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 |
references/quality-rubric.md
View on GitHub# Humanize Korean Quality Rubric
## Required Gates
- Korean source ratio: at least 0.2 Hangul characters over all letters.
- Protected tokens: 100% preserved.
- Change rate: pass at 0.3 or lower, warn above 0.3, fail above 0.5.
- S1 AI-tell count: lower after rewrite, target zero when possible.
- S2 repeated-pattern count: lower after rewrite unless the pattern is genre-appropriate.
- Em dash (M-1): zero `—`/`–` in Korean prose after rewrite; dashes inside code spans and quoted spans are exempt.
- N-1 — misplaced modifier target: 정확성은 시간·수치·사양·정보·식별·일치처럼 확인 가능한 대상에 붙인다. 정확한 컴퓨터/보드/펌웨어 이미지는 공급된 관계에 따라 대상 컴퓨터 확인, 보드 모델 확인, 보드와 일치하는 펌웨어로 고친다. 정확한 시간/수치/사양/정보는 보존한다. N-1 has no deterministic audit pattern or grade effect.
- R family (paragraph and document shape, R-1 to R-7): judged by the rewriter from `references/quick-rules.md`; no deterministic audit pattern or grade effect. An R-7 reply cut is declared with `--cut R-7`: the final must be shorter than the source, protected tokens and links still gate it, and the change-rate failure line moves from 0.5 to 0.8 with a warning that names the cut; any other grade cause still applies.
- Modality preservation (서법): deontic (`~해야 한다`) and hedge (`~일 수 있다`) marker counts must not decrease from source to final; a decrease warns for review because A-10/G-2 repairs may only drop a hedge when the source itself is certain. Repositioning (D-6) keeps counts identical.
- Antithesis repetition (C-8): two or more remaining `~가 아니라 ~` / `~인가, ~인가` pairs warn; keep the strongest one. Human writers do use the device (31 of 532 clean human documents upstream), so a chain of the same frame is the signal, not presence.
- Injection guard: any counted pattern that rises from source to final warns by id. A repair that removes one tell must not create another elsewhere; upstream measured rewrites planting `결국`, `~하는 이유다`, `더 이상 A가 아니라 B`, and connective commas in sentences that lacked them.
- Register: unchanged.
- Added claims: zero.
## Grades
- A: all gates pass, S1 after count is zero, change rate is 0.1 to 0.3.
- B: all required gates pass, S1 after count is zero, S2 after count is four or fewer.
- C: protected tokens pass, but change rate warns or some S1 remains after a real reduction.
- D: protected tokens changed, non-Korean input, change rate above 0.5, S1 count is not reduced, or new claims were added.
## Notes
- Quoted spans (J-2) are protected byte-for-byte, so the audit reports them without letting them affect the grade.
- The change rate divides edit distance by the longer of source and final, so explanation-style repairs that expand text are not over-penalized.
## Judgment Boundaries (adapted from upstream v2.6 false-positive principles)
- The audit reports pattern counts, deltas, and a rewrite grade. It is not a document-level "AI or human" verdict, and no single id is evidence of one: upstream reproduced a pre-ChatGPT human essay scored maximum-risk by one comma metric alone, because that writer's personal comma rate (83%) exceeded the model average. Tells count when they overlap.
- Individual style variance can exceed the gap between the human and AI distributions. When a rewrite touches a writer's recurring habit (a favored connective, a rhetorical pair, a hedge), report it as a judgment call rather than a defect, and prefer the writer's other texts as the baseline when they exist.
- Flawless register is itself a machine signal. Small human deviations (a colloquial word in formal prose, an idiosyncratic spacing habit) are preserved, not corrected; the skill removes tells, it does not polish toward a ceiling.
- Discourse-level absences (no time anchor, no named people, no emotional spike, no speaker self-limitation) are observations, never repairs: filling them invents facts. The one allowed move is relocating a dated fact or quotation the source already contains to the opening.
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: 471k | 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: 322k | 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: 322k | 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: 322k | Updated | Claude CodeCodex | Verified | Install |