I built K-Humanizer from 7,500+ edits across 120,000 Korean responses

evergreenRok — Sun, 23 Aug 2026 12:14:07 +0000

AI-written Korean can be grammatically correct and still feel awkward. English-shaped word order lingers, the tone is more formal than the situation calls for, and abstract wording makes it hard to tell what actually happened.

I reviewed more than 120,000 Korean responses and made over 7,500 edits. The same phrase could sound natural in one context and awkward in another. That work taught me what to change, when to leave a sentence alone, and how to preserve the writer's meaning.

I put that experience into K-Humanizer, an open-source Agent Skill for Cursor, Claude, and Codex. It helps AI drafts sound closer to the way people actually write in Korean while keeping facts, numbers, names, uncertainty, and domain terms unchanged.

한국어 글을 자연스럽게 쓰고 싶거나 AI 초안을 평소 쓰는 문장에 가깝게 다듬고 싶은 분에게 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 tells the reader what changed without the abstract wording.

What the skill checks

K-Humanizer looks for English-like word order, unnecessary passive voice, long noun-heavy phrasing, 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.

Each change depends on context. For example, the skill does not replace every instance of "~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 7,500+ edits and the public validation set serve different purposes. The edits informed how the skill handles context. The repository's public validation set contains 200 synthetic examples across resumes, everyday writing, chat, email, documents, product copy, dialogue, and code review. No real resumes, private messages, customer data, or internal company text are included.

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.