontier.tech/#audit). Paste a Russian text, get a 0-100 cleanliness score and highlighted markers. Same scanner as in the skill, runs in the browser. **Quick start**, one command for any agent (Claude Code, Cursor, Codex, Copilot, Cline and dozens more): ``` npx skills add ilyautov/humanizer-ru ``` The CLI detects your installed agents and asks where to put the skill. In Claude Code you can also install it as a plugin, then updates arrive via `/plugin`: ``` /plugin marketplace add ilyautov/humanizer-ru /plugin install humanizer-ru@ilyautov-plugins ``` Claude.ai, DeepSeek Harness and team rollout are covered in [Install](#install). 📖 **Docs & write-ups (RU):** [humanizer-ru.aifrontier.tech](https://humanizer-ru.aifrontier.tech/): do AI detectors work on Russian, the 67 markers, plagiarism vs AI detection. ## What you get 67 patterns across 16 categories: канцелярит, English calques, emotional sterility, persuasion tricks, information rhythm, hedging specifics, plus the 2025-2026 stylistic fingerprints (jagged-meditation single-word sentences, pseudo-Socratic Q-A chains, decorative emoji per list item, pseudo-therapeutic register) and the 2026 formulas (inanimate subject, Title Case headings, mid-sentence truncation, negation triad), plus a discourse layer (explicit final moral, portrait-style character introduction, emotion conveyed only through the body, seamless causal chains, strictly linear chronology, no direct reader address: narrative signals that survive stylistic rewriting, per StoryScope / COLM 2026). 21 hard-banned constructions that scream "GPT wrote this", including em-dashes (detectors count their frequency). A research-backed section on how detectors actually work (perplexity, burstiness, morphology) with verified numbers from DivEye, PIFE, AINL-Eval 2025 (every citation checked, see SOURCES.md). Five article formulas. Voice calibration. Local editing instead of rewriting: mark the exact spot, strip the wrapper around the claim, verify the result against the source (who did what, certainty, conditions, numbers with units). Nothing is added for "liveliness": no invented facts, emotions or particles; a 195-sample blind run (eval/LOCAL_EDIT_CHECK.md) showed that voice quotas produce foreign tone and staged rhetoric. ## Try it without installing The [site's front page](https://humanizer-ru.aifrontier.tech/#audit) runs the same scanner in JavaScript: paste a text, get a 0-100 cleanliness score, highlighted markers and the install command for your agent. Everything runs in the browser, nothing is uploaded. Rules are exported from `markers.py` by `scripts/export_web_rules.py`; `scripts/test_web_parity.py` keeps the web and Python engines in step in CI. The same scanner ships as a [Chrome extension](https://chromewebstore.google.com/detail/iaelpnagdohdahkcahippeadohpgobad): select text on any page, pick "Проверить на шаблонные обороты" in the context menu, and the popup shows the score and highlights. Two permissions only, context menu and local storage; no host access, no network. Source in `extension/`, built by `scripts/build_extension.py`. ## Install Three deployment channels: upload to Claude.ai web UI, roll out across an organization, or install into local agents (Claude Code, Cowork, API). ### 1. Claude.ai (Web UI) 1. Download the packaged skill: [humanizer-ru.zip](https://github.com/ilyautov/humanizer-ru/releases/latest/download/humanizer-ru.zip) 2. Open Claude.ai → **Settings** → **Capabilities** → **Skills**. 3. Click **Upload skill** and select the ZIP. Do not use `archive/refs/heads/main.zip`: the uploader expects the skill folder at the top level of the archive, and in the repo archive it sits inside `humanizer-ru-main/skills/humanizer-ru/`. To build the archive yourself, zip the skill folder itself: ```bash git clone https://github.com/ilyautov/humanizer-ru.git cd humanizer-ru/skills zip -r ../../humanizer-ru.zip humanizer-ru -x '*/__pycache__/*' ``` ### 2. Organizations (Enterprise & Team) Workspace admins can roll the skill out to the whole team via **Admin Console → Workspace Skills → Add skill**. Upload the same ZIP, no per-user installation needed. ### 3. Claude Code, Cowork, API (local agents) **Plugin marketplace** (recommended): ``` /plugin marketplace add ilyautov/humanizer-ru /plugin install humanizer-ru@ilyautov-plugins ``` **skills.sh CLI** (universal across Claude agents): ```bash npx skills add https://github.com/ilyautov/humanizer-ru/tree/main/skills/humanizer-ru ``` The CLI drops `SKILL.md` into `~/.claude/skills/humanizer-ru/` and registers the skill on [skills.sh](https://skills.sh/ilyautov/humanizer-ru/humanizer-ru). The short form works too: `npx skills add ilyautov/humanizer-ru` discovers the skill inside the repository and offers it for install; add `--skill humanizer-ru` to install without prompts. **API (`/v1/messages` and equivalents):** pass the skill via the `container.skills` parameter. See your client's docs. **Manual:** ```bash git clone --depth 1 https://github.com/ilyautov/humanizer-ru /tmp/humanizer-ru mkdir -p ~/.claude/skills cp -r /tmp/humanizer-ru/skills/humanizer-ru ~/.claude/skills/ ``` Copy the whole folder, not just SKILL.md: the skill ships with a deterministic scanner, `scripts/scan.py` (the machine half of Audit mode; needs `pip install razdel pymorphy3`, and the skill gracefully falls back to a manual audit without them). ### Package install smoke Before a release, build the ZIP with the same script used by CI, then run the install smoke against that archive. The smoke check copies `skills/humanizer-ru` into a clean temporary skills directory, runs the scanner from the installed copy, and checks that the ZIP has no `.DS_Store`, `__pycache__`, or `*.pyc` files: ```bash python3 scripts/build_release_zip.py --output dist/humanizer-ru.zip . python3 scripts/install_smoke.py . --zip dist/humanizer-ru.zip ``` Expected result: ```text RESULT: PASS build-release-zip RESULT: PASS install-smoke ``` ### 4. Codex CLI (OpenAI) Codex uses the same Agent Skills format, so no separate build is needed: ```bash git clone --depth 1 https://github.com/ilyautov/humanizer-ru mkdir -p ~/.codex/skills cp -r humanizer-ru/skills/humanizer-ru ~/.codex/skills/ ``` Or from inside Codex via `skill-installer` with the path `ilyautov/humanizer-ru/skills/humanizer-ru`. Restart Codex after installing; invoke with `$humanizer-ru` or let it auto-trigger. For a per-project install, put the same folder under `.codex/skills/` in your repository. ### 5. Other agents (shared SKILL.md standard) The Agent Skills format is now cross-platform, so humanizer-ru runs beyond the packaged stacks. Officially packaged and verified: Claude Code, Codex CLI, Cursor, Gemini CLI, DeepSeek Harness. Other agents read the same `SKILL.md`, no separate build needed. | How it connects | Agents | What to do | |---|---|---| | Read `SKILL.md` natively | GitHub Copilot, Cline, Roo Code, Kilo Code, Goose, OpenCode, OpenWork, Kimi Code CLI, OpenClaw, OpenHuman, Hermes | Copy the `skills/humanizer-ru` folder into the agent's skills dir (e.g. `~/.hermes/skills/`, `~/.kimi/skills/`, `.agents/skills/`) | | Installer conversion | Windsurf, Trae, Junie | Install via their skill installer, pointing at `ilyautov/humanizer-ru` | | Manual paste | Zed, Aider, Continue.dev | Paste the `SKILL.md` body into the agent's rules or instructions file | Generic path for native readers: drop the skill folder into the directory the agent scans (usually `/.agents/skills/` or `~/.config/agents/skills/`) and restart it. Same activation triggers. > Several of these agents (OpenClaw, Kimi, Hermes) have public skill registries (ClawHub and similar). Listing there adds reach with no extra code: the format is shared. ### 6. DeepSeek Harness (dsh) DeepSeek Harness reads the same Agent Skills format. Two ways to install. As an npm bundle, one command into the profile you use (`web`, `tui` or `headless`): ```bash dsh plugin --profile web add humanizer-ru ``` The same bundle straight from GitHub, for a specific branch or the freshest main: `dsh plugin --profile web add github:ilyautov/humanizer-ru`. The bundle is text only: `package.json` with a `dsh.bundle` field and `cordis.patch.yml`, which mounts the stock `dsh-skill-filesystem` provider on the package's `skills/` folder. No executable code, no build step. Verify without running a model: `dsh --profile web --dump-config` shows a `# == humanizer-ru` layer. Remove with `dsh plugin --profile web remove humanizer-ru`. As a copy into the user skill root: ```bash cp -r humanizer-ru/skills/humanizer-ru ~/.agents/skills/ ``` dsh scans `~/.agents/skills` and `~/.dsh/skills` on its own, no restart needed. It ignores the `allowed-tools` frontmatter key and resolves `references/` relative to the skill folder, so the catalog and the edit log open the same way as in Claude Code. ### 7. Scanner on its own: pip, MCP server, GitHub Action The scanner also ships without the skill, as the [`ru-humanizer`](https://pypi.org/project/ru-humanizer/) package on PyPI (the `humanizer-ru` name on PyPI is taken). Same `scan.py`, as a command: ```bash pip install ru-humanizer ru-humanizer article.md --genre academic ru-humanizer edited.md --before original.md ``` The same package runs an MCP server with two tools: `scan_text` returns the score, penalties, hard bans and markers with positions; `compare_texts` reports "was N, now M" plus the fact lock between source and edit. Works with Claude Desktop, Cursor and any MCP client; in Claude Code it is one command: ```bash claude mcp add humanizer-ru -- uvx ru-humanizer mcp ``` In CI the scanner is a GitHub Action: a cleanliness gate over Markdown and text files that fails below the threshold. ```yaml - uses: ilyautov/humanizer-ru@main with: files: "docs/**/*.md" genre: marketing min-score: 60 ``` One-time publishing setup for PyPI and the MCP registry lives in `PUBLISHING.md`. ## Modes - **Edit** (default): scanner diagnosis, local fixes by priority, verification against the source, "was N, now M" report. - **Audit**: diagnosis only, returns detected patterns with priority A-D. - **Targeted fix**: works on a specific category only. - **Own draft** (silent, unprompted): while the skill is installed, the agent runs the Russian prose it writes itself past the 21 hard bans before you ever see it. No report, no ask. Text that is not the agent's own (your message, a file, a quote, someone's code) is never edited silently: it goes through the three modes above, where the edit is visible. Saying «не правь» turns the mode off for the rest of the conversation. ## Usage Ask Claude in Russian: ``` Очеловечь этот текст: [paste text] Перепиши, звучит как робот: [paste text] ``` Triggers: "очеловечь", "убери следы нейросети", "сделай живым", "звучит искусственно", "перепиши как человек". ## Before / After Before: > В современном мире искусственный интеллект играет всё более важную роль в различных сферах деятельности. Стоит отметить, что данная технология является мощным инструментом для оптимизации рабочих процессов. After (author's facts): > За последний год я внедрил AI-инструменты в три проекта. Два ускорились вдвое. Третий развалился, потому что команда перестала проверять то, что выдаёт модель. AI работает, когда понимаешь его ограничения. Five hard bans triggered in two sentences. Typical. ## Do AI detectors work on Russian? Short answer: poorly, and it matters. GPTZero, Originality.ai, ZeroGPT and most popular AI-text detectors are trained mostly on English. On Russian they are unreliable and fail both ways: - **False positives:** text written by a real human routinely gets flagged as "AI-generated". English perplexity thresholds aren't calibrated for Russian, and Russian's rich morphology inflates "unpredictability" on its own. - **False negatives:** careful AI text passes as human. Even the best Russian-specific detector (RuRoBERTa on the AINL-Eval 2025 benchmark) lands around 86% accuracy, and one verdict in seven is wrong. And detectors don't generalize across domains (verified, see [SOURCES.md](SOURCES.md)). **What this means for humanizing.** Chasing "detector bypass" on Russian means tuning text to an unreliable, moving target. So humanizer-ru optimizes genuine text quality (removing bureaucratic noun-chains, calques and clichés, restoring author voice and live rhythm) rather than gaming a classifier. Those are measurable language properties (see [`eval/`](eval/)) independent of how bad any given detector is. Perplexity and burstiness rise as a side effect, which is what detectors try (and often fail) to measure. We run detectors against real human Russian texts in the harness and publish the false-positive rate: [eval/RESULTS.md](eval/RESULTS.md). ## Sources Patterns drawn from 15+ Russian-language sources (Habr, vc.ru, Gramota.ru, HSE stylometry research, Dialog Conference RuATD, TechInsider, Kokoc.com), verified academic papers (DivEye, PIFE, MASH, Antislop, AINL-Eval, RuATD, NeurIPS 2025), and the Wikipedia AI Cleanup Project. Every citation and figure is checked against primary sources; unverified ones were removed. Full provenance: [SOURCES.md](SOURCES.md). Changelog: [CHANGELOG.md](CHANGELOG.md). Metrics and eval harness: [`scripts/`](scripts/) and [`eval/`](eval/). Full Russian documentation: [README.md](README.md). ## Author Ilya Utov. I write about AI and working with text on Telegram: [Under the Hood](https://t.me/gorilla_under_hood). ## Contributors - **Filipp Zarubin** ([@zarubinphil](https://github.com/zarubinphil)): category N "Discourse layer", patterns 59-64 from StoryScope (COLM 2026). Signals that survive a style rewrite. Plus Russian numeral agreement in the release gate. - **iKonushok** ([@ikonushok](https://github.com/ikonushok)): install smoke for the installed surface and release ZIP packaging. It checks what actually reaches the user, not the folder in the repo. Full list, including bug reports that changed the code: [CONTRIBUTORS.md](CONTRIBUTORS.md). How to get on it: [CONTRIBUTING.md](CONTRIBUTING.md). ## License MIT --- ## Who built this [Ilya Utov](https://github.com/ilyautov), the [AI Frontier](https://aifrontier.tech) lab. I write about how these tools work inside on [Telegram](https://t.me/gorilla_under_hood) and [LinkedIn](https://www.linkedin.com/in/ilyautov). **Nearby:** - [**marketplaces-mcp-ru**](https://github.com/ilyautov/marketplaces-mcp-ru): Wildberries, Ozon, Yandex Market and Avito straight from the agent - [**small-business-ru**](https://github.com/ilyautov/small-business-ru): 34 skills for Russian small business, the numbers computed in code - [**consilium-principis**](https://github.com/ilyautov/consilium-principis): a board of thinkers where every quote is checked word for word - [**hefest**](https://github.com/ilyautov/hefest): chemical safety for an industrial plant, kept inside the plant's own network - [**cordon**](https://github.com/ilyautov/cordon): a deterministic layer between untrusted content and agent actions Everything else: [github.com/ilyautov](https://github.com/ilyautov). Useful? Star it, that is how other people find it.