/humanizer-kr

name: humanizer-kr
description: Rewrite, edit, or review Korean text so it sounds natural, human, and source-grounded rather than AI-generated or translation-like. Use for Korean product copy, posts, emails, public notices, documentation, essays, proposals, scripts, and other drafts that need Korean-specific register control, plain-language discipline, removal of AI-writing tells, or voice matching from a user-provided Korean writing sample.

Humanizer KR

Use this skill to make Korean writing sound like it was written by a person who understands the audience, not by a chatbot translating an English outline.

Do not help users hide plagiarism, impersonate a real person's authorship, or bypass academic or hiring integrity checks. It is fine to improve clarity, register, voice, and readability when the user owns or is allowed to edit the text.

Core Workflow

1. Identify the genre and reader: product copy, public notice, email, essay, documentation, proposal, social post, or spoken script. 2. Preserve facts, intent, paragraph coverage, and constraints. Rewrite the language, not the underlying claims. 3. Choose one register and keep it stable:

• formal public/professional: `합니다`, `하십시오`
• polite product/support: `해요`, `해주세요`
• concise internal note: plain declarative Korean with minimal honorific padding

4. Rewrite AI-like wording into Korean that is concrete, direct, and rhythmically varied. 5. Run the Korean audit pass before returning the final answer. 6. Run one naturalness pass: check whether the rewrite still feels flat, generic, vague, or mismatched to the genre.

When factual claims are weak, do not invent supporting detail. Either keep the claim modest, mark the uncertainty, or ask for a source when the source materially affects the rewrite.

Naturalness Rubric

Use this after removing obvious AI-writing tells. A rewrite can be clean and still feel unnatural.

• Genre fit: product copy should be sharper than documentation; support email should tell the reader what to do next; public notices should separate issue, affected reader, action, and timing when those facts exist.
• Reader action: if the reader needs to do something, name the action. Avoid vague endings such as `필요한 조치를 진행해 주세요` unless the source genuinely withholds the action.
• Concrete subject: prefer the actual actor or object over generic `이 도구`, `이 서비스`, `본 기능` when the source gives enough context.
• Rhythm: vary sentence length. Do not make every sentence land with the same `합니다` rhythm unless the genre requires formal uniformity.
• Warmth: remove ceremonial padding, but do not make support or public-facing text cold. Keep enough respect for the reader relationship.
• Claim control: do not add benefits, examples, numbers, deadlines, app names, institutions, or causal claims that the source did not provide.
• Voice match: if a sample is provided, the sample wins over generic polished Korean.

When the first rewrite is safe but bland, improve once more. Prefer a smaller, more specific sentence over a broader sentence with polished wording.

Failure and Uncertainty Policy

Fail visible when the source is too thin to rewrite safely.

• If an action, timeline, affected audience, or cause is missing from a notice or support reply, do not invent it. Keep the final text modest and add a short note asking for the missing detail.
• If a claim needs evidence, do not dress it up. Lower the claim or ask for the source.
• If the requested voice conflicts with facts, legal wording, quoted text, or safety constraints, preserve the constraint and explain the tradeoff briefly.
• If the user asks for an integrity-bypassing rewrite, refuse that use and offer a clarity/readability edit that does not misrepresent authorship.

Voice Calibration

If the user provides a Korean writing sample, read it before editing and match:

• sentence length and paragraph rhythm
• ending style (`-다`, `-합니다`, `-해요`, mixed casual)
• common connectors and how often they appear
• tolerance for English terms, parentheses, slashes, and emoji
• level of emotional color, humor, or bluntness

Do not "upgrade" the user's voice into generic polished Korean. If the sample is plain, stay plain. If the sample is slightly rough but intentional, keep that pulse.

Korean AI-Writing Tells

Look for clusters, not isolated words. The five groups below are the editing frame; the 20 patterns are observation signals, not banned words. Apply them after identifying the genre, reader, and register.

Read `references/rewriting-playbook.md` when a rewrite needs a concrete edit strategy or when the first pass is safe but still bland. Do not promote a new tell into the main workflow until it has evidence, a clean negative example, and regression coverage.

1. Evidence and Claim Strength

Fix sentences where the claim is larger than the evidence. Lower broad claims into verified facts, specific actions, conditions, or explicit uncertainty.

• Inflated praise: when `혁신적인`, `최적화된`, `강력한`, or similar praise appears before evidence, replace it with the actual feature, result, or behavior.
• Vague authority: when `많은 전문가들은`, `업계에서는`, or `여러 연구에 따르면` lacks a source, add a specific source, narrow the claim, or remove the authority frame.
• Knowledge-gap disclaimer: when `정보는 제한적이지만` or `알려진 바는 많지 않지만` leads into speculation, separate what is known from what is unknown.
• Generic positive conclusion: replace `앞으로의 행보가 기대됩니다` or `좋은 결과가 기대됩니다` with a known next step, plan, or modest close.
• Hedging stack: when `상당히`, `비교적`, `~로 보입니다`, and `~할 수 있을 것으로 예상됩니다` pile up, keep the strongest claim the evidence supports.

Before: `다음 업데이트에서 검색 필터와 저장 기능을 먼저 개선하며 앞으로의 행보가 기대됩니다.` After: `다음 업데이트에서는 검색 필터와 저장 기능을 먼저 개선합니다.`

2. Translationese and Vocabulary Density

Fix wording that sounds like a translated outline or business-English draft. Prefer Korean relationships, natural collocations, and terms the reader actually uses.

• Translationese connectors: when