I built K-Humanizer, an open-source Agent Skill for natural Korean writing

Korean text from LLMs is often grammatically correct but still sounds translated. The word order follows English, abstract nouns hide the action, and the tone is often more formal than the situation needs.

I kept seeing those patterns while cleaning Korean text data every day for a month, so I turned the repeated edits into an Agent Skill.

K-Humanizer works with Cursor, Claude, and Codex. It rewrites Korean for Korean readers while keeping facts, numbers, names, uncertainty, and domain terms unchanged.

If you want your Korean writing to sound more natural, especially after using an LLM for a first draft, this skill is meant for that use case.

한국어 글을 자연스럽게 쓰고 싶거나 AI 초안의 번역투와 과한 격식을 줄이고 싶은 분께 추천합니다.

A small example

Before:

고객 문의 분류 체계 고도화를 통해 주간 평균 처리 시간을 18시간에서 11시간으로 크게 개선했습니다.

After:

고객 문의 분류 기준을 정리해 주간 평균 처리 시간을 18시간에서 11시간으로 줄였습니다.

Both sentences make the same factual claim. The second removes the abstract wording and states the action directly.

What the skill checks

K-Humanizer looks for English-like word order, unnecessary passive voice, long noun phrases, repeated connective words, and formality that does not fit the channel. It also distinguishes between a resume, email, chat message, product screen, and everyday writing.

The rules are conditional. For example, it does not remove every instance of a phrase such as "~를 통해". It keeps the phrase when it describes a real path or intermediary and changes it only when it makes the sentence longer without adding meaning.

What it protects

The skill is designed to keep:

• facts and numbers
• names and sources
• professional terminology
• the original level of certainty
• sentences that already sound natural

It does not promise to bypass AI detectors. It also does not invent resume achievements or make uncertain claims sound definite.

Try it

npx skills add evergreentree97/K-Humanizer --skill k-humanizer --full-depth

Then give the skill the purpose of the text and its intended reader:

Use $k-humanizer to revise the Korean text below for a product notice.
Keep facts, numbers, names, and product terms unchanged.
Fix only translation-like phrasing and unnecessary formality.

[Your text]

Validation so far

The repository includes 200 synthetic examples across resumes, everyday writing, chat, email, documents, product copy, dialogue, and code review.

The current public report checks dataset structure, domain distribution, duplicate IDs, required fields, and repository hygiene. It does not claim a naturalness score yet. A separate manual evaluation for naturalness and meaning preservation is planned.

The repository is MIT licensed. If you find a Korean sentence that still sounds translated, a synthetic before-and-after example would be useful. Please do not submit real resumes, private messages, customer data, or internal company text.

GitHub: https://github.com/evergreentree97/K-Humanizer