Im-not-ai
lastmod 2026-11-17created 2026-11-172 edits
im-not-ai (Humanize KR) is a coding agent skill for editing Korean prose, made by Liam Lee (GitHub epoko77-ai). It finds traces such as translationese, mechanical parallelism, and AI-typical stock phrases in Korean text written by ChatGPT, Claude, Gemini, and others, and rewrites only the style and rhythm into natural Korean while leaving facts, claims, and figures untouched. It started from the observation that English-language humanizers such as QuillBot handle Korean poorly, and it holds that most of the AI feel in Korean AI-written text comes from English translationese.
About
- The skill is named
humanize-korean. It divides AI tells into 10 categories and 85 sub-patterns, detected at the span level: translationese (A), excessive English quotation (B), structural patterns (C), AI stock phrases (D), uniform rhythm (E), redundant modifiers (F), hedging (G), overused conjunctions (H), excessive formal nouns (I), and visual decoration (J) - Examples: "AI 기술을 통해 효율을 높일 수 있다" ("efficiency can be improved through AI technology") → "AI로 효율을 높일 수 있다" ("AI can improve efficiency"); "~에 의해 생성된" ("generated by ~") → "~가 만든" ("made by ~"); "결론적으로, 이는 시사하는 바가 크다" ("In conclusion, this has significant implications") → deleted
- Each pattern gets a severity: S1 (a strong signal even if it appears once), S2 (removed when repeated 3 or more times), or S3 (a problem only when overlapping with other patterns). The edited result is graded A to D, and D goes to human review
- Facts, figures, proper nouns, and direct quotations stay as in the original; only detected spans are changed, and the genre is not altered. Warns when the change rate exceeds 30% and stops above 50%. Figures and dates, proper nouns, double-quoted quotations, and legal provisions are excluded from detection
- A script first scores the input to choose a path: light (1 call), standard (diagnosis + editing, 2 calls), or heavy (diagnosis, editing, finalize, 3+ calls). Measurements showed that a 10,000-character text took 610K tokens when chunked into 7 calls versus 134K tokens in a single call, so texts of 15,000 characters or fewer are not split. Finally,
verify_gates.pychecks the change rate, preservation of figures, and so on - v2.4 added a vocabulary layer that fixes, depending on context, 111 words (115 as of v2.4.6) that arose from translating English, such as "렌즈" (lens), "축" (axis), and "쥔" (holding). Reports that the share of such expressions remaining in real-world paragraphs fell from 94% to 34%
- The companion skill commit-ko, which only strips phrases like "~을 수행함" ("performed ~") from AI-suggested commit messages, is installed separately with
./install.sh --extras - Claude Code:
/plugin marketplace add epoko77-ai/im-not-ai,/plugin install humanize-korean@im-not-ai, then/humanize-korean. GitHub Copilot CLI, Codex CLI ($humanize-korean), and a Gemini CLI extension are also provided, but the Gemini extension is a separate rulebook containing only 47 patterns - States that the goal is improving the quality of Korean writing, not evading AI detectors. The web version imnotai.kr runs on a separate codebase
- MIT license
See also
- Skills
- Korean language
- Translation
- docs-suit - engine that assembles and checks Korean writing guidelines for the document type and audience
- Claude Code