Verified project record

epoko77-ai/im-not-ai

im-not-ai is a CLI skill that rewrites AI-written Korean text into more natural Korean while preserving facts, claims, numbers, proper nouns, and direct quotations. It classifies Korean AI-writing patterns and rewrites style, rhythm, and expression through a routed workflow.

Project overview

It targets Korean specifically because English-language humanizers are weak at Korean and Korean AI-writing patterns often arise from translationese, and its workflow reduces model calls for well-written text through deterministic light, standard, and heavy routing rather than changing the model.

Project type: AI Agent · Prompt Engineering
Use cases: Documents & Office
Deployment: Refer to project documentation
License: MIT

Best for

Users who produce Korean text with AI and need it to read naturally while keeping facts, numbers, proper nouns, and quotations unchanged.
Developers who maintain short Korean commit messages and want an opt-in naturalization step.

Key capabilities

Classifies Korean AI-writing patterns across 10 categories and 85 sub-patterns, including translationese, mechanical parallelism, AI idioms, and other stylistic signals.
Rewrites Korean style, rhythm, and expression while preserving facts, claims, numbers, proper nouns, and direct quotations.
Uses quantitative pre-scoring to select a light, standard, or heavy rewriting route with different numbers of model calls.
Checks change rate, diagnosis goal achievement, parallelism preservation, and golden or numeric verification across routes.
Provides a humanize-scan command to assess whether a full rewriting pass is worthwhile.
An opt-in extra skill that removes bureaucratic or translationese wording from short Korean commit messages, typically one or two lines.

Limitations and risks

The documented taxonomy and rewriting rules are specialized to Korean text.
The reported false-positive remeasurement used the same corpus used to derive the cells, so out-of-sample validation remains pending.
A prior plugin layout caused some loaders to omit the quantitative shim and diagnostics while still producing apparently normal output; verify your loader picks up the full plugin before relying on results.

Getting started

Add the repository as a plugin marketplace, install humanize-korean, then start a supported CLI session and request Korean text rewriting; the README documents natural-language invocation after installation.

Evidence and sources

README: > **Claude Code**, **GitHub Copilot CLI**, **OpenAI Codex CLI**, **Gemini CLI**를 지원합니다. 전체 가이드: [`INSTALL.md`](INSTALL.md)
README: AI(ChatGPT · Claude · Gemini 등)가 쓴 한글 글을 **내용은 한 글자도 건드리지 않고** 문체 · 리듬 · 표현만 자연스러운 한국어로 되돌리는 CLI 스킬입니다. 번역투, 과도한 영어 인용, 기계적 병렬 ("첫째 · 둘째 · 셋째"), "결론적으로 / 시사하는 바가 크다" 같은 AI 특유 관용구,…
README: 영어권 humanizer(QuillBot · Hix · Undetectable AI)는 한국어에 약합니다.
README: 입력을 shim(`prepare_monolith_input.py`)이 먼저 정량 채점하고, 그 점수로 **`route_hint`(light | standard | heavy)** 를 결정적으로 산출합니다. 글의 상태가 경로를 정하고, 경로가 콜 수를 정합니다.
README: [verify_gates.py] ── 문자율·목표달성·대구 전멸·golden/수치 검증 — 모든 경로 공통

Source updated at: 2026-09-25T05:15:57Z

Open original GitHub project