korean-writing-style for Claude Code: a skill that filters translationese from Korean

korean-writing-style is a single skill that filters translationese and AI stock phrases out of Korean documents. This post covers the problem it solves, who it suits, how to install it and run it for the first time, and the one check worth making before you trust its verification step.

What problem it solves

Korean text drafted by an agent often reads like a translation. Sentences follow English word order, stock phrases pile up, modifiers stack on each other, and hedges soften every claim. Most people try to fix this after the draft is finished, but by then the sentence structure already follows the translated pattern, so the edited version still sounds awkward. The skill’s README makes exactly this point, and the skill is built around the opposite order: read the style rules before drafting, then check afterward.

The rules come as tables. One section lists rough colloquial verbs with alternatives to use instead. Another groups translationese and AI stock phrases into six families: literal English word order, AI stock phrases, excessive modifiers, excessive hedging, overused conjunctions, and dependent nouns. The patterns are graded so the result does not turn stiff. High-severity patterns are removed immediately, while lighter ones are allowed twice and removed from the third occurrence onward. A punctuation section lists special symbols that are prohibited outright.

Who it fits

The chapter rates four situations. Three are enough to decide.

Situation	Fit	Reason
Polishing external reports and proposals	Good fit	Requires zero prohibited expressions
Creating shared team style rules	Good fit	Add one row at a time to the table
Cleaning up code comments or configuration files	Poor fit	Excluded from its scope

Internal notes and meeting records sit in the middle. The skill works there, but it tolerates one or two occurrences rather than demanding zero.

Install and first run

Clone the repository, then create the skill folder and copy SKILL.md into it.

# From the cloned repository folder, create the skill folder and copy SKILL.md (Claude Code)
mkdir -p ~/.claude/skills/korean-writing-style
cp 02-skills/korean-writing-style/SKILL.md ~/.claude/skills/korean-writing-style/SKILL.md

After installation, call it by name or include the skill name in a document-writing request.

/korean-writing-style Polish the style of this report draft: report.md

The run follows three steps. It reads the style rules table before drafting, edits rough expressions, translationese, and punctuation, and then checks the remaining matches with grep and keeps revising until the count is zero. Completion is defined by commands rather than by the agent’s judgment: the work is done only when both grep checks, one for prohibited punctuation and one for rough expressions, return zero matches.

What trips people up

The punctuation check relies on a specific grep option, and not every environment’s grep supports it. Check whether your grep supports it before you rely on the zero count. Keep in mind, too, that a passing grep is not the same as a clean document. The commands catch prohibited punctuation and rough expressions, not every translationese pattern, so read the body once against the table as well. The tables themselves are the author’s defaults. Add the expressions your team dislikes and your rules for consistent internal terminology before you treat the skill as complete.

Where this comes from

This post adapts part of chapter 02-07 of The Claude Code & Codex Skills Guidebook. The full chapter covers the before-and-after examples for each of the six pattern groups and a section-by-section map of what the skill file contains, from scope by document type to how to expand the tables, and the book has 52 more skills like this one. Read the full chapter in The Claude Code & Codex Skills Guidebook on Gumroad.