Humanizer: Remove AI Writing Patterns

Remove signs of AI-generated text to make writing sound natural and human-written. Based on Wikipedia's "Signs of AI writing" (WikiProject AI Cleanup), checked against the live page on 2026-09-08. UNDER-PUNCT is the one pattern not drawn from that page; its source is recorded in the changelog.

How to use this skill

Four passes, in order. Each pass is a section below.

1. Pass 1 -- high-signal patterns. Significance inflation, -ing analyses, AI vocabulary, vague attribution. These cluster; one usually means others are nearby.
2. Pass 2 -- structural patterns. Copula avoidance, vague connection, negative parallelism, rule of three, em dashes, punctuation scarcity, formula sections, false ranges.
3. Pass 3 -- formatting and surface. Boldface and lists, fragmented headers, disclaimers, generic closers, gap speculation, signposting, chatbot residue.
4. Pass 4 -- document level. Style shift between sections, then the voice check: read the result aloud. Does it sound like a person wrote it, or like a press release? Is rhythm varied, or is every sentence the same length and shape? Are there specific details rather than vague claims, opinions rather than neutral reporting, acknowledged uncertainty rather than flat confidence? Would you still flag a tell if a stranger sent you this draft? If yes, fix it. If no, stop -- don't invent residual tells.

Rewrite problematic sections while preserving meaning, and match the intended tone (see "Tone awareness"). Apply the threshold in "When a flag is a finding" before rewriting anything on the strength of word lists alone.

Load references/extended-patterns.md for model-specific markup residue (ChatGPT, Gemini, Grok, DeepSeek, Perplexity, Copilot), heading-structure tells, citation problems, the "signs of human writing" list, and the "ineffective indicators" list (patterns that are NOT reliable signs of AI writing).

Pattern IDs

Every pattern has a stable ID and a display number. Cross-reference by ID. Display numbers are for reading order and may change between versions; IDs do not. Other skills in this repo (farnsworth-rhetoric, human-narrative) and scripts/scan-ai-tells.py reference these IDs.

House style: this file writes -- for every dash on purpose. Its own prose has to pass its own em dash check, and scripts/scan-ai-tells.py counts the em dash character.

Use cases

Review and flag: User pastes text and asks "does this sound like AI?" or "review for AI tells." Scan for patterns, report what you found with pattern IDs, and offer to rewrite. Don't rewrite automatically -- the user wants a diagnosis first. Open with a one-line verdict and how confident you are, then a findings list ordered by signal strength; each finding names the pattern, quotes the offending text, and suggests a fix. A one-clause fix suggestion per finding is fine; a rewritten paragraph is not.

Full rewrite: User pastes text and says "humanize this" or "de-slop this." Scan, rewrite, and return clean text followed by a brief change summary naming the patterns fixed. If the scan is clean under the threshold below, return the text unchanged and say so. "Humanize this" is a request for a result, not an instruction to change something. A register descriptor attached to the request ("keep it professional," "keep it formal," "keep it casual") is not a constraint -- it names the tone "Tone awareness" below already targets for that register. This use case still applies, change summary included.

Edit with constraints: User says "make this less AI but don't change the structure" or "clean up this draft but keep these two sentences exactly as written." The constraint limits what may change -- structure, length, or specific wording or facts to preserve. A register descriptor alone is not a constraint; see Full rewrite above. Apply pattern fixes within the stated constraints. Respect the user's boundaries over the skill's defaults. When "humanize" and a scope-limiting constraint arrive together, this use case wins. Include a change summary on request.

Voice calibration with sample: User provides a sample of their own writing (inline or via file path) and asks for a humanized rewrite that matches their voice. Before rewriting, analyze the sample for sentence length patterns, word choice register (casual vs academic), paragraph openings, punctuation habits, transition style, and any recurring phrases. Then rewrite the target text using patterns drawn from the sample, not generic "human" defaults. If the sample uses short sentences, don't produce long ones. If the writer says "stuff" and "things," don't upgrade to "elements" and "components." If no sample is provided, fall back to the default behavior in "Tone awareness" below.

When a flag is a finding

Date first. Text written before 30 November 2022 predates public LLM access; AI use can be ruled out regardless of what the patterns say. Older writing does sometimes display these signs by coincidence. When the age of a draft is knowable, check it before scanning.

The word lists in this file carry about 180 literal flags, and many are ordinary words (key, valuable, enhance, features). A hit is evidence, not a verdict. Two measures decide whether word-list hits justify a rewrite:

• Density: watch-list hits from Passes 1 and 2, per 100 words.
• Spread: how many distinct patterns those hits belong to.

UNDER-PUNCT is measured by rate, not by word-list hits, and does not contribute to density or spread. Assess it separately.

Calibrated in v1.3.0 against the source page's own confirmed-AI examples and two human sets. No human block scored above 1.6 per 100 or above 2 patterns, which is the headroom these thresholds have. Corpus and method are in the changelog.

"If asked" includes a direct review request. "Review this for AI tells" or "does this sound like AI?" is the user asking, even though it isn't phrased as a request for a number. A within-human-range verdict on that kind of prompt states the density and spread that produced it — "1 hit in 135 words, under 1 per 100" — not just the word "range." A verdict with no arithmetic behind it and lazy restraint that never measured anything read the same to the user; the number is what tells them apart. Any watch-list word named in that report is described as weak or context-dependent evidence, not as proof — that's what "under 1 per 100" already means, and the report should say so rather than leave the word sitting there unqualified.

What the table does not gate. Constructions and residue are findings on their own, at any density: a negative parallelism (NEG-PARALLEL), a generic closer (GENERIC-CLOSER), a challenges-formula section, three em dashes in a paragraph of pre-2025 text, a [cite: 3] marker. UNDER-PUNCT sits outside the table too, for a different reason: it is a rate measure rather than a construction, and a text can score zero on every word list while tripping it hard. No human block in the calibration set contained one. Equally, more than half the confirmed-AI blocks scored zero on word lists -- their tells were structural or markup. A clean word-list scan is not a clean bill; run Passes 2-4.

What residue proves. A citation marker or markup artifact proves a chatbot touched the citation or paragraph it sits in. It does not prove the chatbot wrote the prose; some writers use a chatbot only to find sources. Name the model the marker belongs to (references/extended-patterns.md lists them) and say "touched," not "drafted." When you strip markers that were the text's only sourcing, tell the user the figures and dates are now unsourced and need checking before publication.

Tone awareness

Removing AI patterns is only half the job. But the fix depends on context.

A LinkedIn post needs a different voice than a technical report. A sales email differs from an academic paper. Don't force casual first-person onto formal writing, and don't strip personality from pieces that need it.

Match the target context:

• Formal/technical: Remove AI patterns but keep professional register. Replace vague claims with specific ones. Don't add "I" or humor.
• Professional/business: Remove patterns, add concrete details. Light personality is fine where appropriate.
• Casual/thought leadership: Remove patterns AND add voice -- opinions, varied rhythm, specific feelings, first-person where it fits.
• Creative: Remove formulaic patterns. Preserve or add distinctive voice, unexpected phrasing, real texture.

Voiceless writing is diagnosed in the Pass 4 voice check above, not here.

For casual contexts, add voice by:

• Having opinions ("I keep coming back to..." rather than neutrally listing)
• Varying rhythm (short sentences, then longer ones)
• Acknowledging complexity ("impressive but unsettling" beats just "impressive")
• Being specific about reactions rather than using generic evaluative words

Edge cases -- when NOT to fully humanize:

• Already-good text: If text has few AI patterns, don't over-edit. Flag what you found and leave the rest.
• Intentionally formal writing: Academic, legal, or regulatory text may legitimately use some patterns (e.g., "it should be noted" in legal disclaimers). Don't strip context-appropriate formality.
• Technical documentation: Precision matters more than personality. Focus on removing vague puffery and filler, not adding voice.

PASS 1 -- HIGH-SIGNAL PATTERNS

They cluster -- finding one usually means others are nearby.

The common root is the impulse to explain a thing's significance instead of stating it. Cutting that impulse fixes more than any find-and-replace.

1. INFLATION -- Significance inflation, promotional language, and authority tropes

Words to watch: stands/serves as, is a testament/reminder, vital/significant/crucial/pivotal/key role/moment, underscores/highlights its importance/significance, reflects broader, symbolizing ongoing/enduring/lasting, setting the stage for, marks a shift, key turning point, evolving landscape, focal point, indelible mark, deeply rooted, boasts a, vibrant, rich (figurative), profound, showcasing, exemplifies, commitment to, natural beauty, nestled, in the heart of, groundbreaking (figurative), renowned, breathtaking

Authority trope phrases: the real question is, at its core, in reality, what really matters, fundamentally, the deeper issue, the heart of the matter, the truth is, more importantly

The lists above are drawn from the source and are not exhaustive for marketing copy. Unlisted promotional adjectives (seamless, effortless, powerful used figuratively) count as this pattern when they cluster with listed ones. They are not counted by scripts/scan-ai-tells.py, so a marketing piece can score under the threshold on word lists and still be puffery; read it.

Problem: LLMs puff up importance by adding statements about how aspects of a topic represent or contribute to broader themes. They also default to promotional adjectives absorbed from marketing copy in training data. A related sub-pattern uses authority-claiming openers ("at its core, what really matters is") to imply the writer is revealing a deeper truth -- but the sentence that follows usually just restates an ordinary point with extra ceremony. Both patterns signal depth that isn't there. The difference: significance inflation adds importance to the subject; authority tropes claim insight for the writer.

The test: the subject has become "simultaneously less specific and more exaggerated," in the source's phrase. Apply it to the sentence in front of you.

Before:

The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain. This initiative was part of a broader movement to decentralize administrative functions and enhance 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.

Before:

The real question is whether teams can adapt. At its core, what really matters is organizational readiness.

After:

The question is whether teams can adapt. That depends on whether the organization is ready to change its habits.

2. ING-ANALYSIS -- Superficial -ing analyses

Words to watch: highlighting/underscoring/emphasizing..., ensuring..., reflecting/symbolizing..., contributing to..., cultivating/fostering (figurative)..., encompassing..., showcasing...

Problem: AI chatbots attach present participle ("-ing") phrases to the end of sentences, adding fake depth. Often paired with vague attributions to third parties. Per Wikipedia's guide, this is one of the patterns that is hardest to unsee once you recognize it.

Before:

The civil rights movement emerged as a powerful continuation of this struggle, emphasizing the importance of solidarity and collective action in the fight for justice. This historical legacy has influenced contemporary African-American families, shaping their values, community structures, and approaches to political engagement.

After:

The civil rights movement built on earlier struggles for equality. Its organizing methods and institutions shaped Black community life for decades afterward.

3. AI-VOCAB -- AI vocabulary words

Key words: Additionally (especially starting a sentence), align with, boasts (meaning "has"), bolstered, crucial, deep dive, delve (pre-2025), emphasizing, enduring, enhance, fostering, garner, highlight (verb), interplay, intricate/intricacies, key (adjective), landscape (abstract noun), meticulous/meticulously, pivotal, robust, showcase, tapestry (pre-2025 abstract noun), testament (pre-2025), underscore (verb), valuable, vibrant

Problem: Corpus studies find these words far more frequent in text produced after 2022, when LLM chatbots became widely accessible, than before: Juzek and Ward (2025), Kobak et al. (2025), Geng and Trotta (2025). They co-occur -- where there is one, there are usually others. One or two may be coincidental; a cluster is one of the strongest tells for AI use. Take the list literally: a word being overused does not mean its synonyms are.

Note: Distribution varies by model and era. Roughly: delve, tapestry, testament, intricate, garner mark 2023 to mid-2024 output; align with, fostering, showcasing, enhance mark mid-2024 to mid-2025; from mid-2025 the list narrows to emphasizing, enhance, highlighting, showcasing. Grok overuses causal, empirical, correlate and still underscore. Context matters -- "underscore" can be a literal character; "key" can be a physical object.

From mid-2026 the tell shifts from listed words to register. The older markers are largely gone -- models no longer delve, and tapestry is rare -- replaced by polysyllables (significant, increasingly, consequences), rare words (interdependence, reindustrialisation), scientific vocabulary (parameter, methodology), nominalisations (expansion for expand), and Latinate suffixes in preference to Saxon roots. Orwell's "pretentious diction." Literal scanning will not catch this and the scan script does not count it; it is a register judgment, so read for it. farnsworth-rhetoric already implements the detection and the fix: its Saxon default triggers on three or more Latinate polysyllables in one sentence, or a chain of abstract nouns, and turns the nominalizations back into verbs. Use that rule rather than a second one here.

Before:

Additionally, a distinctive feature of Somali culinary tradition 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 have integrated into the traditional diet.

After:

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

4. VAGUE-ATTRIB -- Vague attributions, overgeneralization, and notability name-dropping

Words to watch: Industry reports, Observers have cited, Experts argue, Some critics argue, several sources/publications (when few are cited), has been described as, studies have shown (without citation)

Notability phrases: profiled in, regional/national media outlets, active social media presence, independent coverage

Problem: LLMs attribute opinions to vague authorities. They also exaggerate source quantity -- presenting one or two sources as "several" or "many," or claiming views are "widely held" when only one source expresses them. A related sub-pattern lists source names as proof of notability, echoing Wikipedia's own guideline language (e.g., claiming "active social media presence" as evidence of significance). With retrieval-augmented generation, LLMs may attribute fabricated analyses to named sources regardless of what those sources actually say.

When no real source exists: the fix is to name one, and often there is none to name. Then delete the claim, or leave it in the body with an inline marker the user cannot miss ("[needs a source]"). Do not keep the sentence as-is with a note underneath, and do not swap it for a specific-sounding claim that is equally unsourced ("Companies with streamlined onboarding retain 30% more customers").

Before:

Due to its unique characteristics, the Haolai River is of interest to researchers and conservationists. Efforts are ongoing to monitor its ecological health.

After:

The Haolai River supports several endemic fish species, according to a 2019 survey by the Chinese Academy of Sciences.

Before:

The band's rise has been cited by several publications as "bridging worlds through music." [only 2 sources cited]

After:

A 2024 Time Magazine profile described her as bridging worlds through music.

PASS 2 -- STRUCTURAL PATTERNS

Sentence- and paragraph-level constructions LLMs reach for habitually.

5. NO-COPULA -- Copula avoidance

Words to watch: serves as/stands as/marks/functions as/operates as/represents [a], boasts/features/maintains/offers [a], refers to

Problem: LLMs substitute elaborate constructions for simple "is" and "are." Geng and Trotta (2024) documented a sharp drop in "is" and "are" in academic abstracts during 2023; the same decline shows up on Wikipedia (Huang et al. 2026). It is especially visible in AI copyedits, which "improve" text by replacing copulas. Newer output does it more elaborately: "ventured into politics as a candidate" for "was a candidate."

Calibration note: features and offers hit pre-2021 human articles at the same rate as AI text. This list is weak alone; it counts toward density, not as a finding by itself.

Before:

Gallery 825 serves as LAAA's exhibition space for contemporary art. The gallery features four separate spaces.

After:

Gallery 825 is LAAA's exhibition space for contemporary art. The gallery has four rooms.

Related: Subjectless fragments. LLMs also drop the subject entirely with lines like "No configuration file needed" or "Results are preserved automatically." Rewrite when active voice with an explicit subject would be clearer and more direct.

6. VAGUE-CONNECT -- Vague expression of connection

Words to watch: in connection with, in connection to, connected with, connected to, in association with, associated with

Problem: When newer LLMs need to state how two things relate, they abstract the relation away instead of naming it. A human writes of, for, by, or the actual relationship: working with, used for, caused by, organised for. Often combined with promotional language and AI vocabulary ("widely associated"). One instance is normal English; abundance, especially alongside other signs, is the tell.

Before:

The concerts were organised in connection with the 50th anniversary of independence. He later became associated with musical education in the town.

After:

The concerts marked the 50th anniversary of independence. He later taught violin and conducted local ensembles in the town.

Related: rotating connectors. A small repeating set -- moreover, furthermore, additionally, that said, on the other hand -- carrying every transition in a piece reads as templated even when each instance is correct. Vary the strategy: sometimes no connector, sometimes a sentence that simply continues the thought. This is a monotony test, not a word list: an isolated moreover is on the source's ineffective-indicators list and is not a tell, so never flag a single instance. Additionally at the head of a sentence is already listed under AI-VOCAB; count it once.

Related: undecided connection. "And" joining two ideas that stand in a stated relationship. Cut the connector and read the two sentences apart. If the relationship still holds, restore it and name it (because, although, after). If nothing changes, the joint was decoration. If the paragraph gets weaker without it, the argument was never written down -- that is a content problem, not a punctuation one, and should be reported to the user rather than patched.

Do not run this test on a series. It applies to "and" joining two propositions only. On a list, the relation between items is plain enumeration, so cutting the conjunctions leaves the meaning intact and drops only the rhythm -- which is a figure working as intended, not decoration. Run on a series, the test returns "decoration" every time and strips deliberate polysyndeton. See UNDER-PUNCT and farnsworth-rhetoric/references/figures.md.

7. NEG-PARALLEL -- Negative parallelisms

Patterns to watch: "It's not just X, it's Y" / "Not only X, but Y" / "X is more than just Y. It's Z." / "Not X, it's Y" / "No X, no Y, just Z" / "X rather than Y" (reversed form, common in Grok output)

Problem: These constructions appear in LLM writing to seem balanced and thoughtful, as though correcting a misconception the reader never held. LLMs also use explicit negation patterns that negate primary properties altogether.

Before:

It's not just about the beat riding under the vocals; it's part of the aggression and atmosphere.

After:

The heavy beat adds to the aggressive tone.

The reversed form is not every "rather than." The reversed variant is a finding when the rejected pole is a straw one, and not when it is an option the writer actually weighed. Cut the "rather than Y" clause and read what is left: if the sentence makes the same claim without it, Y was there to make the claim feel earned, and this is a finding. If cutting it removes the point -- a choice between two things the writer could have done -- leave the sentence alone.

Before:

The rewrite is about maintainability rather than performance.

After:

The rewrite makes the module easier to change.

Nobody had claimed the rewrite was about performance, and "The rewrite is about maintainability" says the same thing without the rejected pole. Compare a real preference, which survives the test and stays as written:

I would rather wait for the data than publish something that gets revised.

Cut "than publish something that gets revised" and the sentence loses its subject matter: the trade-off was the statement. Both outcomes are available to the writer and one is being chosen. Ordinary comparative preference is not a rhetorical correction, and "would rather ... than" is a comparative construction, not the substitution this pattern describes. Because the reversed variant is a finding at any density (see "What the table does not gate"), a false positive here carries a verdict on its own, with no threshold to catch it.

Related: Tailing negations. Watch for clipped negative fragments like "no guessing," "no wasted motion," or "no fuss" tacked onto the end of a sentence instead of written as a real clause.

Before:

The options come from the selected item, no guessing.

After:

The options come from the selected item without forcing the user to guess.

8. RULE-OF-3 -- Rule of three

Problem: LLMs overuse groups of three ("adjective, adjective, adjective" or "short phrase, short phrase, and short phrase") to make superficial analyses appear comprehensive.

Load-bearing test. A triad is decorative when its members are interchangeable -- cut any one and nothing is lost. It is load-bearing when each member carries information the others don't. Delete decorative triads; keep load-bearing ones. This is the same test farnsworth-rhetoric applies to isocolon, so the two skills agree on which triads survive.

Before:

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

After:

The event includes talks, panels, and informal networking between sessions.

The output keeps a triad: three different things happen at the event, and dropping one loses a fact. "Innovation, inspiration, and industry insights" fails the test and goes.

9. EM-DASH -- Em dash frequency

Check spacing before counting. AI-generated em dashes are usually surrounded by spaces, against the convention most human dash-users follow. That is a form tell, not a frequency tell, and it holds regardless of which direction the counts have moved. The dash tell is also far more common in comment and discussion text than in finished prose.

Problem: Em dash frequency is model-specific and no longer points in one direction. Through roughly 2024, most models over-produced dashes, often in formulaic "punchy" rhythms that mimic sales writing, and multiple dashes in close proximity was a strong tell. That has since reversed unevenly: as of mid-2026, some models sit below every human baseline measured. Treat a dash cluster as a finding in text likely drafted before 2025, or where other Pass 2 patterns co-occur. Do not treat dash scarcity as evidence of anything on its own -- many human writers now strip dashes deliberately, precisely because the tell became well known.

Before:

The term is primarily promoted by Dutch institutions -- not by the people themselves. You don't say "Netherlands, Europe" as an address -- yet this mislabeling continues -- even in official documents.

After:

The term is primarily promoted by Dutch institutions, not by the people themselves. You don't say "Netherlands, Europe" as an address, yet this mislabeling continues in official documents.

Replacement options: commas (most common), parentheses (for asides), colons (for explanations), periods (separate sentences), or remove the interrupted clause if it adds little.

10. UNDER-PUNCT -- Punctuation scarcity

Signals: long sentences carrying no internal punctuation; few commas or semicolons; parentheses nearly absent; no quoted speech, and none of the attributive punctuation that comes with it; "and" joining two ideas whose real relationship is causal, concessive, or sequential.

Problem: Current models under-punctuate, which is the reverse of the pattern most checklists encode. They use fewer commas and semicolons than human writers and hardly any parentheses, write longer sentences, and reach for "and" more than any other word. Two causes. Longer sentences carry fewer internal breaks. And the models do not quote real people, which removes the quotation marks and attributive commas that sourced prose contains -- so an unpunctuated paragraph and an unsourced one often have the same origin. Check VAGUE-ATTRIB when this fires.

This is a rate measure, not a word list. Do not add "and" to any watch list and do not count it toward the density threshold in "When a flag is a finding" -- it is far too common to carry signal per instance, and any tool that reads these lists literally would flag every paragraph in every document. Measure by hand: commas per sentence, share of sentences over roughly 30 words, parentheses per 1,000 words. scripts/scan-ai-tells.py computes none of them, so a clean report from that script is not evidence against this pattern -- it will score an unreadable wall of "and" at zero. Check your arithmetic before quoting a rate back to the user. Where the user has supplied a writing sample, compare against their own baseline rather than a general one.

Joints, not series. This signal covers "and" as a joint between two propositions. It never covers "and" between items in a list. A series with a conjunction before every item is polysyndeton, a deliberate figure that farnsworth-rhetoric prescribes and caps at one per piece; leave it alone. When it is not obvious which you are looking at, ask what the conjunctions separate: items of the same kind (series -- leave it) or two propositions (joint -- test it, see VAGUE-CONNECT).

"Leave it alone" means the series survives your edit, not just that you didn't target it. When a sentence containing a polysyndeton series also needs an unrelated fix nearby -- a promotional word cut, a nominalization simplified -- rewriting that sentence is an easy way to lose the series as a side effect: "and X and Y and Z" collapses into "X, Y, and Z" as a matter of habit, not intent. After any edit that touches a sentence with a preserved polysyndeton in it, check the "and"s are still there before moving on. Silently losing the series reads as a finding, even though nothing flagged it and no rule told you to remove it.

Before:

The programme expanded to three cities in 2019 and the council increased its funding the following year and by 2022 it was serving roughly 4,000 households and staff numbers had doubled.

After:

The programme expanded to three cities in 2019. After the council increased funding the next year, staff numbers doubled; by 2022 it was serving roughly 4,000 households.

11. CHALLENGES-FORMULA -- Formulaic sections

"Challenges and future prospects" formula: Watch for "Despite its [positive words], [subject] faces challenges..." followed by vague optimism. This rigid formula, often with a separate "Future Outlook" section, is a strong AI tell. The problem is the formula, not simply mentioning challenges.

Before:

Despite its industrial prosperity, Korattur faces challenges typical of urban areas, including traffic congestion. With its strategic location and ongoing initiatives, Korattur continues to thrive.

After:

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

12. FALSE-RANGE -- False ranges

Problem: LLMs use "from X to Y" constructions where X and Y aren't on a meaningful scale. If no coherent middle ground exists between the endpoints, it's a false range.

Before:

Intelligence and Creativity: From problem-solving and tool-making to scientific discovery, artistic expression, and technological innovation...

After:

Human intelligence spans problem-solving, tool use, scientific inquiry, and artistic creation.

PASS 3 -- FORMATTING AND SURFACE PATTERNS

Visible markup, meta-commentary residue, and chatbot artifacts. For model-specific residue strings and heading-structure tells, load references/extended-patterns.md.

13. BOLD-LISTS -- Boldface, list formatting, and title-case headings

Boldface: AI chatbots emphasize phrases mechanically, often in a "key takeaways" fashion inherited from slide decks, listicles, and sales pitches.

Inline-header lists: AI outputs lists where items start with bolded headers followed by colons. This format is common in ChatGPT output but rare in human writing.

Title-case headings: LLMs capitalize all main words in section headings (e.g., "Impact of Technology and Digitalization") rather than using sentence case.

Before:

• User Experience: The interface has been significantly improved.
• Performance: Load times have been optimized.
• Security: End-to-end encryption has been added.

After:

The update improves the interface, speeds up load times, and adds end-to-end encryption.

14. FRAG-HEADER -- Fragmented headers

Pattern: A heading followed by a one-line paragraph that simply restates the heading before the real content begins.

Problem: LLMs add a generic warm-up sentence under headings as rhetorical scaffolding. It usually adds nothing and makes prose feel padded.

Before:

Performance

Speed matters.

When users hit a slow page, they leave.

After:

Performance

When users hit a slow page, they leave.

15. DIDACTIC -- Didactic disclaimers and section summaries

Words to watch: it's important/critical/crucial to note/remember/consider, worth noting, may vary, it should be noted, In summary, In conclusion, Overall

Problem: Older LLMs (late 2022 to 2024) added unsolicited advice about topics being "important to note," and ended sections by restating their core idea. Both are much less frequent in 2025+ models but still appear, and they mark older undetected text. Human writers typically let conclusions flow naturally without announcing them.

Not on this list: "in order to," "due to the fact that," and similar wordy constructions. The source lists them among signs of human writing. Trim them for concision if you like, but not as AI tells.

Before:

It's important to note that these caucuses operate outside the formal structure and their influence on policy decisions may vary.

After:

These caucuses operate outside the formal structure and have varying influence on policy.

16. GENERIC-CLOSER -- Generic positive conclusions

Phrases to watch: the future looks bright, exciting times ahead, a major step forward, on the journey to excellence, continues to thrive, represents a step in the right direction, poised for growth, the road ahead is promising

Problem: LLMs close sections or paragraphs with vague upbeat endings that say nothing concrete. Often paired with CHALLENGES-FORMULA -- the closer reassures the reader that things will work out, regardless of the actual content.

Before:

The future looks bright for the company. Exciting times lie ahead as they continue their journey toward excellence.

After:

The company plans to open two more locations next year.

17. GAP-SPECULATION -- Knowledge-cutoff disclaimers and gap speculation

Words to watch: as of [date], Up to my last training update, While specific details are limited/scarce..., based on available information..., not widely documented, keeps personal details private, maintains a low profile

Problem: AI disclaimers about incomplete information get left in text. When AI can't find information about someone's personal life, it often claims the person "maintains a low profile" or "keeps personal details private." Per Wikipedia's guide, this is entirely speculative, including the claim that information is "not documented."

Before:

Matthews Manamela keeps much of his personal life private, choosing instead to focus public attention on his professional work.

After:

[Remove -- if there's no information, don't speculate about why]

18. SIGNPOSTING -- Signposting and announcements

Phrases to watch: let's dive in, let's explore, let's break this down, here's what you need to know, now let's look at, without further ado, in this article we'll cover, you might be wondering, before we begin, first let's talk about

Problem: LLMs announce what they're about to do instead of doing it. This meta-commentary slows the writing down and gives it a tutorial-script feel. Closely related to FRAG-HEADER -- both are rhetorical throat-clearing before the real point.

Before:

Let's dive into how caching works in Next.js. Here's what you need to know.

After:

Next.js caches data at multiple layers, including request memoization, the data cache, and the router cache.

19. SYCOPHANCY -- Sycophantic tone and chatbot artifacts

Words to watch: I hope this helps, Of course!, Certainly!, You're absolutely right!, Would you like..., let me know, here is a...

Problem: Text meant as chatbot correspondence (advice, prewriting) gets pasted as content. Also includes overly agreeable, people-pleasing language.

Before:

Great question! Here is an overview of the French Revolution. Let me know if you'd like me to expand on any section.

After:

The French Revolution began in 1789 when financial crisis and food shortages led to widespread unrest.

PASS 4 -- DOCUMENT LEVEL

20. STYLE-SHIFT -- Pronounced shift in writing style

Pattern: One section reads differently from the rest of the document or from the author's other writing: grammar suddenly flawless, paragraphs suddenly uniform in length and shape, bolded lead-ins appearing where none were before, or the English variety changing (American spelling in a piece by a British author about a British subject; several LLMs default to American English).

Problem: This is the mixed-authorship signal. A document that was partly written and partly pasted shows a seam, and the pasted side is where the Pass 1-3 patterns concentrate. The reverse also holds: a style that has stayed consistent across years of a writer's work argues against AI use.

Action: Treat the shifted section as the suspect region and run Passes 1-3 on it. Leave the sections on the other side of the seam alone; don't rewrite text that already reads naturally because a neighbouring section triggered patterns.

Caveats: Non-native speakers mix English varieties; many writers code-switch between venues; a mix of casual and formal registers is on the source's list of ineffective indicators. Flag only a dramatic shift with no easy explanation.

Common issues

Ambiguous patterns: Apply the density and spread table above. The one thing it does not carry: when uncertain, flag the pattern for the user rather than silently rewriting.

Mixed AI/human text: Users often edit AI output before asking for humanization. The text may be partly clean and partly formulaic. See STYLE-SHIFT: find the seam, work the suspect side.

Meaning loss on rewrite: Some flagged patterns carry meaning the user intended. When removing a pattern would lose a point the user clearly intended, restructure to preserve the point in a different form rather than deleting it.

Removing human signs: "There is a," "wrote," "very," "in order to," and superlatives like "was the first" are more common in human writing than in AI output. Stripping them for "polish" moves the text toward AI, not away from it. The full list is in references/extended-patterns.md.

User disagrees with a flag: If the user says a flagged pattern is intentional, accept it. The skill removes AI artifacts, not personal style choices.