SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillshumanize-korean-ai-text

humanize-korean-ai-text

General & other

View repository

Updated 10-15 23:41

About this skill

Humanize KR — Korean AI-tell remover (im-not-ai)

Skill by ara.so — Daily 2026 Skills collection

Korean text generated by AI (ChatGPT · Claude · Gemini, etc.) while keeping the content unchanged is a Claude Code skill that converts only the style, rhythm, and expressions into natural Korean. It detects and fixes 10 major categories × 40+ patterns, including translationese, mechanical parallel structures, AI-specific idioms, and overuse of the passive voice.

Installation

git clone https://github.com/epoko77-ai/im-not-ai.git
cd im-not-ai
claude

You must im-not-ai inside the folder claudefor the skill to load. If you run it from another path, it will operate as regular Claude Code.

Prerequisites: Claude Code CLI installation required

claude --version  # 설치 확인

Basic usage

Method A — Natural language (easiest)

이 AI 글 자연스럽게 윤문해줘:

[ChatGPT / Claude / Gemini 초안 붙여넣기]

Using any of the expressions below will automatically trigger the skill:

  • "AI 티 없애줘"
  • "GPT 문체 제거해줘"
  • "사람이 쓴 것처럼 윤문해줘"
  • "번역투 제거"
  • "한글 AI 윤문"

Method B — Slash command

/humanize [텍스트 또는 파일 경로]

# 옵션 예시
/humanize ./draft.txt 장르: 칼럼
/humanize ./report.md 강도: 적극 최소심각도: S1
/humanize ./essay.txt 원문 톤을 더 살려줘

# 결과가 맘에 안 들면 재실행
/humanize-redo "번역투만 다시"
/humanize-redo "관용구만 재처리"

Method C — Plugin installation (forked version)

# gaebalai 포크 — Claude Code Plugin 규격
/plugin install humanize-korean@epoko77-ai-plugins

# 또는 스크립트로 다른 프로젝트에 설치
./scripts/install.sh --target ~/my-project

Processing modes

Fast mode (default — under 5,000 characters)

humanize-monolith Handles detection, polishing, and self-verification all at once with a single agent call.

  • Time required: 약 2~3분
  • Tool calls: Capped at 4~5 (Read input → Read rulebook → Write final → Write summary)
  • Auto-selection conditions: Input under 5,000 characters
입력 텍스트
    ↓
[humanize-monolith]  ← 탐지 + 윤문 + 자체검증 일괄
    ↓
_workspace/{날짜-번호}/final.md
_workspace/{날짜-번호}/summary.md

Strict mode (--strict or 8,000+ characters)

A 5-person pipeline processes step by step, saving each step's output to a separate file.

/humanize ./long-report.md --strict

입력 텍스트
    ↓
[ai-tell-detector]          → 02_detection.json
    ↓
[korean-style-rewriter]     → 03_rewrite.md
    ↓
[content-fidelity-auditor]  → 04_fidelity_audit.json  (병렬)
[naturalness-reviewer]      → 05_naturalness_review.json (병렬)
    ↓
[오케스트레이터 종합]
    ↓
final.md + summary.md

Output file structure

_workspace/
└── 20260426-001/          # 실행마다 새 폴더 생성
    ├── 01_input.txt       # 원문 그대로
    ├── final.md           # 윤문 결과
    └── summary.md         # 메트릭 · 탐지 결과 · 등급 · 주요 변경

# Strict 모드 추가 파일
    ├── 02_detection.json          # span 단위 AI 티 탐지 리포트
    ├── 03_rewrite.md              # 윤문본 초안
    ├── 04_fidelity_audit.json     # 의미 동등성 감사 (13항 체크리스트)
    └── 05_naturalness_review.json # 자연도 재측정 결과

AI-tell classification system — 10 major categories

ID Major category Example main patterns
A Translationese ~를 통해, ~에 있어서, ~되어진다 (double passive), 가지고 있다
B excessive English quotation Using translatable English as-is, excessive parenthetical duplication
C Structural AI patterns Mechanical 첫째/둘째/셋째, excessive bullets/headings/emoji
D AI-specific idioms 결론적으로, 시사하는 바가 크다, 주목할 만하다, 혁신적인
E Rhythm uniformity Low standard deviation of sentence length, repeated identical sentence endings
F Modifiers/redundancy 매우, 정말, synonym double modification, ~적/~성/~화 overuse
G Hedging overuse ~할 수 있을 것으로 보인다 multiple hedges
H Conjunction overuse Sentence-initial 또한/따라서/즉/나아가 in succession
I Formal noun overuse 것이다, 점, 수, 바, ~할 필요가 있다
J Visual decoration overuse Excessive bold, "따옴표", dashes (—) overuse

40+ sub-patterns in total: .claude/skills/humanize-korean/references/ai-tell-taxonomy.md

Severity criteria

Grade Criteria Action
S1 Definitive AI certainty even if it appears only once Remove unconditionally
S2 Strong Allowed 1~2 times, repeated 3+ times Remove on repetition
S3 weak Harmless on its own Remove only when overlapping with other patterns

Quality grade (check summary.md after revision)

Grade Criteria Follow-up action
A S1 0 cases, S2 ≤2 cases, score improvement 70%+ Complete
B S1 0 cases, S2 ≤4 cases, improvement 50%+ Complete
C S1 1–2 cases or 2 over-editing signals Second revision starts automatically
D S1 3+ cases or severe over-editing Human review recommended

The Four Cardinal Rules (rules the agent always follows)

  1. Meaning invariance — 100% preservation of the original text for facts, claims, figures, proper nouns, and direct quotes
  2. Evidence-based — surgical edits only to detected spans, no changes to undetected sections
  3. Genre preservation — no converting columns into literature or reports into essays
  4. No over-editing — warning if the change rate exceeds 30%, forced stop if it exceeds 50%

Excluded from detection and editing

The following items are never modified:

  • Figures, units, dates
  • Proper nouns, personal names, product names, model names
  • Direct quotes inside double quotation marks
  • Legal and regulatory provisions
  • Academic terminology (when unavoidable)

Actual conversion examples

Translationese (Category A)

Before: AI 기술을 통해 효율을 높일 수 있다.
After:  AI로 효율을 높인다.

Before: 이에 있어서 중요한 점은
After:  여기서 중요한 건

Before: ~에 의해 생성된
After:  ~가 만든

Before: 가지고 있다
After:  있다

AI clichés (Category D)

Before: 결론적으로, 이는 시사하는 바가 크다.
After:  (삭제 또는 문맥에 맞게 구체화)

Before: 이는 주목할 만한 혁신적인 접근법이다.
After:  이 방식은 눈길을 끈다.

Hedging overuse (Category G)

Before: 이는 효과적일 수 있을 것으로 보인다.
After:  이 방식은 효과적이다. (또는 문맥에 따라 적절히 완곡화)

Mechanical parallelism (Category C)

Before:
첫째, 비용이 절감된다.
둘째, 시간이 단축된다.
셋째, 품질이 향상된다.

After:
비용과 시간이 줄고 품질도 오른다.

Re-run · correction commands

If you are not satisfied with the result, just say so in natural language:

"이 문단만 다시 윤문해줘"          → 해당 구간만 재시도
"번역투만 더 손봐줘"               → 카테고리 A만 재처리
"관용구만 다시"                    → 카테고리 D만 재처리
"윤문 강도 낮춰줘"                 → S1 결정적 패턴만 제거
"원문 톤을 더 살려줘"              → 변경률 상한 낮춤
"2차 윤문해줘"                     → 현재 결과를 한 번 더 다듬기

Partial re-run commands automatically switch to Strict mode.

Agent configuration (reference)

Agent Mode Role
humanize-monolith Fast (default) Batch processing of detection, editing, and self-verification
ai-tell-detector Strict Span-level JSON detection report
korean-style-rewriter Strict Finding-based surgical editing, change rate monitoring
content-fidelity-auditor Strict Meaning equivalence audit (13-item checklist)
naturalness-reviewer Strict Residual AI markers, over-editing, naturalness judgment, grades A–D
korean-ai-tell-taxonomist Separate command Classification system (SSOT) management, review of new patterns
humanize-web-architect Option Next.js 15 + Vercel web service extension design

Key reference files

.claude/skills/humanize-korean/
├── references/
│   ├── ai-tell-taxonomy.md      # 40+ 서브 패턴 전체 목록 및 처방
│   ├── rewriting-playbook.md    # 카테고리별 윤문 규칙
│   ├── quick-rules.md           # Fast 모드용 슬림 룰북 (~150줄)
│   └── web-service-spec.md      # 웹 서비스 확장 설계
└── commands/
    ├── humanize                 # /humanize 커맨드 정의
    └── humanize-redo            # /humanize-redo 커맨드 정의

Frequently asked questions

Q: The skill doesn't load and it behaves like regular Claude
→ im-not-ai inside the folder claudeCheck that you ran

pwd  # /path/to/im-not-ai 이어야 함
claude

Q: Processing takes too long (25 min+)
→ This is a problem with versions before v1.5. git pullGet the latest version with . Fast mode completes within 2–3 minutes.

Q: The change rate exceeded 50% and it was forcibly stopped
→ This is an over-editing safety mechanism. "윤문 강도 낮춰줘" Or "S1만 제거해줘"Re-run with.

Q: Numbers or proper nouns changed
→ This is a bug. _workspace/{날짜-번호}/01_input.txt와 final.mdCompare and report: Issues

Q: Can it be used with the web version or API?
→ Currently, it is exclusive to Claude Code CLI. The web service is on the v1.6 roadmap (humanize-web-architect being designed as an agent).

Q: Does it support Japanese and Chinese?
→ Currently, it is Korean-only. The v4 roadmap includes expansion to Japanese and Chinese.

Version history summary

Version Major changes
v1.5 Monolith Fast Path added, v1.2~v1.4 rollback, quick-rules.md added
v1.4 Role-based model distribution (deprecated due to performance issues)
v1.3 Candidate pool introduced (deprecated)
v1.2 Voice profile introduced (deprecated)
v1.1 5-agent pipeline (preserved in strict mode)