kwrite — Korean Prose Polishing (윤문)
SkillDev toolsMUST USE for Korean prose polishing (윤문) — revising existing Korean text to read like a person wrote it: removing translationese and AI idioms, fixing register/tone breaks, varying rhythm, replacing abstract endings, while preserving meaning exactly. Applies to docs, articles, announcements, READMEs, UI copy, and any Korean output the agent writes. Triggers: 윤문, 다듬어, 다듬어줘, 자연스럽게, 매끄럽게, 교정, 고쳐줘, AI투, 번역투, Korean polish, kwrite, proofread Korean.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the kwrite — Korean Prose Polishing (윤문) skill
What this skill tells your AI
The instructions your AI receives, as published by lidge-jun/codexclaw in plugins/codexclaw/skills/kwrite/SKILL.md and read by ahel’s review.
Use this skill when the task is to REVISE existing Korean text (or polish Korean text you are about to output): remove the patterns that make it read machine-written, keep everything it says intact. This is a revision discipline, not a generation pipeline — genre, format, and audience stay whatever the source text already is.
Prime directives
- Meaning is frozen. Facts, claims, numbers, proper nouns, quotes, and causal links are preserved to the letter. Ambiguity is never resolved by invention.
- Edit detected spans only. Sentences without a detected pattern are not touched. Polishing is not rewriting.
- No over-polishing. If more than ~30% of the text changes, stop and re-check: you are probably rewriting, not polishing.
- Register is part of meaning. 문어체 stays 문어체, 구어체 stays 구어체, 존댓말 stays 존댓말. Fix breaks IN the register; never migrate the text to a different one.
Revision protocol (run in order, re-run a pass after fixing it)
Pass 1 — Register/tone consistency (S1)
- First and last sentence carry the same register and speech level.
- No "~합니다/~입니다" islands inside 반말/평서 text; no "~해/~야" inside 존댓말 text; no "~하겠습니다" formality spikes in informal prose.
Pass 2 — Translationese + AI idioms (S1)
Scan against references/ai-tell-taxonomy.md CAT-1 (번역투), CAT-3 (AI 관용구), CAT-5 (톤 파괴). Anything S1 is fixed immediately; the taxonomy carries register-preserving corrections.
Pass 3 — Mechanical structure (S2)
- "첫째/둘째/셋째" enumeration, "A, B, 그리고 C" parallelism, identical sentence frames in sequence (CAT-2).
- Sentence-initial connective stacking: 또한/한편/더불어/아울러/이에 따라 (CAT-4). Default fix is deletion.
- Over-explanation: restating the obvious, "다시 말해"/"즉," repetition, parenthetical glossing of things the reader knows (CAT-7).
Pass 4 — Rhythm + endings (S2)
- Uniform sentence length or repeated final endings three times in a row: break with a short sentence, vary the ending (CAT-8).
- Abstract escape-hatch closings — 기대된다/주목된다/~할 것으로 보인다/귀추가 주목된다 (CAT-6): end on a concrete statement, a specific consequence, or simply end.
Verdict
All four passes clean -> output. Any FAIL -> fix that span, re-run that pass. Report noteworthy changes briefly when the user asked for 윤문 explicitly; apply silently when polishing your own outgoing Korean prose.
Scope guard
- This skill owns revision of existing Korean text. Generating new content for a specific platform, audience, or format is a different job — use the skill that owns that surface (or plain judgment) and then run this protocol on the draft.
- Structure is out of scope here. When the draft is a report or explainer, apply Reader documents first (두괄식: 결론이 맨 앞에 있는지), then run this protocol on the sentences.
- Non-Korean text: out of scope. Mixed-language docs: polish only the Korean prose, leave code/English untouched.
Signals
- GitHub stars
- 37
- Forks
- 7
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
cxc-kwrite- Source
- github.com/lidge-jun/codexclaw