SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillshumanize-writing

humanize-writing

Writing & research

View repository

Updated 10-15 16:46

About this skill

Humanize Writing

A Skill that detects AI-generated-ness and restores text to a form that reads as if written by a human.

The core is not adding more explanation, butrestoring traces of the writer's judgment. Rather than lining up polished truisms, prioritize writing in which "why it was done that way" and "what was emphasized" are visible.

When to Use

  • AI-ness audit of a draft (review type)
  • Humanize / style revision of a draft (conversion type)
  • Creating generation instructions to make writing human-like from the start (generation-instruction type)
  • Japanese articles (Qiita / WordPress / note), English articles (Medium), technical documentation, README
  • Keywords: AIっぽさ, humanize, 人間らしく, 文体チェック, AI感, ChatGPT tells

Scope

  • Bold text, bullet lists, sentence length, commas, and consecutive sentence endings are not judged by impression but counted with scripts/text_metrics.py (standard library only. Code blocks, tables, and link URLs are excluded). Thresholds are guidelines for candidate issues; judge by looking at the genre and quotations.
  • If another Japanese proofreading/style-adjustment Skill is active at the same time, the instructions interfere and the output becomes unnatural. When concurrent use is suspected, stop one and run again.
  • A style policy to keep active throughout the entire article-writing process should be placed in the workspace's always-applied instruction (if any).
  • For the detailed tell dictionary, refer to references/ai-likeness-signals.md .
  • For long additional instructions in generation-instruction mode, refer to references/generation-style-prompt.md .

Operating Rules

  • Do not change the meaning or facts. After revising, separately check for the omission of conditions, numbers, and constraints that were in the original, and for the addition of subjects, specifications, or generalities that were not in the original.
  • Do not add facts or rewrite into unverified assertive terms.
  • Increase assertions only when the basis exists in the existing body text.
  • For topics that change over time, such as pricing, specifications, and previews, briefly indicate the public Docs and the point of investigation as needed. If you compared using assumed token amounts or exchange rates, make the conditions and the exclusion of caching and long-context fees explicit, and do not confuse them with the actual billed amount.
  • Match polite/plain style, emoji, and level of warmth to the existing dominant style.
  • If the user provides past articles or example sentences, prioritize that voice.
  • Do not judge as AI based on a single symbol, polished style, or polite structure alone. See whether it can be explained as a cluster of multiple tells.
  • Do not uniformly cut metaphors or characteristic words based on impression. Judge word by word by looking at the number of occurrences across the whole work, distribution within chapters, and role in context. Words that function across multiple chapters should be kept as voice, while repetition concentrated in the same chapter, or one-off phrases that stand out from the surrounding style and level of abstraction and merely restate nearby specifics, should be cut or made concrete. For metaphors that are set up in the introduction and paid off in the final chapter, you can keep both ends and tidy up only the middle. Even for two words in the same paragraph, it may be a judgment to cut one and keep the other.
  • Do not mechanically delete all cautionary terms. Words that are intentionally effective on social media or in spoken language may be kept.
  • Do not fool yourself into thinking you have fixed something by replacing synonyms one-to-one. Swapping a forbidden word for another vague word does not fix the syntax. Applying rules excessively increases exaggerated coinages, so fix by restoring the subject-predicate and the agent and checking whether the meaning makes sense in one sentence. In particular, for Japanese 速い / 早い , first separate whether you mean speed, timing, or ease of understanding, and if necessary しやすい や 早く〜できる unravel into.
  • For terms around products, billing, and contracts, prioritize natural industry Japanese over literal translation. If necessary, add the official English in parentheses. The same applies to technical terms; do not translate heartbeat as 心拍 but as ハートビート . On the other hand, for words that work in IT (決定論的 , etc.), do not force a paraphrase; add a plain paraphrase in parentheses on first occurrence (e.g., deterministic (rule-based processing)).コンテキスト トークン プロンプト Do not re-translate established katakana terms like文脈 either (changing it to
  • The original article's experiment/statistics jargon (Treatment A / B / Cohort / Segment / Variant / Bucket) is the same. In Japanese prose, use natural Japanese (パターン A / B / グループ / セグメント), and only on first occurrence add a parenthetical such as パターン A (元記事内では Treatment A) . Technical labels such as scorecard column labels / headings / implementation names / feature flags technical labels remain in English. AI tends to leave English as is, so monitor with a humanize pass.
  • After fixing, check whether it has drifted too far toward safe, edifying prose.
  • In verification articles, separate screen displays, measured results, completion results, official docs, and unverified scope. Do not turn unimplemented operational proposals into habits with 'usually' or 'I look at ~', and do not write disclaimers like 'not measured' or 'could not confirm' that do not change the reader's judgment. For unconfirmed items that affect the reader's judgment (such as pricing), do not stack negative sentences; consolidate them in one place in the form of 'where to confirm' and 'what to ask'. Keep the main text focused on the conclusion and necessary caveats, and move detailed logs to tables, appendices, or notes.
  • If the same caveat or premise (e.g., 1 回分の表示、試算であり実請求額ではない) is repeated in multiple sections, steps, or summaries, count the occurrences and, except for the TL;DR summary, consolidate to the one place where it is most necessary. Before deleting, confirm that the same condition is actually written in the place you are keeping, and move conditions that exist only in the deleted sentence (usage cautions, differences in comparison conditions) to the place you are keeping.
  • Before cutting evidence, measurement conditions, or caveats because of exceeding the character limit, confirm the medium's upper limit and the length the user allows. If the allowable range is unclear, confirm rather than compressing on your own. When shortening, keep sentences with concrete examples, actions, and results, and delete abstract summary sentences. Deleting concrete sentences and keeping only summaries makes it shorter but increases the AI smell. Also re-read sentences newly written during shortening or rewriting against the tell dictionary.
  • Lack of experience or judgment that cannot be fixed by style correction alone should not be fabricated; return it as a gap such as 【← ここに〜】 .
  • Expressions that refer to oneself by a third-person nickname（◇◇視点 ◇◇個人の ◇◇的にはIf you find ) 振り返り 個人の 私は replace it with a neutral expression such as . Only greetings and signatures directly below the title are permitted. Since this is a long-term trigger that users find painful, do not overlook even a single instance.

Workflow

  1. Determine the purpose of the input.
    • Audit only: Return the problem locations, categories, and suggested replacements.
    • Fix: Return before / after or a diff proposal.
    • Generation instruction: generation-style-prompt.md Create a prompt based on .
  2. Check the medium, intended readers, and voice.
    • Qiita / WordPress / note / Medium / README, etc.
    • Only for technical terms that the intended readers cannot be assumed to know, add a brief explanation at first mention.
    • If there are sample sentences, match the sentence length, paragraph length, connectives, punctuation, and strength of assertion.
  3. Scan in the following order.
    • Have the facts changed?
    • Are the writer's subject and judgment visible?
    • Is it over-explaining with abstract words or connectives?
    • Is the structure too uniform?
    • Are there any remaining glyph corruptions mixed into long-form generation (Generation Artifacts）。
    • Is the subject a thing and the verb a metaphor (break, knock down, melt, crush),これ 両者 片方 Can what refers to be returned to a concrete noun (Japanese Signals）。
  4. If detailed tells are needed, use ai-likeness-signals.md .
  5. If multiple tells overlap in the same location, combine them into one item as stacking pattern .
  6. In conversion mode, after the first draft, briefly list 'reasons it still looks AI-like,' and fix only the necessary locations in the final draft.
  7. Finally, when read aloud, check whether the sentences are ones you could say as-is to a colleague or acquaintance.
  8. For important drafts where same-model self-review is insufficient, run one different-model or subagent skeptic pass. Verify quoted text and line numbers against the actual draft before acting on feedback.

Fix Order

If time is limited, process from top to bottom. Fixing only the lower items will still leave an AI smell.

  1. Stance — Is there a falsifiable, concrete claim?重要だ 本質的だ Do not end with .
  2. Agency — Eliminate false agency (things as subjects with human verbs), and write who did what.
  3. Structure — Break proposition-style headings (H2 / H3), template introductions, and lack of variation.
  4. Vocabulary — Remove favored words, loanword metaphors, and adverb stacks.
  5. Symbols — Tidy up em dashes, unnecessary「」, middle-dot parallelism, and decorative emoji.

High-priority Signals

  • Stock connectives: つまり 要は 言い換えると Summarize the explanation with .
  • Abstract words: かなり 十分 活用 効率化 価値 Round it off with .
  • Pivot construction: A ではなく B A というより B Place a contrast instead of an observation with . Starting a clause with A だけでは決めません and a negation is of the same kind. Even in comparisons, do not place negative evaluations like 〜級とは言えません 〜に届かない見込み ; write only the facts of the comparison target and the numbers.
  • Example bias: In an article about how to use different tools, the examples lean only toward code fixes, tests, and log investigation. Also include readers' daily work such as summarizing, research, and document creation, and for examples the author has not tried, write them as recommendations in the form 'if ..., ... is enough.'
  • Binary contrast of abstract metaphors: 治療になるか加速剤になるか Strong metaphors are placed as binary choices to make claims look bigger. If you can write what happens as it is, prioritize the phenomenon over the metaphor. The same goes for headings; instead of phrasing like〜と同じ線上にあります 〜の延長です , state the content directly (e.g., クラウドの安いモデルに回すのも手です). Pointed out by a user.
  • Judgment-log exposure: 元の図を採用しなかった理由 Do not let production decisions or verification processes like this flow into the TL;DR / headings / body / sources section. Keep the source URLs, measurement conditions, and caveats that readers need.
  • Declarations of effect/attitude: 効きます 正面から 多角的に 掘り下げる are placed in place of observation or judgment.
  • Subjective-verb-type lead-in: ここで一気に効いたのが ここからが一番ピンと来た話 Instead of a conjunction, place a subjective verb such as 効いた ピンと来た .ここまでの話の核になるのが ここで関わってくるのが unravel into, etc.
  • Subject confusion: what you did, what you had AI do, and quotes from official information get mixed together.
  • Structural uniformity: paragraph lengths, heading endings, and bullet counts are too uniform. Adding an addendum section with **ラベル**: 説明 3 bullet points is of the same kind. If it is one flow, write it in short paragraphs.
  • Adverb-favoring cluster: within one article, if the same adverb such as 個人的に 本当に ちゃんと ものすごく appears 3 or more times, even if each is not High on its own, they accumulate into an AI smell. Make a list of locations and reduce at least 3 of them.
  • English tell: delve, leverage, robust, Let's dive into, In conclusion, Not just X but Y。

Output

For the detailed templates (Review / Rewrite / Self-scoring / Anti-patterns / Example Prompts), refer to references/output-templates.md . Just the key points:

  • Audit: return High / Medium / Low + Structure in the AI-likeness Audit format
  • Conversion: before / after or revised sentence. Do not do bulk replacement
  • Generation instruction: return a style-instruction prompt (generation instruction mode)

Done Criteria

Review type:

  • Problem locations are listed with line numbers or quotes
  • Each problem has a category and a suggested replacement
  • Duplicate signals for the same location are not padded out
  • You are not deciding based on a single tell, but checking whether it can be explained as a cluster / stacking pattern
  • The temperature (polite style / plain style / emoji) is consistent throughout the article
  • Facts, impressions, and intentions are not mixed up
  • You check whether the writer's judgment is visible, rather than generalities
  • Measured results, official sources, and AI-supplied formatting results are not mixed together
  • The recommendations and conclusions in the TL;DR, body, and summary do not conflict, and there are no assertions stronger than the original source (including paraphrases with quotation marks)

Generation/conversion type:

  • The same opening or the same sentence ending does not occur 3 times in a row. However, where having a uniform form is itself correct, such as numbered steps or listings of specifications, do not fix it unless it overlaps with other tells
  • There is no dropping of conditions or numbers from the original, and no addition of subjects or specifications not in the original
  • It is written with concrete words rather than abstract words
  • What you did, what you had AI do, and what you quoted are separated at the subject level
  • If there are examples, it leans toward that person's voice
  • After the first draft, self-audit the remaining AI-ness exactly once, and eliminate it in the final draft
  • Even after revision, just as much motivation, hesitation, and judgment as necessary remains

References