hashgraph-online/awesome-codex-plugins / plugins/beefiker/superloopy/skills/humanize-korean
humanize-korean
VerifiedKorean prose humanizer for Codex and Superloopy that rewrites Korean text so it sounds naturally human while preserving meaning, register, facts, protected tokens, and evidence. Use when the user asks to remove Korean AI tells, make Korean copy sound human, fix 번역투, remove ChatGPT/Claude/Gemini tone, polish Korean without changing meaning, or says "AI 티 없애줘", "AI 윤문", "번역투 고쳐", "사람이 쓴 것처럼", "humanize Korean", "한글 AI 티 제거", "글 자연스럽게 다듬어줘". Inspired by Korean AI-tell ideas from https://github.com/epoko77-ai/im-not-ai. Handles Korean rewriting only; not for translation, fact expansion, SEO rewriting, legal drafting, or generic proofreading.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 9.4 KB |
| agents/openai.yaml | config | 315 B |
| references/golden-set.md | markdown | 13.6 KB |
| references/quality-rubric.md | markdown | 4.1 KB |
| references/quick-rules.md | markdown | 20.6 KB |
| references/upstream-notice.md | markdown | 3.3 KB |
| scripts/audit-humanize-output.mjsexecutable | script | 16.6 KB |
| scripts/calque-patterns.mjs | script | 7.2 KB |
scripts/calque-patterns.mjs
View on GitHub// Korean adverb and compound-term calques (English carried across word-for-word) that mark
// LLM-written technical prose. Measured on GitHub issues 2026-09-10: 조용히 grew 381x from
// pre-ChatGPT to 2026 against 4x for control vocabulary, with a third of hits carrying an
// AI-authorship footer. See docs/specs/2026-09-10-korean-calque-rules-design.md.
//
// Single source for the P family. The audit imports it; any Stop-time nudge or doctor check
// must import it too rather than copying a regex, so the tiers cannot drift apart.
// Program actions a system performs "silently". Verb stems (잘리, 멈추, 죽) match any inflection;
// Sino-Korean nouns (실패, 무시, 누락, 중단, 종료, 삭제) must carry verbal morphology (실패한다,
// 무시됨, 종료된) so the bare noun in 실패 원인 or 무시 사례 is not a program action. Up to two
// intervening tokens absorb an object or adverb (조용히 H2로 기동된다).
const PROGRAM_ACTION = "실패(?:한|했|합|해|하|다|로|함)|무시(?:한|했|합|해|하|된|됐|됩|되|당|함)|누락(?:된|됐|됩|되|한|했|시|됨)|중단(?:된|됐|됩|되|한|했|시|됨)|종료(?:된|됐|됩|되|한|했|시|됨)|삭제(?:된|됐|됩|되|한|했|됨)|사라|잘리|잘린|잘림|잘라|건너뛰|죽|깨지|깨진|멈추|멈춘|멈춤|비어|빠지|빠진|넘어가|덮어|no-?op|skip";
// Human actions and settings that 조용히 legitimately modifies. Anything else after 조용히 is
// the calque: programs do not act 조용히, people do.
const HUMAN_ACTION = "앉|말|얘기|이야기|대화|속삭|물러|떠나|지내|기다리|듣|들었|웃|울|걷|살|잠|쉬|지켜|바라|생각|읊|읽|되짚|돌아보|살펴|회고|분석|검토|정리|기록|관찰|고민|준비|공부|연습|하세요|하자|해라|밤|아침|분위기|공간|마을|동네|사람|성격|목소리";
export const P1A_SILENCE_PATTERN = new RegExp(`조용(?:히|한)\\s*(?:[가-힣A-Za-z0-9_.%]+\\s+){0,2}(?:${PROGRAM_ACTION})`, "gu");
export const P1B_SILENCE_PATTERN = new RegExp(`조용(?:히|한)(?!\\s*(?:[가-힣]+\\s+){0,2}(?:${HUMAN_ACTION}))`, "gu");
export const P2_GRACEFUL_PATTERN = /우아(?:하게|한)\s*(?:[가-힣]+\s+)?(?:종료|실패|축소|처리|중단|저하|대응|폴백|재시도)/gu;
export const P3_TRANSPARENT_PATTERN = /투명(?:하게|한)\s*(?:[가-힣]+\s+)?(?:처리|동작|작동|지원|전달|연결|통합|전환|교체|대체|적용|마이그레이션|프록시|캐시|재시도|폴백|위임|포워딩|라우팅)/gu;
// 투명 in a sentence about rendering is alpha transparency, not the "invisible to the caller"
// calque. Sentences carrying these words are dropped before P-3 is counted.
export const P3_ALPHA_CONTEXT_PATTERN = /배경|색|알파|투명도|이미지|레이어|opacity|png/iu;
export const P4_SSOT_PATTERN = /단일\s*진실\s*(?:공급원|소스|원천|의\s*원천)/gu;
// No trailing boundary: particles attach directly (정본을, 정본으로). The leading boundary
// already excludes 수정본 and 개정본.
export const P5_CANONICAL_PATTERN = /(?<![가-힣])정본/gu;
export const P6_FAILSAFE_PATTERN = /안전하게\s*실패|(?<![가-힣])첫날부터/gu;
// Quoted spans, inline code, and fenced code blocks are protected byte-for-byte, so a calque inside
// them is the source's, not the writer's. Every consumer strips these before counting; shared here so the audit and the
// write-time nudge cannot disagree about what a quote is.
export const PROTECTED_PROSE_SPAN_PATTERN = /```[\s\S]*?```|`[^`\r\n]+`|"[^"\r\n]+"|“[^”\r\n]+”|‘[^’\r\n]+’|「[^」\r\n]+」|『[^』\r\n]+』/gu;
export function removeProtectedProseSpans(text) {
return text.replace(PROTECTED_PROSE_SPAN_PATTERN, "");
}
// Hangul share of all letters; the audit's Korean-text threshold is 0.2.
export function koreanRatio(text) {
const hangul = text.match(/[\u3131-\u318E\uAC00-\uD7A3]/gu)?.length ?? 0;
const letters = text.match(/[\p{L}]/gu)?.length ?? 0;
return letters === 0 ? 0 : hangul / letters;
}
// Fail tier: grade A/B requires these at zero in the rewrite.
export const CALQUE_GATING_PATTERNS = [
["P-1a", P1A_SILENCE_PATTERN],
["P-2", P2_GRACEFUL_PATTERN],
["P-4", P4_SSOT_PATTERN]
];
// Warn tier: reported, never graded. P-3 is advisory because developer prose uses 투명하게 for
// alpha and disclosure far more often than for the calque; P-5 keeps its legal sense; P-6 has
// unproven AI attribution (its footer share sits below baseline) and 첫날부터 can be literal.
export const CALQUE_ADVISORY_PATTERNS = [
["P-1b", P1B_SILENCE_PATTERN],
["P-3", P3_TRANSPARENT_PATTERN],
["P-5", P5_CANONICAL_PATTERN],
["P-6", P6_FAILSAFE_PATTERN]
];
export const CALQUE_PATTERN_IDS = [...CALQUE_GATING_PATTERNS, ...CALQUE_ADVISORY_PATTERNS].map(([id]) => id);
// One ladder step per id, quoted in audit warnings. Deletion first; a symptom stated from
// supplied facts second; never a stock replacement phrase (정본, 페일세이프, 그레이스풀 and the
// ~없이 family measure as AI tells themselves).
export const CALQUE_REPAIR_HINTS = {
"P-1a": "조용히 is for people. Delete it when the verb already implies unnoticed behavior (조용히 무시된다 → 무시된다), or state the observed symptom from supplied facts.",
"P-1b": "조용히 modifies a program action here. Say what the program did and which signal it did not give; do not swap in 아무 표시 없이 or 티 안 나게.",
"P-2": "우아하게 is a calque of gracefully. Delete it and keep the plain verb (종료한다) unless the source states what happens to in-flight work; never assert on your own that requests finish, drain, or are cancelled. A developer genre may keep `graceful shutdown` in backticks.",
"P-3": "투명하게 means disclosed in Korean, not unnoticed by the caller. If the source means the latter, say 사용자가 손댈 것 없이 or 코드 변경 없이; if it means alpha or disclosure, keep it.",
"P-4": "단일 진실 공급원 is a calque of single source of truth. Use 기준 데이터 or state the fact (설정은 이 파일 하나만 본다); developer genres may keep `SSOT` in backticks.",
"P-5": "정본 in the canonical-data sense is itself an AI tell (4,631x since 2022). Prefer 기준 데이터 or name the file; keep the legal sense.",
"P-6": "안전하게 실패 reads as a contradiction in Korean; say what state the failure leaves (실패하면 기본값으로 되돌린다). 첫날부터 is filler unless a real date replaces it."
};
// Count tier-B silence hits that tier-A did not already claim, so one 조용히 is never
// reported twice. Both patterns anchor on the same 조용 token, so a shared start index is
// the whole overlap test.
export function countP1bOnly(text) {
const claimed = new Set([...text.matchAll(P1A_SILENCE_PATTERN)].map((match) => match.index));
return [...text.matchAll(P1B_SILENCE_PATTERN)].filter((match) => !claimed.has(match.index)).length;
}
// Drop sentences that talk about rendering before counting P-3. Sentence boundaries are
// terminal punctuation or a line break; Korean prose rarely needs more than that here.
export function stripAlphaContextSentences(text) {
return text
.split(/(?<=[.!?。])\s+|\n/u)
.filter((sentence) => !P3_ALPHA_CONTEXT_PATTERN.test(sentence))
.join("\n");
}
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 468k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |