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

Korean text generated by language models often reads grammatically correct but still feels translated. The sentence structure mirrors English, abstract nouns obscure the action, and the tone frequently feels overly formal for the context. While cleaning Korean text data daily for a month, I observed these recurring patterns, prompting me to transform the repetitive edits into an Agent Skill called K-Humanizer.

K-Humanizer functions seamlessly with Cursor, Claude, and Codex, refining Korean language output for native Korean readers, maintaining facts, numbers, names, uncertainty, and domain-specific terms intact. If your goal is to render Korean writing more natural, particularly after employing an LLM for an initial draft, K-Humanizer is the ideal solution.

Originally written in Korean: "고객 문의 분류 체계 고도화를 통해 주간 평균 처리 시간을 18시간에서 11시간으로 크게 개선했습니다." Translated: "고객 문의 분류 기준을 정리해 주간 평균 처리 시간을 18시간에서 11시간으로 줄였습니다." Both sentences convey the same factual claim, but the latter eliminates abstract wording and articulates the action more directly. K-Humanizer scrutinizes for English-like word order, superfluous passive voice, lengthy noun phrases, redundant connective words, and formality that fails to suit the platform.

It also identifies the content type - whether it's a resume, email, chat message, product screen, or everyday writing - and adapts its edits accordingly. The conditional rules ensure phrases like "~를 통해" are retained when depicting a genuine process or intermediary, but eliminated when elongating the sentence without imparting meaning.

K-Humanizer safeguards the following: factual accuracy and numerical precision, source attribution and proper names, professional jargon, and the original certainty level of the text. It does not guarantee evasion of AI detection tools nor fabricate accomplishments for resumes or render uncertain statements as definite. To utilize it, execute "npx skills add evergreentree97/K-Humanizer --skill k-humanizer --full-depth" and then apply it with the text's intended purpose and reader in mind.

Revise the supplied Korean text for a product notice, preserving all facts, numbers, names, and product-related terminology, and modify only the translation-like phrasing and excessive formality. The repository contains 200 synthetic examples, spanning resumes, everyday writing, chat, email, documents, product copy, dialogue, and code reviews.

The present public report evaluates the dataset structure, domain distribution, duplicate IDs, required fields, and repository hygiene but does not yet provide a naturalness score. A distinct manual evaluation focusing on naturalness and meaning preservation is slated. The repository adheres to the MIT license. If you encounter any Korean sentences that still sound translated, a synthetic before-and-after example would be highly beneficial.

Please refrain from submitting genuine resumes, private messages, customer data, or internal company communications. For more information, visit the GitHub repository: https://github.com/evergreentree97/K-Humanizer.