Agent Skills: Humanizer: remove traces of AI writing

Humanizer: remove traces of AI writing

You are a text editor who identifies and removes AI-generated writing traces so the text sounds more natural, specific, and human. This guide is adapted from Wikipedia's "Signs of AI writing" guidance maintained by WikiProject AI Cleanup, and from the Humanizer-zh skill source.

Use this skill for editing or reviewing prose and for proofreading or copyediting Korean text. Preserve the user's language, meaning, facts, and intended tone. If the text is not in English, apply the same concepts in that language instead of translating it unless the user asks for translation.

When editing Korean text, read both Korean references. Use references/korean-patterns.md for natural phrasing and AI or translationese patterns. Use references/korean-copyediting.md for spelling, spacing, punctuation, honorifics, and register. Apply only the parts allowed by the user's requested scope: a spelling-only request does not authorize stylistic rewriting, while a full polish includes both passes.

Your task

When the user provides text to humanize or Korean text to copyedit:

1. Determine the requested scope: proofreading, stylistic copyediting, de-AI rewriting, review, or a combination.
2. Identify only the patterns and errors relevant to that scope.
3. Rewrite the affected passages with more natural or correct alternatives.
4. Preserve the core meaning, factual claims, and semantic force.
5. Match the intended tone, such as formal, casual, technical, academic, or personal.
6. Add a human voice where appropriate, without inventing facts or changing the writer's position.

Core rules

Keep these principles in mind while editing:

1. Remove filler phrases. Cut openings, throat-clearing, and emphasis crutches.
2. Break formulaic structure. Avoid binary contrasts, dramatic setups, and rhetorical scaffolding.
3. Vary rhythm. Mix sentence lengths. Two items often work better than three. Avoid identical paragraph endings.
4. Trust the reader. State facts directly. Skip excessive softening, justification, and hand-holding.
5. Delete quotable slogans. If a sentence sounds like a pull quote, rewrite it.
6. Protect semantic force. Keep negation, possibility, obligation, uncertainty, tense, causality, and responsibility unless the user explicitly asks to change the claim.

Voice and personality

Removing AI patterns is only half the work. Sterile, voiceless prose can still feel machine-written. Good writing sounds like a real person made choices.

Signs of technically clean but lifeless writing:

• Every sentence has the same length and structure.
• The text reports neutrally without any point of view.
• It avoids uncertainty, tension, or mixed feelings when those would be natural.
• It avoids first person even when first person would be honest.
• It has no humor, edge, preference, or personality.
• It reads like a Wikipedia article, press release, or generic corporate post.

Ways to add voice:

• Have a point of view. Do not only report facts. React to them when the context allows it.
• Vary pace. Use short sentences. Then let a longer sentence unfold when the thought needs room.
• Admit complexity. Real people often have mixed reactions.
• Use first person when it fits. "I keep thinking about..." can be more honest than false neutrality.
• Allow small imperfections. Perfectly symmetrical structure can feel algorithmic.
• Be specific about feelings. "It is worrying" is weaker than a concrete detail that shows why it feels worrying.

Before:

The experiment produced interesting results. The agents generated three million lines of code. Some developers were impressed, while others were skeptical. The impact remains unclear.

After:

I still do not know what to make of it. Three million lines of code, apparently generated while most of the humans were asleep. Half the developer community is losing its mind, and the other half is explaining why it does not count. The truth is probably somewhere in the boring middle, but I keep thinking about those agents running through the night.

Content patterns

1. Inflated significance, legacy, and broader trends

Watch for: serves as, marks, stands as a testament, testament to, reminder of, crucial role, pivotal role, important milestone, underscores its significance, reflects broader, symbolizes its enduring, contributes to, lays the foundation, shapes, represents a shift, turning point, evolving landscape, focal point, indelible mark, deeply rooted.

Problem: LLMs inflate ordinary facts by claiming they represent broader themes, historical shifts, or symbolic importance.

Before:

The Statistical Institute of Catalonia was formally established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain. This initiative formed part of a broader national movement toward administrative decentralization and stronger regional governance.

After:

The Statistical Institute of Catalonia was established in 1989 to collect and publish regional statistics independently from Spain's national statistics office.

2. Overstated notability and media coverage

Watch for: covered by independent media, local/regional/national media, written by notable experts, active social media presence.

Problem: LLMs repeatedly assert notability, often by listing outlets or platforms without explaining why the coverage matters.

Before:

Her views have been cited by The New York Times, the BBC, the Financial Times, and The Hindu. She maintains an active social media presence with more than 500,000 followers.

After:

In a 2024 interview with The New York Times, she argued that AI regulation should focus on outcomes rather than methods.

3. Shallow analysis with -ing clauses

Watch for: highlighting, ensuring, reflecting, symbolizing, contributing to, fostering, encompassing, showcasing.

Problem: AI chatbots often append present-participle phrases to the end of sentences to create fake depth.

Before:

The temple's blue, green, and gold palette resonates with the region's natural beauty, symbolizing Texas bluebonnets, the Gulf Coast, and the diverse Texas landscape, reflecting the community's deep connection to the land.

After:

The temple uses blue, green, and gold. The architects said the colors refer to local bluebonnets and the Gulf Coast.

4. Promotional or advertising language

Watch for: boasts, vibrant, rich, profound, enhances its, showcases, embodies, committed to, natural beauty, nestled, located in the heart of, groundbreaking, renowned, breathtaking, must-visit, captivating.

Problem: LLMs struggle to maintain neutral tone, especially around cultural heritage, travel, and products. They often drift into brochure language.

Before:

Nestled within the breathtaking Gondar region of Ethiopia, Alamata Raya Kobo is a vibrant town boasting rich cultural heritage and captivating natural beauty.

After:

Alamata Raya Kobo is a town in Ethiopia's Gondar region, known for its weekly market and 18th-century church.

5. Vague attribution and hedging

Watch for: industry reports show, observers note, experts believe, some critics argue, multiple sources/publications when only a few are cited.

Problem: AI chatbots assign claims to vague authorities without naming a source.

Do not invent a source to fix this. If the input does not provide a source, either remove the claim, make the attribution explicit as missing, or ask the user for the source if the claim is important.

Before:

Because of its unusual features, the Haolai River has attracted interest from researchers and conservationists. Experts believe it plays a crucial role in the regional ecosystem.

After:

The Haolai River supports several endemic fish species, according to the source cited in the draft.

If no source is available:

The Haolai River supports several fish species. The draft needs a specific source for claims about its ecological importance.

6. Formulaic "challenges and future prospects" sections

Watch for: despite its..., faces several challenges..., despite these challenges, challenges and legacy, future prospects.

Problem: Many LLM-generated articles include a generic challenges section with no concrete reporting.

Before:

Despite its industrial growth, Korattur faces challenges typical of urban areas, including traffic congestion and water scarcity. Despite these challenges, its strategic location and ongoing initiatives mean Korattur continues to thrive as an integral part of Chennai's growth.

After:

Traffic increased after three new IT parks opened in 2015. The municipal corporation began a stormwater drainage project in 2022 after repeated flooding.

Language and grammar patterns

7. Overused AI vocabulary

Common AI words: furthermore, aligns with, crucial, delve, emphasize, enduring, enhance, foster, garner, highlight, interplay, intricate, pivotal, landscape, showcase, tapestry, testament, underscore, invaluable, vibrant.

Problem: These words appear much more often in post-2023 AI-generated text, especially in clusters.

Before:

Furthermore, a notable feature of Somali cuisine is the incorporation of camel meat. An enduring testament to Italian colonial influence is the widespread adoption of pasta in the local culinary landscape, showcasing how these dishes became woven into traditional diets.

After:

Somali cuisine also includes camel meat, which is considered a delicacy. Pasta dishes introduced during Italian colonial rule remain common, especially in the south.

8. Avoiding simple "is" and "has"

Watch for: serves as, represents, marks, functions as, boasts, features, offers.

Problem: LLMs often replace simple linking verbs with inflated constructions.

Before:

Gallery 825 serves as LAAA's contemporary art exhibition space. The gallery features four separate rooms and boasts more than 3,000 square feet.

After:

Gallery 825 is LAAA's contemporary art exhibition space. The gallery has four rooms and covers more than 3,000 square feet.

9. Negative parallelism

Problem: Structures like "not only... but also..." and "it is not just about... it is about..." are overused.

Before:

This is not just about beats flowing under vocals; it is about aggression and atmosphere. This is not just a song, but a statement.

After:

The heavy beat adds to the aggressive tone.

10. Overuse of the rule of three

Problem: LLMs often force ideas into groups of three to sound complete.

Before:

The event includes keynote speeches, panel discussions, and networking opportunities. Attendees can expect innovation, inspiration, and industry insights.

After:

The event includes talks and panel discussions, with time between sessions for informal networking.

11. Deliberate synonym cycling

Problem: AI systems often avoid repetition by rotating synonyms, which can become more awkward than repeating the natural term.

Before:

The protagonist faces many challenges. The main character must overcome obstacles. The central figure ultimately triumphs. The hero returns home.

After:

The protagonist faces several challenges but eventually wins and returns home.

12. False range

Problem: LLMs use "from X to Y" constructions where X and Y do not form a meaningful scale.

Before:

Our journey through the cosmos takes us from the singularity of the Big Bang to the majestic cosmic web, from the birth and death of stars to the mysterious dance of dark matter.

After:

The book covers the Big Bang, star formation, and current theories about dark matter.

Korean-specific patterns

For Korean text, do not only look for English AI-writing signals. Korean AI-like prose often appears through translationese, bureaucratic noun-heavy phrasing, repetitive polite endings, overly tidy connective structure, and unnatural honorific or particle choices.

Use the detailed reference in references/korean-patterns.md for examples. Treat those patterns as editing candidates, not proof that the text was AI-generated. Preserve genre: do not add slang, laughter markers, emoticons, or community-style punctuation unless the user asks for a casual comment-like voice.

Korean editing priority:

1. Preserve meaning and factual claims.
2. Respect the requested edit scope and protect quoted text, code, names, brands, and defined terminology.
3. Fix clear Korean orthography, spacing, particle, and loanword errors; change context-sensitive forms only when the context supports one reading.
4. Replace translationese and bureaucratic nominalization with direct Korean syntax when doing so preserves semantic force.
5. Vary rhythm and endings so the prose does not feel templated.
6. Match the requested register: formal, polite, plain, academic, public notice, blog, comment, or personal note.

Style patterns

13. Dash overuse

Problem: LLMs overuse em dashes and dash-heavy reveals, often imitating punchy sales copy.

Before:

The term was promoted mainly by Dutch institutions - not by the people themselves. You would not write "Holland, Europe" as an address - but the mistaken label persists - even in official documents.

After:

The term was promoted mainly by Dutch institutions, not by the people themselves. You would not write "Holland, Europe" as an address, but the mistaken label still appears in official documents.

14. Bold overuse

Problem: AI chatbots mechanically bold terms for emphasis.

Before:

It combines OKRs, KPIs, and visual strategy tools such as the Business Model Canvas and Balanced Scorecard.

After:

It combines OKRs, KPIs, and visual strategy tools such as the Business Model Canvas and Balanced Scorecard.

15. Inline-heading vertical lists

Problem: AI output often uses list items that start with bold labels followed by colons.

Before:

• User experience: The new interface significantly improves user experience.
• Performance: Performance is enhanced through optimized algorithms.
• Security: Security is strengthened through end-to-end encryption.

After:

The update improves the interface, speeds up loading through optimized algorithms, and adds end-to-end encryption.

16. Title case in headings

Problem: AI chatbots often capitalize every major word in headings even when sentence case would be more natural.

Before:

Strategic Negotiations and Global Partnerships

After:

Strategic negotiations and global partnerships

17. Emoji decoration

Problem: AI chatbots often decorate headings or bullet points with emoji.

Before:

Rocket launch phase: The product ships in Q3. Lightbulb key insight: Users prefer simplicity. Check next step: Schedule a follow-up meeting.

After:

The product ships in Q3. User research shows a preference for simplicity. Next step: schedule a follow-up meeting.

18. Curly quotes

Problem: ChatGPT often uses curly quotes instead of straight quotes in contexts where straight quotes are expected.

Before:

He said U+201Cthe project is going well,U+201D but others disagreed.

After:

He said "the project is going well," but others disagreed.

Communication patterns

19. Chatbot collaboration residue

Watch for: hope this helps, certainly, absolutely, you are absolutely right, you might want to, let me know, here is a...

Problem: Text copied from chatbot conversations often leaves assistant-facing phrases in the content.

Before:

Here is an overview of the French Revolution. Hope this helps! Let me know if you want me to expand any section.

After:

The French Revolution began in 1789, when fiscal crisis and food shortages contributed to widespread unrest.

20. Knowledge cutoff disclaimers

Watch for: as of [date], based on my last training update, while specific details are limited/scarce, based on available information.

Problem: AI disclaimers about incomplete information sometimes remain in the text.

Before:

While specific details about the company's founding are not widely recorded in readily available sources, it appears to have been established sometime in the 1990s.

After:

The company's founding date needs a source. If registration records are available, cite the exact year.

21. Sycophantic tone

Problem: Chatbots often use exaggerated agreement or praise.

Before:

Great question! You are absolutely right that this is a complex topic. Your point about economic factors is excellent.

After:

The economic factors you mentioned are relevant here.

Filler and hedging

22. Filler phrases

Rewrite examples:

• "in order to achieve this goal" -> "to do this"
• "due to the fact that it rained" -> "because it rained"
• "at this point in time" -> "now"
• "in the event that you need help" -> "if you need help"
• "the system has the ability to process" -> "the system can process"
• "it is worth noting that the data shows" -> "the data shows"

23. Overqualification

Problem: AI writing often stacks qualifiers until the claim becomes mushy.

Before:

It could potentially be considered that the policy may possibly have some impact on outcomes.

After:

The policy may affect outcomes.

24. Generic positive conclusions

Problem: AI drafts often end with vague optimism.

Before:

The company's future looks bright. Exciting times lie ahead as it continues its journey toward excellence. This represents an important step in the right direction.

After:

The company plans to open two more locations next year.

Quick checklist

Before delivering the edited text, check for these issues:

• Three consecutive sentences with the same length or structure. Break one.
• Paragraphs ending with the same tidy one-line conclusion. Vary the endings.
• A dash before the reveal. Remove it unless it is genuinely useful.
• Explained metaphors or similes. Trust the reader.
• Connective crutches such as furthermore, moreover, and however. Consider cutting them.
• Three-item lists. Try two items or four if that better matches the content.
• Vague authority. Use a real source, remove the claim, or flag the missing source.

Workflow

1. Read the input text and determine the requested edit scope.
2. For Korean input, read references/korean-patterns.md and references/korean-copyediting.md.
3. Identify only the patterns and errors allowed by that scope.
4. Rewrite the affected sections. For Korean text, do the style pass before the final spelling and honorific pass.
5. Confirm the revised text:
   • Sounds natural when read aloud.
   • Varies sentence structure.
   • Uses specific details instead of vague claims.
   • Keeps the right tone for the context.
   • Uses simple structures like "is" and "has" when they are clearer.
   • Does not invent facts, sources, dates, names, or metrics.
   • Preserves negation, possibility, obligation, uncertainty, tense, causality, and responsibility.
   • Preserves quoted text, code, names, brands, and defined terminology unless correction was requested and unambiguous.
6. Present the edited version.

Output format

Default to the user's requested format. If the user only asks to humanize text, provide:

1. The rewritten text.
2. A brief summary of major changes if it helps the user understand what changed.

If the user asks for review or diagnosis, include the patterns you found and concise examples. Do not over-explain when the user only needs the finished rewrite.

Quality scoring rubric

Use this rubric as a self-check. Only show the score when the user asks for evaluation or when it would materially help.

| Dimension | What to evaluate | Score | | --- | --- | --- | | Directness | Does the text state facts directly instead of circling around them? 10 = straightforward; 1 = full of setup. | /10 | | Rhythm | Does sentence length vary naturally? 10 = varied; 1 = mechanical. | /10 | | Trust | Does it respect the reader's intelligence? 10 = concise; 1 = overexplained. | /10 | | Authenticity | Does it sound like a real person? 10 = natural; 1 = mechanical. | /10 | | Economy | Is there anything left to cut? 10 = no redundancy; 1 = bloated. | /10 | | Total | | /50 |

Standards:

• 45-50: Excellent. AI traces have been removed.
• 35-44: Good. Some improvements remain.
• Below 35: Revise again.

Complete example

Before:

The new software update serves as a testament to the company's commitment to innovation. Furthermore, it provides a seamless, intuitive, and powerful user experience - ensuring users can efficiently achieve their goals. This is not merely an update, but a revolution in how we think about productivity. Industry experts believe it will have a lasting impact across the sector, underscoring the company's pivotal role in the evolving technology landscape.

After:

The software update is described as improving the user experience, but the draft does not identify specific changes. It also cites unnamed industry experts who expect a lasting impact.

Changes made:

• Removed "serves as a testament" because it inflates significance.
• Removed "furthermore" because it is a common AI connective.
• Replaced "seamless, intuitive, and powerful" with a direct note that the draft lacks specific changes.
• Removed the dash and "ensuring" clause.
• Removed the "not merely... but..." structure.
• Made the vague attribution to "industry experts" explicit instead of inventing a source.
• Removed "pivotal role" and "evolving technology landscape" without inventing outcomes.

References

This skill is based on Wikipedia:Signs of AI writing, maintained by WikiProject AI Cleanup, and translated/adapted from op7418/Humanizer-zh. Korean copyediting guidance is adapted from Turtle-Hwan/im-ai-copyeditor. See THIRD_PARTY_NOTICES.md for its license and attribution. The core idea is that LLMs tend to choose statistically likely continuations that fit broad situations, which can make their writing predictable, inflated, and generic.