Humanizer

A bilingual, portable AI-agent instruction for rewriting AI-sounding text so it reads as human-written text.

It can be used as a Claude Code/OpenCode skill, but it is not limited to coding agents. You can also load or paste SKILL.md into ChatGPT, Claude, Codex, a Notion page, a custom GPT, an agent framework, or any other AI assistant that follows written instructions.

The skill supports two separate language modules:

• English
• Russian

The modules are intentionally separate. The agent detects the input language first and then uses only the matching module. This avoids the common failure mode where English AI-writing heuristics are applied to Russian text and produce unnatural or inaccurate edits.

Installation

Claude Code

Clone the skill into Claude Code's skills directory:

mkdir -p ~/.claude/skills
git clone https://github.com/timsmykov/humanizer.git ~/.claude/skills/humanizer

Or copy the skill file manually if you already have this repo cloned:

mkdir -p ~/.claude/skills/humanizer
cp SKILL.md ~/.claude/skills/humanizer/

OpenCode

Clone the skill into OpenCode's skills directory:

mkdir -p ~/.config/opencode/skills
git clone https://github.com/timsmykov/humanizer.git ~/.config/opencode/skills/humanizer

Or copy the skill file manually:

mkdir -p ~/.config/opencode/skills/humanizer
cp SKILL.md ~/.config/opencode/skills/humanizer/

OpenCode also scans ~/.claude/skills/, so one clone into ~/.claude/skills/humanizer/ can work for both tools.

Usage

/humanizer

[paste your text here]

You can also ask directly:

Please humanize this text:
[text]

For Russian:

Гуманизируй этот текст:
[текст]

Voice calibration

To match a personal writing style, provide a sample:

/humanizer

Here's a sample of my writing for voice matching:
[2-3 paragraphs]

Now humanize this text:
[text]

For Russian:

/humanizer

Вот пример моего стиля:
[2-3 абзаца]

Теперь гуманизируй этот текст:
[текст]

How the skill works

SKILL.md is the source of truth.

The agent follows this sequence:

1. Detect the input language.
2. Choose exactly one module: English or Russian.
3. Preserve facts, quotes, terms, numbers, dates, citations, code blocks, and source register.
4. Rewrite only AI-sounding fragments.
5. Run a final audit pass in the same language.
6. Return only the rewritten text unless the user asked for commentary.

English module

The English module is based on Wikipedia's "Signs of AI writing" guide and the original blader/humanizer skill.

It covers patterns such as:

• significance inflation
• notability name-dropping
• superficial "-ing" analysis
• promotional language
• vague attribution
• formulaic challenges/future outlook sections
• AI vocabulary clusters
• copula avoidance
• negative parallelism and tailing negation
• rule of three
• synonym cycling
• false ranges
• passive voice and subjectless fragments
• em dash overuse
• boldface and inline-header lists
• title case headings
• chatbot artifacts
• cutoff disclaimers
• sycophantic tone
• filler phrases
• excessive hedging
• generic positive conclusions
• signposting announcements
• fragmented warm-up headers

Russian module

The Russian module is the main extension in this fork.

It is designed around Russian-specific LLM markers rather than translated English rules:

• псевдо-важные вводные: важно отметить, следует подчеркнуть
• канцелярит and nominalization chains: осуществление анализа в целях выявления
• overuse of является, представляет собой, выступает в качестве
• bureaucratic demonstratives: данный, указанный, вышеупомянутый
• Russian symmetric constructions: не просто X, а Y, не только X, но и Y
• тройные перечисления with parallel grammar
• ChatGPT calques: в рамках, в разрезе, на сегодняшний день
• dramatic long dashes and ellipses
• parasite intensifiers: действительно, по-настоящему, в целом
• universalizing formulas: в современном мире, сегодня как никогда
• participial and adverbial-participial cascades
• passive and impersonal constructions where an actor is needed
• vague authority references: исследования показывают, эксперты считают
• LinkedIn-style motivational endings
• modal bureaucratic phrasing: необходимо, следует, рекомендуется
• empty universal adjectives: эффективный, комплексный, ключевой
• mechanical logical transitions: таким образом, кроме того, однако
• register drift
• terminology and case-government traps

The module also includes separate guidance for Russian registers:

• academic/scientific
• business/publicistic
• conversational/blog
• technical/instructional
• legal/regulatory

Russian example

Before:

Важно отметить, что Notion является инструментом, представляющим собой систему для управления знаниями, которая, в рамках современных реалий, играет ключевую роль в осуществлении планирования, структурирования и организации информации. Это не просто приложение — это экосистема.

After:

Notion — система для управления знаниями. В нём удобно хранить документы, задачи и решения команды. По сути, это уже не отдельное приложение, а рабочая среда.

Research basis

The English module keeps the original empirical base:

• Wikipedia: Signs of AI writing
• WikiProject AI Cleanup

The Russian module is informed by Russian and multilingual NLP work showing that Russian needs separate evaluation and detection heuristics:

• MERA: A Comprehensive LLM Evaluation in Russian
• REPA: Russian Error Types Annotation for Evaluating Text Generation and Judgment Capabilities
• Findings of the RuATD Shared Task 2022 on Artificial Text Detection in Russian
• Detecting LLM-Generated Text with Trigram-Cosine Stylometric Delta
• LLMTrace: A Corpus for Classification and Fine-Grained Localization of AI-Written Text

Version history

• 3.0.1 - Clarified that the instruction is portable beyond Claude Code/OpenCode. Expanded the Russian module with modal bureaucratic phrasing, empty universal adjectives, and mechanical logical transitions.
• 3.0.0 - Reworked the skill as a bilingual English/Russian humanizer with strict language routing. Added a dedicated Russian module with register-specific guidance, sense-preservation checks, terminology safeguards, and Russian self-checks.
• 2.5.1 - Added passive-voice / subjectless-fragment rule in the original English-only skill.
• 2.5.0 - Added persuasive framing, signposting, and fragmented headers in the original English-only skill.
• 2.4.0 - Added voice calibration.
• 2.3.0 - Added hyphenated word pair overuse.
• 2.2.0 - Added final "obviously AI generated" audit and second-pass rewrite.
• 2.1.0 - Added before/after examples for English patterns.
• 2.0.0 - Complete rewrite based on Wikipedia's AI writing guide.
• 1.0.0 - Initial release.

License

MIT