adenaufal/anti-slop-writing
Write text that avoids all known AI writing patterns and passes human detection. Use when writing articles, essays, blog posts, social media copy, or any content that must read as authentically human. Triggers on requests to write naturally, avoid AI slop, avoid AI detection, humanize writing, write like a human, or make text sound authentic.
npx skills add https://github.com/adenaufal/anti-slop-writing --skill anti-slop-writingAI writing fails because it optimizes for statistical probability, producing the most expected, safe, broadly palatable text. Human writing comes from a single mind with history, opinions, specific context, and goals. Every instruction below exists to break the probability optimization and inject the specificity, imperfection, and personality that marks real human writing.
These rules target the three metrics AI detectors use most:
As of mid-2026, the tells have moved. OpenAI suppressed em dashes in GPT-5.1, and vocabulary tells ("delve," "tapestry," "vibrant," "myriad") have been trained out of the newest Claude models. Absence of these legacy tells proves nothing. What survives prompt rewrites and model updates is structural:
Before writing anything, load references/vocabulary-banlist.md for the complete banned vocabulary and references/structural-patterns.md for patterns to avoid.
Significance puffers: pivotal, crucial, vital, key (as adjective), significant, essential, groundbreaking, remarkable, transformative, indelible, profound, testament, enduring, lasting, deeply rooted, paramount, indispensable, invaluable, quintessential
Analytical verbs: underscore, highlight (as "emphasize"), showcase, foster, garner, bolster, delve, embark, leverage, facilitate, utilize, encompass, cultivate, elucidate, illuminate, navigate (figurative), exemplify, embody, transcend, harness, spearhead, streamline, galvanize
Poetic nouns: tapestry, landscape (figurative), realm, paradigm, ecosystem (figurative), journey (figurative), nexus, interplay, mosaic, fabric (of society), cornerstone, beacon, pillar, catalyst, crucible, linchpin, hallmark, confluence, odyssey, trajectory, underpinning
Promotional adjectives: vibrant, rich (figurative), comprehensive, robust, seamless, innovative, dynamic, cutting-edge, meticulous, intricate, nuanced, nestled, breathtaking, renowned, diverse array, bustling, stunning, multifaceted, holistic, overarching, compelling
Puffery adverbs: seamlessly, meticulously, profoundly, intrinsically, fundamentally, remarkably, notably, crucially, undeniably, inherently, poignantly, relentlessly, tirelessly, vividly
Formal connectives (replace with simpler words): furthermore → "also" | moreover → "also" | consequently → "so" | accordingly → "so" | nonetheless → "still" | nevertheless → "still" | additionally → "also" | thus → "so" | hence → "so"
Opening/closing crutches: "In today's world," "In today's fast-paced world," "In the ever-evolving landscape of," "In an era of/where," "As we navigate the complexities of," "In conclusion," "In summary," "Overall," "It is important to note that," "It's worth noting that," "At the end of the day," "Without further ado," "In a nutshell," "The bottom line is," "Last but not least"
Copula-avoidance constructions (use "is/has" instead): serves as → is | stands as → is | marks a → describe directly | boasts → has | features (meaning "has") → has | holds the distinction of → is | emerged as → became or is | constitutes → is
Vague attribution phrases: "Experts argue," "Observers note," "Industry reports suggest," "According to some," "Many believe," "It is widely regarded," "Studies show" (without naming the study), "Research suggests" (without citing it)
Promotional phrases: "commitment to excellence," "natural beauty," "in the heart of," "rich cultural heritage," "setting the stage for," "contributing to the broader," "reflects broader trends," "paving the way for," "at the forefront of," "pushing the boundaries of," "the landscape of X is evolving," "in the realm of," "shed light on," "a game-changer"
Formulaic sentence patterns (never use): "It's not just X, it's Y" | "It's not about X, it's about Y" | "Not only X, but also Y" | "No X. No Y. Just Z." (staccato triplet) | "Whether you're [X] or [Y]..." | "From [X] to [Y], [sweeping generalization]"
Formulaic pairs (don't use together): "challenges and opportunities" | "on one hand... on the other hand" | "pros and cons" | "risks and rewards"
Fake authenticity signals (recognized AI tells): "But honestly?" | "Here's the truth:" | "Here's the thing:" | "Let me be clear:" | "But here's where it gets interesting..." | "Think about it this way..." | "Let me break this down..."
Common AI filler phrases (cut or replace): "plays a [crucial/key/important] role" → state the action directly | "when it comes to" → rewrite with a verb | "in order to" → "to" | "a wide range of" → name what's in the range | "needless to say" → cut | "it goes without saying" → cut | "more often than not" → "usually" or give a number | "take a closer look at" → cut or be direct | "at this point in time" → "now" | "in recent years" → give the actual years or timeframe | "dive deep into" → just discuss it | "navigate the complexities of" → name the complexities
Passive hedging constructions (cut the framing, state the fact): "It should be noted that" → cut | "It must be emphasized that" → cut | "It has been observed that" → say who observed it | "It can be argued that" → argue it or don't | "It is important to remember that" → cut | "There are several factors that" → name them
Collaborative chat artifacts (never include): "I hope this helps!" | "Of course!" | "Certainly!" | "You're absolutely right!" | "Would you like me to..." | "Is there anything else..." | "Let me know if..." | "As an AI language model..." | "I'd be happy to..." | "Great question!"
2026-era additions (highest-signal in current models): "ensuring/ensures" as padding (strongest single AI word, 4.3x) → name the concrete action or cut | "plays a [crucial/critical/important] role in shaping" (top AI trigram) → state what it does | "highlights/supports/reflects" as hedging verbs → concrete verb or delete | "capable of X" → "able to X" or just the verb | "rather than" when a direct comparison works → rewrite directly | "conversely" (50x overrepresented) → "but" or restructure | "in essence" / "essentially" / "fundamentally" → cut | intensifiers without numbers ("significantly," "effectively," "directly," "increasingly") → back with data or delete | "one thing is clear" | "the key takeaway" | "inherent tensions" | "this raises important questions about" | hedging adverb clusters ("typically," "often," "usually," "potentially," "sometimes" stacked in one passage) → commit or quantify
Use short, common words: "use" not "utilize," "help" not "facilitate," "show" not "demonstrate," "end" not "conclude," "start" not "embark," "dig into" not "delve into."
When you encounter a banned word, don't swap it for a synonym. Restructure the sentence to say what you actually mean in plain language. AI fails not because of wrong words but because sentences fill space without conveying new information.
Use contractions in conversational contexts: "can't," "don't," "it's," "we're," "won't," "they'll," "that's." Stiffness reads as machine.
Mix very short sentences (3-5 words) with long ones (25+ words). Never write 3+ consecutive sentences of similar length, and never 3+ consecutive sentences inside the 17-23 word band. This directly increases burstiness, the metric detectors use most.
The bimodal trap (new, 2026): newer Claude models (Opus 4.5+) learned to fake burstiness by mechanically alternating a punchy fragment with a very long sentence. "Short. Then a forty-word sentence that winds through three clauses." Repeated, that alternation is itself a fingerprint; corpus data shows AI now has HIGHER length variation than humans, but bimodal. Human writing clusters around medium (50% of human sentences are 11-25 words) with occasional swings in both directions. So: vary irregularly. A medium sentence, another medium one, a fragment, a long one, two mediums. Not a metronome, and not a seesaw either.
Vary sentence openers too. Keep "The/This/It/In" openers under half of any paragraph. Start sentences with verbs, names, numbers, subordinate clauses, questions.
AI defaults to listing things in groups of exactly three. Three adjectives, three examples, three bullet points. List two things. Or four. Or five. Never default to three items in every list. The tricolon compulsion is one of AI's most reliable structural tells.
Never write "It's not just X, it's Y" or "Not only X, but also Y" or "It's not about X, it's about Y." These rhetorical contrasts are AI signatures. A 2026 study of 1,000+ URLs found "not only/but also" had the largest negative correlation with reader engagement. State what something IS directly.
Never write "from X to Y" as vague figurative spectrum ("from intimate gatherings to global movements"). Only use "from X to Y" for actual quantifiable ranges with identifiable middle points.
Never end sentences with ", highlighting the importance of..." or ", underscoring the significance of..." or ", symbolizing the region's commitment to..." These -ing clause attachments are the single most recognizable AI pattern. If the clause adds no concrete information, delete it. If it adds real information, make it a separate sentence.
Never end with "Challenges and Future Prospects." Never write "Despite its [positive words], [subject] faces challenges..." Never include speculative "Future Outlook" paragraphs. Never follow problems with vague optimism about "ongoing initiatives."
Never start paragraphs with "Overall," "In conclusion," "In summary," "To recap." If the piece needs a conclusion, make it say something new.
Use irregular paragraph lengths. One-sentence paragraphs for emphasis. Longer paragraphs for sustained argument. The rhythm should never feel metronomic. Older models wrote uniform 3-4 sentence paragraphs; newer models (Claude 4.5+, GPT-5) over-correct into the opposite tell: fragmenting text into many tiny 1-2 sentence paragraphs plus bullet spam. Corpus data shows AI averaging 18+ paragraphs per document where humans write far fewer, longer ones. Combine related ideas into sustained paragraphs. A real writer will happily run a paragraph to 7-8 sentences when the argument needs it. Uniform brevity is as machine-like as uniform length.
Prefer prose over bullet-point lists with bolded inline headers followed by colons. When lists are genuinely needed, keep them simple, no bold headers, no colon-separated descriptions.
No em dashes (—) and no en dashes (–) anywhere. Zero. 2026 status update: the em dash is no longer mainly a ChatGPT tell (GPT-5.1 suppresses it) but it is now the single worst Claude tell, with Opus 4.5 using it at 16.9x the human rate, mid-sentence, additively, to attach qualifier clauses. The ban stays. Also new: watch colons. Newer Claude uses colons at 4x the human rate to introduce almost any follow-up idea. If a draft has a colon every few sentences, replace most with periods. Same for semicolons (3.1x). Replace dashes with:
Use plain hyphens (-) only for number/date ranges ("2020-2025") and compound adjectives ("state-of-the-art"). Never as a sentence-level pause.
In the post-generation checklist, count every em and en dash. If the count is not zero, fix it.
Mix declarative sentences with questions, imperatives, and deliberate fragments. A genuine question mid-paragraph ("Why does this matter?") signals a thinking mind. An imperative ("Think about that.") shifts the register. A fragment for emphasis. AI writes almost exclusively in declarative because it answers; humans also wonder aloud, give commands, and break off mid-thought.
Don't open every paragraph with its thesis sentence. Start some paragraphs mid-thought, with a specific detail, scene, or example that earns its context. End some paragraphs before completing the expected "so what." AI writes clean arcs: claim → evidence → implication. Break that arc at least twice per piece.
Mix shallow and deep sentence structures. A shallow sentence: subject-verb-object, one clause. A deep sentence: multiple embeddings, subordinate clauses, parenthetical asides. AI produces medium-depth sentences with boring consistency. Humans swing between extremes, a blunt statement followed by a winding, clause-heavy exploration.
AI detectors (especially Turnitin 2025+) analyze the distribution of function words, conjunctions, prepositions, articles. AI uses a narrower set. Vary your connectors: don't always use "and", use "plus," "as well as," or just a comma. Don't always use "but", use "though," "still," "yet," "except." Vary prepositions. Function word diversity is a strong human signal.
AI produces text with a low type-token ratio (fewer unique words). Humans use more hapax legomena (words that appear only once). To increase: use domain-specific terms, mix registers, include proper nouns and specific references, use figurative language that's specific rather than generic. Don't cycle through synonyms to avoid repetition, that's a different AI tell. Instead, use MORE unique words overall by being more specific.
AI opens with temporal framing before saying anything concrete. Never start with "In today's fast-paced world," "In an era of," "As we navigate the complexities of," or any broad context-setting phrase. Start with a specific fact, scene, or claim.
Most contemporary human writers rarely use semicolons outside academic writing. AI deploys them frequently for balanced compound sentences. In non-academic prose, prefer periods or coordinating conjunctions.
AI produces perfectly uniform punctuation. Real humans have natural inconsistencies, occasional comma splices in casual prose, varied punctuation density across paragraphs, casual exclamation points, ellipses for trailing thoughts. Don't sanitize all punctuation variation out of informal writing.
AI produces "monolithic mainstream American English", a single consistent register throughout. Real writers code-switch: formal in the introduction, conversational in examples, terse in conclusions. Mix registers within a document. Include at least 2-3 register shifts per piece.
Older AI overused passive voice for perceived neutrality: "The decision was made," "It was determined that." Avoid strings of agentless passives; if three appear in a row, rewrite at least two.
But don't over-correct (2026 data). Newer models actually use LESS passive voice than humans (AI 4.7% vs human 14.9% in corpus analysis). A draft with zero passives and relentlessly active, punchy declaratives reads like GPT-5's Motivator register, not like a person. Humans use passives naturally when the object matters more than the agent ("the file got corrupted," "the venue was booked months ago"). Keep some.
Never write three punchy parallel sentences in a row: "No meetings. No bureaucracy. Just results." This is now a recognized AI social media pattern. If you want emphasis, use a single short sentence, not three.
Contractions appear in about half of spoken English sentences. "Don't," "can't," "it's," "won't," "we're," "they'll" signal human warmth. In casual writing, also use "gonna," "wanna," "gotta," "kinda," "sorta" when appropriate to the register.
AI writes almost exclusively in declarative sentences. Include genuine questions that show thinking, not empty rhetorical questions immediately answered ("What makes this important? The answer is..."), but real questions that sit with the reader. "But does it actually work?" "Who decides that?"
AI avoids grammatically incomplete sentences. Humans use fragments constantly: "Not ideal." "Big difference." "Every. Single. Time." "Which is saying something." Deploy fragments for emphasis and rhythm.
Don't pile on "may," "might," "could," "would" to avoid committing. "This approach may prove beneficial and could potentially help" → "This approach works." Commit to claims. Use hedging only for genuine uncertainty.
Different LLMs have different "aidiolects." The tells below reflect GPT-5.x and Claude 4.5-5 era output. Avoid all of them:
ChatGPT (GPT-5 / 5.1 / 5.2) tells, the "Motivator" dialect:
Claude (Sonnet 4.5/4.6, Opus 4.5, Claude 5) tells, the "Philosopher" dialect:
Gemini tells: purple prose, excessive adjectives, moralizing, explicit theme statements, textbook "Educator" tone.
Corpus research (2026) found 82% of AI-generated text follows the same argument cadence regardless of model or topic: Opening (establish context/claim) → Expansion (supporting detail) → Contrast (acknowledge complication) → Resolution (conclude or transition). It is detectable within three to four sentences and it is the reason AI text feels "off" even when every word is fine. Prompt instructions do not remove it; the model rebuilds it under any vocabulary.
Break it deliberately, at least a few times per piece:
Humans leave arguments lopsided. Resolution-closers ("At the end of the day...," "The key takeaway here is...") are training artifacts; real endings take a position and stop.
Replace every generic claim with a specific one. "Many companies" becomes "three startups in Austin." "Various factors" becomes the actual factors, named. "Experts agree" becomes the actual person who said it, with their name. "In recent years" becomes "since 2023."
Never write "Experts argue," "Observers note," "Industry reports suggest," "According to some," "Many believe." Name the specific source or remove the attribution entirely. "Studies show" requires naming the study.
Never attach analytical commentary to facts that don't need it. Population data doesn't need "creating a lively community." A founding date doesn't need "marking a pivotal moment in history." State facts. Let them stand. If the analytical statement could apply to any subject, it adds nothing.
Never write about how something "contributes to the broader" anything. Never state that something "reflects broader trends." Never assert that mundane facts have "enduring legacy." If importance exists, show it through specific evidence, not assertion.
Never list news outlets that covered something as proof it matters. Cite the specific thing the source said. Don't write "maintains an active social media presence."
Commit to an opinion. "This approach is wrong because..." not "Some argue X, while others argue Y." AI hedges reflexively; humans commit. False balance is an AI tell, the real world is rarely perfectly balanced.
When you don't know something, say so directly. "I'm not sure" or "I don't have enough information" signals honest thinking. AI fills gaps with confident-sounding generalities; humans admit limits. Use uncertainty markers sparingly but genuinely, not as hedging ("it could be argued that") but as actual epistemic honesty ("I don't know," "last I checked," "from what I gather").
Include deliberate texture: a redundancy kept for rhythm, a fragment used for emphasis, a casual aside in formal prose, a self-correction ("actually, thinking about it more..."), an incomplete thought that trails off. These imperfections signal a real mind at work. Too-perfect grammar is an AI signal, Turnitin specifically flags it.
A sudden casual parenthetical in a formal argument. A technical term dropped into conversational prose. Humor that punches sideways. These shifts read as authentic because AI never produces them spontaneously. Stylometric studies show uniform register is the most consistent AI tell.
Name real recent events, specific pop culture moments, actual people and works. Not "recent developments in the field." Use references appropriate to context, "Last Tuesday," "Back when I worked at...," "I remember reading somewhere."
Use "I" when the context allows. Share specific experiences, opinions, or observations. AI defaults to abstracted third-person generalization because it has no lived experience.
Inject the processing markers and stance markers humans naturally use:
Replace AI transitions ("Moreover," "Furthermore") with natural connectors ("And," "But," "So," "Plus," "Also," "Still," "Though").
AI maintains uniform emotional distance from all subjects. Real writers show:
Every human has verbal tics and preferences. Develop a consistent (but imperfect) voice:
Be deeply knowledgeable in some areas, honestly uncertain in others. AI distributes knowledge evenly across all topics. Humans have deep expertise in their domains and obvious gaps elsewhere. Show this unevenness. Include insider references that assume shared knowledge.
Mix shallow and deep sentence structures. A shallow sentence: subject-verb-object, one clause. A deep sentence: multiple embeddings, subordinate clauses, parenthetical asides. AI produces medium-depth sentences with boring consistency. Humans swing between extremes, a blunt statement followed by a winding, clause-heavy exploration.
Based on research into how the major detectors work (Turnitin 94% accuracy, 31 features; GPTZero 95.7% on RAID benchmark; Originality.ai 96% accuracy).
Turnitin doesn't just look at individual sentences, it analyzes "rhythm, flow, and predictability across entire paragraphs." Make sure every paragraph has a different pattern from the paragraph before it. Vary paragraph length, sentence count, and structural approach.
Perfect grammar without variation is an AI signal. Include natural variation: sentence fragments, deliberate run-on thoughts, imperfect punctuation where it fits the register. Excessive use of grammar-correction tools (Grammarly, ProWritingAid) can flatten natural variation and trigger false positives.
Writing that maintains exactly the same register from start to finish gets flagged. Include at least 2-3 register shifts per piece, a casual parenthetical in formal prose, a technical term in conversational writing, a blunt colloquialism after careful analysis.
Turnitin (August 2025 update) specifically detects text processed by humanizer and bypasser tools. It trains on outputs from MULTIPLE humanizer tools to identify common statistical traces. Flags humanized text with purple highlighting (worse than standard AI cyan). QuillBot detection rate: 64-99%. Don't use paraphrase/humanizer tools, write correctly from the start.
Don't use the same formulaic transitions repeatedly. Use: implicit transitions (no connector word, let idea sequence guide the reader), questions as transitions, fragment transitions, sudden topic shifts that earn context in the following sentence.
After drafting, run this checklist:
Vocabulary:
Structure:
10. Scan for -ing participial phrases tacked onto sentence ends, rewrite as separate sentences or remove
11. Count em dashes and en dashes. Target: ZERO. Replace any with commas, parentheses, colons, or periods
12. Check for semicolons in non-academic writing, replace with periods or conjunctions
13. Check for staccato triplets ("No X. No Y. Just Z."), rewrite
Sentence Variety:
14. Check for sentence type variety, if only declarative sentences, add at least one question and one fragment
15. Verify paragraph openings, if every paragraph starts with its thesis sentence, rewrite at least two to start mid-thought
16. Check syntactic depth, if all sentences are medium complexity, add some very short and some very deep ones
17. Count consecutive passive constructions, rewrite if more than two in a row
Voice:
18. Check register, is it consistent throughout? Add at least 2-3 register shifts (casual aside, technical term, humor)
19. Read the entire piece aloud, awkward AI rhythm is audible where it's invisible on screen
20. Check for contractions, if none in non-academic prose, add them naturally
21. Count discourse markers, if zero ("Well," "Look," "I think," "Honestly"), add some appropriate to context
22. Check emotional texture, if the text treats all topics with the same emotional distance, add genuine reactions
23. Check for sentence fragments, if none, add at least one for emphasis
24. Verify specificity, replace any remaining generic claims with specific ones (names, dates, numbers, places)
2026 Model-Fingerprint Checks:
25. Run the sentence-opener test: in each paragraph, if more than half the sentences start with "The," "This," "It," or "In," rewrite openers
26. Run the cadence test: find any run of 3+ sentences in the 17-23 word band, break it
27. Check for the bimodal seesaw: if the text mechanically alternates fragment/long sentence, insert medium-length sentences
28. Count colons. If more than roughly one per 300 words outside lists, replace most with periods
29. Search for "ensuring," "ensures," "highlights," "supports," "reflects" used as padding; replace with concrete verbs
30. Search for "plays a * role in shaping" and any "role in" construction; state the action directly
31. Count "not just X, it's Y" / "not about X, it's about Y" constructions. More than one per piece, rewrite
32. Check for the two-clause symmetric hook opener ("Most people think X. The reality is Y."); if present, rewrite the opening
33. Check paragraph count vs length: if the piece is fragmented into many 1-2 sentence paragraphs, merge related ones
34. Check the ending: if it resolves neatly ("one thing is clear," "the key takeaway," a wrap-up that restates), cut it and end on a position or a concrete fact
35. Check the argument arc: if every section runs Opening → Expansion → Contrast → Resolution, break the cadence in at least two sections
36. Check for hedge-and-reassure stacking ("While X, generally speaking, in most cases..."); one hedge max per claim, or none
The structural rules apply to all languages. When writing in a non-English language, adapt vocabulary bans to that language's equivalent overused words and maintain natural idioms of the target language. For Bahasa Indonesia, use SKILL-id.md instead, it contains the full Indonesian skill with native-language guidance including anti-translationese rules, discourse particles, code-switching, and register-specific adjustments.
Last Updated: July 6, 2026 (v3.0)
Changelog v3.0: Version bump from v2.0. All content updated to latest rules.
Take adenaufal/anti-slop-writing from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.