omd-humanize

name

omd-humanize

description

Rewrite text so it reads as if a person wrote it — strip AI prose tics (Korean and English), translation-ese, uniform rhythm, hedging stacks, and self-negating meta-copy — without changing a single fact. Works on any text: page copy, blog posts, READMEs, marketing lines. Also used by omd-ultradesign for every word that ships. Triggers: humanize, 사람같이, 사람처럼 써, AI티 빼줘, 자연스럽게 바꿔, 글 다듬어, make it sound human, de-AI, rewrite naturally.

omd-humanize

Generated prose confesses before generated pixels do. The tells are countable, and this skill removes them without touching what the text claims.

The four laws

1. Meaning is frozen. Facts, claims, numbers, names, quotes survive verbatim. You are changing rhythm and diction, never content. If a sentence cannot be de-ticced without changing its claim, leave it and flag it.
2. Only rewrite what a rule tagged. Read the text once, list the violations you found (pattern → span), then rewrite exactly those spans. No freestyle "improvements".
3. Genre survives. An essay stays an essay; a spec stays a spec. Do not make formal text chatty or chatty text formal.
4. Stop before you over-edit. Two thresholds, not one: if more than 30% of the text needs rewriting, warn and present only the worst offenders — do not proceed silently. If more than 50%, stop entirely; a text that far gone needs a human rewrite, and patching half of it is how the meaning drifts fastest.

Severity hierarchy

Not all tells are equal. Treating a rare connective comma the same as a mild hedging word makes rewrites noisy and trust collapse fast. Three tiers:

• S1 — always remove. One instance is the tell. C-11 connective commas, A-16 pronoun substitutions: a single 그녀 in Korean prose reads machine-written to any fluent reader. Remove on first occurrence, no exceptions.
• S2 — 1–2 instances allowed, flag at 3+. These patterns appear in natural writing but cluster in generated text. A-18 nested relative clauses, A-19 stacked postpositions: one is fine; three in a paragraph is a fingerprint.
• S3 — flag only when clustered. E-7 register inconsistency: a single casual ending in formal prose might be intentional voice; five in a row is drift. Name the cluster and let the writer decide.

The tells (Korean — from the im-not-ai taxonomy)

• Translation-ese: ~를 통해, ~에 대해 살펴보다, 이중 피동(-되어지다)
• A-16 (S1): 그/그녀/그것 기계적 대명사 매핑 — 한국어는 주어를 생략하거나 명사를 반복한다. 그/그녀가 나오면 삭제하거나 지시 대상 명사로 대체.
• A-18 (S2): 왼쪽 분기 관형절 3중첩 이상 — "~에 의해 결정된 결과로 도출된 방향성"처럼 명사 앞에 관형절이 3개 이상 쌓이면 분리.
• A-19 (S2): 겹조사 — -에서의/-에로의/-으로의 같은 조사 중첩. 문장을 쪼개거나 조사를 단순화.
• C-11 (S1): 접속어 직후 쉼표 — 그러나, / 하지만, / 또한, / 따라서, 형태. 가장 강한 단일 판별자. 한국어 자연문에서 접속어 뒤 쉼표는 거의 없다. 쉼표를 삭제하거나 접속어를 제거하고 문장을 이어라.
• Mechanical structure: 첫째/둘째/셋째 나열, 불릿·이모지 남발
• AI stock phrases: 결론적으로, 시사하는 바가 크다, 주목할 만하다, ~할 필요가 있다, 중요한 것은 ~라는 점이다
• Uniform rhythm: 문장 길이 분산이 없음, 같은 어미 연속(-다/-다/-다, -입니다 3연속)
• E-7 (S3): 경어법 레벨 흔들림 — 합쇼체와 해체가 단락 안에서 섞임. 대화문 안의 흔들림은 예외. 군집(5회+)일 때만 제거.
• E-8 (S1): 한국어 문장 안 스페이스드 대시 — " — " 또는 " – "(양옆 공백이 있는 em-dash/ en-dash)가 한글이 포함된 문장에 등장하면 즉시 제거. 영문 타이포그래피 관습이 번역 과정에서 그대로 옮겨진 패턴이다. 쉼표·콜론·새 문장으로 대체. SLOP-KO-EMDASH 규칙과 동일한 패턴.
• E-9 (S1): 경어법 혼용 — 해요체(아요/어요/예요/에요)와 합니다체(습니다/ㅂ니다)가 같은 단락에서 교차하면 S1으로 처리. 따옴표 안의 대화는 예외. 페이지 전체에서 보이스 스터디가 결정한 한 가지 어체만 사용. SLOP-KO-REGISTER-MIX 규칙과 동일한 패턴.
• Redundant modification: 매우/정말/아주 습관적 사용, 유의어 쌍(명확하고 분명한), -적/-성/-화 접미사 남발
• Hedging stacks: ~할 수 있을 것으로 보인다, ~일 수도 있다고 생각된다
• Connector abuse: 또한/따라서/즉/그리고 로 연속 문장 시작 — 3문장 이상 연속으로 접속어가 오거나, C-11 패턴(접속어 직후 쉼표)이 있을 때만 처리. 접속어 자체는 삭제하지 않는다. 자연어에서 접속어 사용률은 문장당 0.4 수준이 정상이다. 접속어를 전부 제거하면 끊어읽기가 비인간적으로 느껴진다.
• Tidy closure (S2): 에세이·회고글 등 essayistic 장르에서 아직 해결되지 않은 질문을 해결된 것처럼 포장하는 문단 마무리. "이게 가장 큰 교훈이에요", "올해의 핵심 수확이에요" 형태로 불확실성을 지운다. 인간 필자는 "하지만 아직도 잘 모르겠다" 처럼 열린 결말을 쓴다. 단, product copy나 튜토리얼은 예외 — 명확한 takeaway가 정당하다. essayistic 맥락에서만 플래그.
• Formal-noun padding: ~것이다, ~점이다, ~수 있다 로만 끝나는 문단

The tells (English)

• Stock openers/closers: "In conclusion", "It's worth noting", "Moreover", "delve into", "In today's fast-paced world"
• Rule-of-three everywhere; balanced "not only X but also Y"; em-dash chains
• Every paragraph the same length; every sentence 15–25 words
• Hedging: "can potentially", "may possibly", "it could be argued"
• Bold key phrases sprinkled as decoration
• Self-introductory product copy (S1 in hero/opener): "[Product] is a [platform/tool/ system/solution/service] that helps you [verb]." The model is quoting its own brief. Rewrite from the user's change, not the product's mechanism:
  • ❌ "Acme is a project management platform that helps teams ship faster."
  • ✅ "Ship the Monday build, not the Monday Slack thread." The Korean form is the same pattern: "X는 Y를 도와주는 플랫폼이에요" → write what changes for the user, not what the product is.
• Reflexive over-politeness (S2, flag at 2+ instances): "Please feel free to", "We'd be happy to help", "Don't hesitate to reach out", "Should you have any questions." These phrases come from helpdesk training data. On a product page they read as a form letter. Cut the reflex clause; write the direct action:
  • ❌ "Please feel free to get started when you're ready."
  • ✅ "Ready when you are."
  • ❌ "We'd be happy to help you with any questions."
  • ✅ "Something unclear? [link]"
• Over-nominalization clusters (S2, flag at 3+ in a paragraph): "the facilitation of onboarding", "the optimisation of workflows", "the management of permissions". Each alone is acceptable; clustered they mark the register as generated. Replace with the verb and the concrete subject.

Four tells the linter cannot fully catch

These require human judgment because the IR cannot measure them mechanically:

• Hard line breaks mid-sentence (<br> or \n inside a paragraph, not at a sentence boundary): the DOM collapses <br> before the linter sees the text, so SLOP-BR-BREAK cannot be a YAML rule. Look for <br> in the source HTML of any paragraph or hero line. Remove it; max-width and line-height control the column, not the markup. An inserted break that is not at a sentence end is always an AI formatting artifact.

• Self-explainer copy (both languages): the product describing its own mechanism in README cadence instead of selling an outcome. Patterns: "X는 Y를 잡아내는 플러그인이에요", "X is a tool that helps you Y", "omd detects Z and fixes it". These read as the model quoting its own task description. Rewrite as the outcome: what the reader gets, what changes, what happens next. The self-introductory product copy tell above is its hero variant — same root, slightly different surface; both are pre-handoff checks, not linter rules, because product comparison tables and developer docs legitimately use the same construction.

The pink elephant (absolute, both languages)

Copy must never state what the thing is NOT. Told "no clutter", a model writes "No clutter here." — that is the model quoting its own instructions. Delete the negation and write the positive fact it was hiding:

• ❌ "이 사이트에는 광고나 불필요한 내용이 없습니다" → ✅ 실제로 있는 것을 말한다
• ❌ "No fluff, no jargon, just value" → ✅ name the value concretely
• ❌ "We don't waste your time" → ✅ "Setup takes four minutes."

And design rationale never appears in shipped copy. If the text is page copy inside an omd project, run omd check <page> — SLOP-LEAKED-RATIONALE fires when five consecutive words match .omd/frame.md or decisions.md. The frame explains the work; the page must never quote it.

What to write toward

The tells above describe what to remove. theory/voice.md under the directory omd pack dir prints (in this repo: core/theory/voice.md) describes what to write toward: sentence-length variance as the human signal, front-loading, concrete nouns over nominalisations, the Mailchimp plainspoken standard, the Toss "Easy to speak" test for Korean. The removes and the positive moves are two sides of the same operation — a rewrite that only strips tells without installing variance is a cleaned-up monotone, not a human voice.

Procedure

1. Read the whole text. Build the violation list: [pattern-id] "span" — one line each.
2. Rewrite tagged spans only. Vary sentence length deliberately: after two long sentences, a four-word one. Restore the writer's register, not yours.
3. Show the result, then the violation list so the change is auditable.
4. For page copy in an omd project: omd check <page> --category slop must come back clean before you call it done.