npx skills add daleseo/humanizer-koREADME
humanizer-ko
An AI agent skill that detects and corrects Korean AI writing patterns to make text sound naturally human-written.
νκ΅μ΄ AI μλ¬Έ ν¨ν΄μ κ°μ§νκ³ μμ°μ€λ¬μ΄ μΈκ°μ κΈμ°κΈ°λ‘ λ³ννλ AI μμ΄μ νΈ μ€ν¬μ λλ€.
Based on scientific research (KatFishNet paper, 94.88% AUC accuracy) β’ κ³Όνμ μ°κ΅¬ κΈ°λ° (KatFishNet λ Όλ¬Έ, 94.88% AUC μ νλ)
English
What is humanizer-ko?
humanizer-ko is an AI agent skill that transforms artificial-sounding Korean text into natural human writing based on scientific linguistic research. It works with Claude Code, Cursor, Windsurf, and other AI coding agents that support the Agent Skills format. The skill analyzes and corrects the distinctive patterns that large language models (LLMs) like ChatGPT, Claude, and Gemini produce when writing in Korean.
The Problem: LLM-generated Korean text exhibits measurable linguistic patterns that differ from human writing. For example, LLMs use commas in 61% of sentences while humans use them in only 26%. LLMs also show overly consistent spacing (standard deviation 0.02 vs natural human variation), higher noun ratios, and poverty in verb/adjective usage.
The Solution: Based on the KatFishNet paper (ArXiv 2503.00032v4), this skill detects 19 linguistic patterns across 5 categories and rewrites text to sound naturally human-written while preserving the original meaning and formality level.
Key Features
- Scientifically grounded: Based on empirical linguistic research (KatFishNet paper)
- Korean-specific: Analyzes patterns unique to Korean language, not translated from English/Chinese
- 19 detection patterns: Organized into 5 categories by priority
- Preserves meaning: Maintains original intent and formality level
- High accuracy: 94.88% AUC for punctuation, 82.99% for POS, 79.51% for spacing
Quick Start
Installation:
Using Skills CLI (recommended):
npx skills add daleseo/humanizer-ko
Or manual installation:
mkdir -p ~/.claude/skills
git clone https://github.com/daleseo/humanizer-ko.git ~/.claude/skills/humanizer-ko
Usage:
/humanizer-ko
[Paste Korean text to humanize]
Detection Patterns
The skill analyzes 5 categories of AI writing patterns:
1. Punctuation (6 patterns) - Highest priority (94.88% AUC)
- Excessive comma usage (LLM 61% vs human 26%)
- English-style comma placement
- Commas after connective endings
- Sentence-ending comma patterns
- Unnecessary list commas
- Em-dash overuse
2. Spacing (3 patterns) - High priority (79.51% AUC)
- Rigid bound noun spacing (standard deviation <0.05)
- Excessive auxiliary verb spacing
- Numeral-bound noun spacing
3. Part-of-Speech Diversity (3 patterns) - High priority (82.99% AUC)
- Noun overuse (>35% ratio)
- Verb/adjective poverty (<25% ratio)
- POS n-gram monotony
4. Vocabulary (3 patterns) - Medium priority
- AI vocabulary overuse (μ€μνλ€, ν¨κ³Όμ , νμ μ , μ§μκ°λ₯ν, etc.)
- Unnecessary Sino-Korean terms (μ§ννλ€βνλ€, μ€μνλ€βνλ€)
- English idiom calques (~μ μ€μ¬μ, ~λ₯Ό ν΅ν΄)
5. Sentence Structure (4 patterns) - Medium priority
- Lack of rhythm (uniform sentence length)
- Rule of three overuse
- Connector overuse (κ·Έλ¬λ, λν, λ°λΌμ)
- Honorific uniformity
Example
Before (AI-generated):
μΈκ³΅μ§λ₯ κΈ°μ μ λ°μ μ λΉ λ₯΄κ² μ§νλκ³ μμΌλ©°, λ€μν μ°μ λΆμΌμ μ μ©λκ³ μμ΅λλ€. μ΄λ¬ν κΈ°μ μ ν¨κ³Όμ μΌλ‘ μ 무 ν¨μ¨μ±μ ν₯μμν€κ³ , νμ μ μΈ μ루μ μ μ 곡νλ©°, μ§μκ°λ₯ν μ±μ₯μ κ°λ₯νκ² ν©λλ€.
After (humanized):
μΈκ³΅μ§λ₯ κΈ°μ μ λΉ λ₯΄κ² λ°μ νκ³ μμΌλ©° μ¬λ¬ μ°μ λΆμΌμ μ μ©λκ³ μμ΅λλ€. μ΄ κΈ°μ μ μ 무 ν¨μ¨μ λμ΄κ³ μλ‘μ΄ μ루μ μ μ 곡νλ©° μ₯κΈ°μ μ±μ₯μ κ°λ₯νκ² ν©λλ€.
Key changes: Removed 6 commas, replaced AI vocabulary (λ€μνβμ¬λ¬, ν¨κ³Όμ μΌλ‘βλμ΄κ³ , νμ μ μΈβμλ‘μ΄, μ§μκ°λ₯νβμ₯κΈ°μ ), replaced unnecessary Sino-Korean terms (μ§νλκ³ βλ°μ νκ³ , ν₯μμν€κ³ βλμ΄κ³ ).
Resources
- π KatFishNet Paper (ArXiv 2503.00032v4)
- π humanizer (English version)
- π¨π³ humanizer-zh (Chinese version)
- π Detailed pattern reference - See reference files for comprehensive pattern explanations
License
MIT License - Free to use, modify, and distribute.
νκ΅μ΄
humanizer-koλ?
humanizer-koλ κ³Όνμ μΈμ΄ν μ°κ΅¬λ₯Ό κΈ°λ°μΌλ‘ μΈμμ μΈ νκ΅μ΄ ν μ€νΈλ₯Ό μμ°μ€λ¬μ΄ μΈκ°μ κΈμ°κΈ°λ‘ λ³ννλ AI μμ΄μ νΈ μ€ν¬μ λλ€. Claude Code, Cursor, Windsurf λ± Agent Skills νμμ μ§μνλ λ€μν AI μ½λ© μμ΄μ νΈμμ μ¬μ©ν μ μμ΅λλ€. ChatGPT, Claude, Gemini κ°μ λν μΈμ΄ λͺ¨λΈ(LLM)μ΄ νκ΅μ΄λ‘ μμ±ν λ λνλλ νΉμ§μ μΈ ν¨ν΄μ λΆμνκ³ κ΅μ ν©λλ€.
λ¬Έμ : LLMμ΄ μμ±ν νκ΅μ΄ ν μ€νΈλ μΈκ°μ κΈμ°κΈ°μ λ€λ₯Έ μΈ‘μ κ°λ₯ν μΈμ΄νμ ν¨ν΄μ 보μ λλ€. μλ₯Ό λ€μ΄, LLMμ λ¬Έμ₯μ 61%μμ μΌνλ₯Ό μ¬μ©νλ λ°λ©΄ μ¬λμ 26%λ§ μ¬μ©ν©λλ€. λν LLMμ μ§λμΉκ² μΌκ΄μ μΈ λμ΄μ°κΈ°(νμ€νΈμ°¨ 0.02 vs μΈκ°μ μμ°μ€λ¬μ΄ λ³ν), λμ λͺ μ¬ λΉμ¨, λμ¬/νμ©μ¬ λΉκ³€μ 보μ λλ€.
ν΄κ²°μ± : KatFishNet λ Όλ¬Έ(ArXiv 2503.00032v4)μ κΈ°λ°μΌλ‘ 5κ° μΉ΄ν κ³ λ¦¬μ 19κ°μ§ μΈμ΄νμ ν¨ν΄μ κ°μ§νκ³ , μλ¬Έμ μλ―Έμ 격μ μμ€μ μ μ§νλ©΄μ μμ°μ€λ¬μ΄ μΈκ°μ κΈμ°κΈ°λ‘ μ¬μμ±ν©λλ€.
μ£Όμ νΉμ§
- κ³Όνμ κΈ°λ°: μ€μ¦μ μΈμ΄ν μ°κ΅¬ κΈ°λ° (KatFishNet λ Όλ¬Έ)
- νκ΅μ΄ νΉν: μμ΄/μ€κ΅μ΄ λ²μμ΄ μλ νκ΅μ΄ κ³ μ ν¨ν΄ λΆμ
- 19κ°μ§ ν¨ν΄: μ°μ μμλ³ 5κ° μΉ΄ν κ³ λ¦¬λ‘ κ΅¬μ±
- μλ―Έ 보쑴: μλ¬Έμ μλμ 격μ μμ€ μ μ§
- λμ μ νλ: λ¬Έμ₯λΆνΈ 94.88% AUC, νμ¬ 82.99%, λμ΄μ°κΈ° 79.51%
λΉ λ₯Έ μμ
μ€μΉ:
Skills CLI μ¬μ© (κΆμ₯):
npx skills add daleseo/humanizer-ko
λλ μλ μ€μΉ:
mkdir -p ~/.claude/skills
...