writing-humanizer-ja

name	writing-humanizer-ja
description	Strips the "AI tells" from Japanese text and finishes it so it reads as if a human wrote it. Built on top of a Zenn article, Wikipedia's "Signs of AI writing", blader/humanizer, and matsuikentaro1/humanizer_academic. Apply as a final pass to any Japanese text generation — briefings, reports, notes, proposals, etc. Triggers: "AIっぽい" / "人間らしく" / "文章を自然に" / "文体改善" / final-pass on text generation.

writing-humanizer-ja — Anti-AI Japanese writing

Detects and strips "AI tells" from LLM-generated Japanese text so it reads as if a human wrote it.

Important: pattern stripping alone produces "sterile-room writing", which is also AI-like. The two-stage core of this skill is: strip the patterns first, then inject "soul".

This skill operates on Japanese text. The detection patterns and corrections below target patterns that show up in Japanese writing, so the examples are kept in Japanese — these are the strings the skill is supposed to catch.

---

1. When to apply

Apply this skill as a final pass to any writing situation:

• Briefings, daily summaries, weekly reviews
• Research reports, analysis notes
• Proposals, planning documents
• Chat notifications, emails
• Blog posts, social media posts
• Documentation, READMEs

skills/long-form-analysis/SKILL.md Step 8, and any text intended for external publication, are designed to always pass through this skill.

---

2. Phase 1: detection — 20-item checklist

A. Content patterns (meaning / content level)

1. Excessive emphasis of significance 🔴

The tic of stating the importance of a fact right after presenting it. The reader can judge for themselves.

• NG: 「HR は 0.65 でした。これは心保護効果の重要性を浮き彫りにしており、今後の治療戦略に大きな示唆を与えています」
• OK: 「HR は 0.65 だった」

Detection words: 「〜を浮き彫りにして」「〜を示唆して」「〜の重要性を示して」「〜を物語って」「〜を裏付けて」

2. Formulaic openings and closings 🔴

An opening of 「ここでは〜について解説します」 and a closing of 「今後の展開が注目されます」 reads as AI-written on contact.

• NG opening: 「本記事では」「ここでは〜について詳しく見ていきます」「〜が注目を集めています」
• NG closing: 「今後の展開が注目されます」「引き続き注視していく必要があります」「〜が期待されます」
• OK: open with a specific question or scene, close with a lingering note or a new angle

3. Generic "challenges and outlook" section 🟡

The template structure of "challenges exist, but future progress is expected".

• NG: 「〜にはまだ課題が残されているものの、今後の技術的進歩により解決が見込まれる」
• OK: write the challenge concretely. Skip the outlook

4. Vague citations / appeals to authority 🔴

Substanceless 「専門家によると」「研究では」.

• NG: 「多くの専門家が指摘するように」「研究によると」「業界関係者の間では」
• OK: 「<著者名> が <媒体> のインタビューで述べたように」

5. Promotional / travel-guide modifiers 🟡

• NG: 「画期的な」「革新的な」「包括的な」「堅牢な」「シームレスな」
• OK: write specifically what is good and how. Do not let adjectives carry the weight

B. Language patterns (vocabulary / grammar level)

6. AI-frequent vocabulary (Japanese edition) 🔴

When three or more of these cluster together, treat as an AI signal:

Watch word	Human-sounding alternative
さらに / 加えて / また (consecutive)	Drop the connective and start the sentence. Or 「それと」「あと」
重要なのは	(drop it and write the substance)
〜において	〜で
〜に関して	〜について、〜の
〜を活用する	〜を使う
〜を実現する	〜をやる、〜にする
〜に寄与する	〜に役立つ
〜を促進する	〜を進める
包括的な / 網羅的な	(write specifically what it covers)
〜の観点から	〜から見ると
多角的に	いろんな角度で
〜が求められる	〜が要る、〜しないといけない

7. Roundabout copula avoidance 🔴

Going long where 「〜です」 would do.

• NG: 「〜として位置づけられています」「〜の役割を果たしています」「〜としての側面を持っています」
• OK: 「〜です」「〜だ」

8. Tacked-on -ing-style addition 🔴

A literal translation of English participle constructions: tacking on the significance at the end of the sentence.

• NG: 「売上は前年比 20% 増でした。これは市場の成長性を示しており、今後のさらなる拡大が期待されます」
• OK: 「売上は前年比 20% 増だった」(let the reader infer the meaning)

Detection pattern: chains of 「〜しており、」「〜を示しており、」「〜を反映しており、」

9. Pointless synonym cycling (elegant variation) 🟡

Re-paraphrasing the same thing every sentence.

• NG: 「ユーザー」→「利用者」→「参加者」→「対象者」 (all referring to the same people)
• OK: use the same word consistently. Do not fear repetition

10. Forced rule of three 🟡

The tic of always lining up three things.

• NG: 「スピード、品質、信頼性の 3 つが重要です」 (are there really three?)
• OK: if two things matter, write two; if four, write four. Do not pad to fit a count

11. Overuse of 「〜だけでなく〜も」 (negative parallel) 🟡

• NG: 「コスト削減だけでなく、品質向上にも貢献します」
• OK: 「コストが下がる。品質も上がる」 (write them as separate sentences)

12. Excessive hedging 🟡

Avoiding assertion by reaching for 「かもしれない」「可能性がある」.

• NG: 「〜の可能性が示唆されていると言えるかもしれません」
• OK: if it is uncertain, state the uncertainty and then state the analyst's view

C. Style patterns (formatting / look level)

13. Overuse of em-dash (—) 🟡

Inserting full-width dashes is a classic AI tic.

• NG: 「SGLT2 阻害薬—比較的新しい薬剤クラス—は」
• OK: 「SGLT2 阻害薬 (比較的新しい薬剤クラス) は」 — substitute parentheses or commas

Rule: full-width dashes (—) are banned. Rewrite with parentheses or commas.

14. Mechanical bold overuse 🟡

Bolding every keyword indiscriminately.

• NG: 「スピード が大幅に向上し、品質 も改善、コスト は削減されました」
• OK: bold the one place that genuinely deserves emphasis. Or skip bold entirely

15. Inline-header bullet list 🔴

A pattern that shows up constantly in LLM output.

• NG: 「- 速度: コード生成が大幅に高速化されました」「- 品質: 出力品質が向上しました」
• OK: integrate as prose. Or drop the inline-header bolding

16. Uniform sentence length 🔴

Every sentence is the same 18-22 words. Human writing has irregular rhythm.

• OK: a long sentence. A short one. Then long again. Very short. That is rhythm.

D. Communication patterns (dialogue residue)

17. Chatbot residual phrasing 🔴

• NG: 「承知しました!」「いい質問ですね!」「ご質問ありがとうございます」 (left over inside the document)
• OK: dialogue acknowledgments do not belong in the document

18. Sycophantic tone 🟡

• NG: 「素晴らしい取り組みですね」「非常に的確なご指摘です」
• OK: when assessment is needed, be specific. 「この切り口は 〇〇 の観点で有効だと思う」

19. Knowledge-cutoff excuses 🟡

• NG: 「最新の情報は確認できていませんが」「私の知識の範囲では」
• OK: write it or do not. Skip the excuse

20. Template-style placeholders 🟡

• NG: 「[ここにデータを挿入]」「〜については別途確認が必要です」
• OK: leave it blank if it cannot be filled, or state specifically what is missing

---

3. Phase 2: correction — pattern removal

Apply the following principles to fix detected patterns.

3.1 The 5 correction principles

1. Deletion is the strongest edit: redundant modifiers / connectives / hedges become better just by being cut
2. Replace abstraction with specificity: abstract adjectives → numbers / proper nouns / anecdotes
3. Do not fear short sentences: 「〜であり、〜しており、〜と言える」 → split into 3 sentences
4. Do not fear repetition: repeating the same word is easier to read than cycling through synonyms
5. Trust the reader: significance does not need to be spelled out. Show the data and the reader understands

3.2 Before / after example

Before (AI-tasting):

さらに、本プロジェクトはコスト削減だけでなく、品質向上にも大きく寄与しており、今後のさらなる発展が期待されます。包括的なアプローチにより、多角的な改善が実現されました。

After (human-sounding):

コストは月額で 12 万円減った。バグの発生率も半分になった。

---

4. Phase 3: inject soul — exit the sterile room

Removing every pattern still produces "clean, but who wrote this?" writing. This is the real work.

4.1 The 4 soul principles

1. Take a position

Do not finish with neutral reporting. Include 「〜だと思う」「〜は間違いだと考えている」.

• NG: 「A と B の両方にメリットがあります」
• OK: 「ぼくは B 派。A は理屈では正しいけど、運用で破綻する」

2. Break the rhythm

Polished writing reads mechanical. Break it on purpose.

• Drop in short sentences. Cut.
• Mix in noun-final endings.
• Sprinkle in colloquial phrasing.

3. Acknowledge complex emotion

AI tends to binary out into "good" or "challenging". Human emotion is messier.

• NG: 「懸念される」
• OK: 「深夜 3 時にエージェントが勝手にコードを書いてるのを想像すると、ちょっと落ち着かない」

4. Sense concretely

Write with the senses and scenes, not abstract impressions.

• NG: 「印象的だった」
• OK: 「スライドの 3 枚目で会場が静まり返った。あの沈黙が答えだった」

4.2 Soul-injection techniques

Technique	Description	Example
Use first person	Make the subject explicit with 「私は」 / 「ぼくは」	「ぼくはこの数字を見て方針を変えた」
Bring your own experience	Anecdote over abstraction	「前職で同じ失敗をした。3 ヶ月無駄にした」
Ask a question	Throw it to the reader	「本当にそうだろうか?」
Show the turn	Leave the thought process visible	「最初は A だと思ったけど、データを見て考えが変わった」
Admit uncertainty	Be honest where you cannot assert	「正直、ここはまだ確信が持てない」
Humor / self-deprecation	In moderation	「3 回目のピボットで、さすがに笑えてきた」

---

5. Phase 4: anti-AI pass — self-audit

Once the writing is finished, run the following as a final check.

5.1 Self-audit procedure

1. Read the opening: is it a formulaic opening? Would the first sentence read as "AI wrote this"?
2. Read the closing: did it land on something like 「今後の展開が注目されます」?
3. Count the connectives: do 「さらに」「加えて」「また」 appear 3+ times in a row?
4. Count bolds: are there 3+ bolds in a single section?
5. Search for em-dashes: any full-width dashes (—) remaining?
6. Check sentence length: is everything roughly the same length? Is there rhythm?
7. Look for an opinion: is there at least one of the author's stance / impression / judgment?
8. Self-question: "where does this look AI-generated?" → fix what you find

5.2 Pass criteria

• 0 items from the Phase 1 checklist trigger
• At least one of the author's opinion / judgment / experience is present
• Sentence length varies
• Opening and closing are not template patterns

---

6. Context-dependent application strength

The skill is not applied at the same strength to every text type:

Context	Phase 1-2 (pattern removal)	Phase 3 (soul)	Phase 4 (audit)
Blog post / social	Full	Full	Full
Research report	Full	Light (one observation note)	Full
Internal briefing	Full	Light (add a judgment)	Light
Proposal / plan	Full	Medium (carry the proposer's conviction)	Full
Chat notification	Light (remove redundancy only)	Skip	Skip
Technical documentation	Full	Minimal	Light

---

7. Continuous update

• Wikipedia's "Signs of AI writing" is continuously updated. As AI writing shifts, new patterns appear
• LLMs pick "the most likely next token" probabilistically, so new tics can emerge even after known patterns are stripped
• Periodically re-read your own output and add new AI signals to this skill when you notice them

---

8. Related skills

• skills/long-form-analysis/SKILL.md — applied as the final pass in Step 8
• skills/structured-minutes/SKILL.md — final style pass on the minutes
• skills/knowledge-extract/SKILL.md — knowledge-extract's output is structured data, so this skill is not applied to it (mechanical formatting takes priority)

---

9. References

• AI 生成文から「AI くささ」を取り除く技術 (Zenn)
• Wikipedia: Signs of AI writing
• blader/humanizer
• matsuikentaro1/humanizer_academic

---

License

MIT (this skill). See LICENSE at the repo root for details.