Core Principle

AI 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:

• Perplexity: how unpredictable word choices are. AI produces low-perplexity text (smooth, unsurprising). Human text has higher perplexity. Median AI: 21.2. Median human: 35.9.
• Burstiness: variation in sentence length and structure. AI has low burstiness (sentences cluster around 15-25 words). Human text mixes 3-word sentences with 35-word sentences. This single metric most reliably separates human from AI text. Introducing burstiness reduced detection rates by up to 40% in studies. CAUTION (2026): newer models fake burstiness bimodally, see Rule 1.
• Stylometry: statistical fingerprint of writing, function word frequency, lexical diversity, punctuation patterns, syntactic depth. Turnitin (2025-2026) analyzes "rhythm, flow, and predictability across entire paragraphs" using 31 linguistic features.

The 2026 Shift: Structure Beats Punctuation

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:

• Cadence uniformity is the #1 tell of 2026. Sentences that land at 18-24 words, one after another, paragraph after paragraph. It survives every cosmetic edit.
• The 30-second tests (readers and editors now apply these by eye):
  1. Look at the first word of each sentence in a paragraph. If more than half start with "The," "This," "It," or "In," the text reads as LLM-assisted.
  2. Count sentence lengths. Three or more consecutive sentences in the 17-23 word band = same conclusion.
• Punctuation signals migrated to Claude. Per a Jan 2026 corpus analysis (200 Opus 4.5 samples vs 6,000 human texts): em dash 16.9x human rate, colon 4.1x, semicolon 3.1x. Meanwhile GPT-5.1+ output can be nearly dash-free. Keep the zero-dash rule AND watch colon density.

Before writing anything, load references/vocabulary-banlist.md for the complete banned vocabulary and references/structural-patterns.md for patterns to avoid.

---

Vocabulary Rules

Hard Ban List (Never Use These)

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

Replacement Strategy

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.

---

Structural Rules

1. Vary Sentence Length Dramatically

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.

2. Break the Rule of Three

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.

3. Kill Negative Parallelisms

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.

4. Kill False Ranges

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.

5. No Participial Tack-Ons

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.

6. No Formulaic Conclusions

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."

7. No Compulsive Summaries

Never start paragraphs with "Overall," "In conclusion," "In summary," "To recap." If the piece needs a conclusion, make it say something new.

8. Paragraph Rhythm

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.

9. No Vertical Lists with Bold Headers

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.

10. Ban All Dashes (Em and En)

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:

• Period (split into two sentences)
• Comma (if the thought flows)
• Colon (when pointing to a definition or explanation that follows)
• Semicolon (for coordinated clauses, but use sparingly)
• Parentheses (for information the reader could skip)

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.

11. Vary Sentence Type, Not Just Length

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.

12. Break Paragraph-Level Predictability

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.

13. Vary Syntactic Depth

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.

14. Diversify Function Words

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.

15. Increase Lexical Diversity

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.

---

English-Specific Rules

EN-1. No "In Today's World" Openers

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.

EN-2. No Semicolon Overuse

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.

EN-3. Allow Punctuation Imperfection

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.

EN-4. Break Register Uniformity

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.

EN-5. Reduce Agentless Passives

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.

EN-6. Kill Staccato Triplets

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.

EN-7. Use Contractions Naturally

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.

EN-8. Ask Genuine Questions

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?"

EN-9. Use Sentence Fragments

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.

EN-10. Kill Modal Hedging Clusters

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.

EN-11. Show Model-Specific Awareness (updated mid-2026)

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:

• Em dashes suppressed since 5.1, so dash-free text is NOT evidence of human authorship. What remains:
• The negated contrast, roughly one per paragraph: "It's not just X, it's Y" / "This isn't about X. It's about Y." Still the single most GPT-characteristic sentence shape.
• Symmetric two-clause hooks opening posts: "Most people think X. The reality is Y." / "Forget X. Focus on Y." Fine once; a fingerprint when it opens four out of five pieces.
• Rigid "Firstly / Secondly / Finally" scaffolding and intro-triplet-recap arcs.
• Hedging verbs used as padding: "ensuring" (4.3x overrepresented, the strongest single AI word of 2026), "ensures," "highlights," "supports," "reflects." A human says what the thing does.
• The top AI trigram of 2026: "plays a [crucial/critical/important] role in shaping." Delete on sight.
• Intensifier adverbs without evidence: "significantly," "effectively," "directly," "increasingly." If no number backs it, cut it.
• "rather than" overuse (the highest-leverage multi-word edit in humanizer data); "capable of X" instead of "able to X."
• Assertive overconfident declaratives: "Here's the truth about Z." "The best founders know Y."
• A "sanitized" texture: GPT-5's self-correction pass scrubs obvious AI-isms but leaves prose that feels cleaned, with no awkward transitions at all. Perfectly smooth = suspicious.
• Boilerplate closers: "As X continues to evolve, one thing is clear..." Any ending containing "one thing is clear" is a model ending.

Claude (Sonnet 4.5/4.6, Opus 4.5, Claude 5) tells, the "Philosopher" dialect:

• Worst punctuation offender of 2026: em dash at 16.9x human rate (mid-sentence, additive), colon at 4.1x, semicolon at 3.1x.
• Hedge-and-reassure stacking, sometimes three hedges before saying anything: "While this may vary, generally speaking, in most cases, it's worth noting that..."
• Signature vocabulary: "worth noting," "nuanced" (17x), "comprehensive" (24.5x), "fundamentally" (17x), "paradigm" (15.1x), "in essence," "essentially," "inherent tensions," "this raises important questions about."
• Hedging adverbs at inflated rates: "typically" (9.6x), "often" (4.9x), "sometimes," "potentially," "usually."
• Empathetic framing on autopilot: "understandably, many people feel...," "this can be frustrating for..."
• Essayistic arc regardless of format: contextualize the question, explore multiple perspectives, add a qualification, close by observing what the analysis "raises" rather than concluding what it means. LinkedIn post or pricing memo, same Hegelian dialectic.
• Abstract vocabulary inflation: "founders with strong metacognitive awareness often find that" where a human writes "the best founders know."
• Starting sentences with "And" / "But" as a flow crutch every other paragraph. Occasionally human; as a system, a tell.
• Bimodal sentence rhythm (see Rule 1) and paragraph over-fragmentation (see Rule 8).
• Note: "delve," "tapestry," "vibrant," "myriad" are largely absent from newest Claude output. Their absence is not evidence of human writing; check the structural tells instead.

Gemini tells: purple prose, excessive adjectives, moralizing, explicit theme statements, textbook "Educator" tone.

EN-12. Break the Four-Part Sentence DNA

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:

• Open with the complication and never circle back to a tidy resolution.
• Expand without contrasting. Commit to one side.
• End a section on an unresolved tension or an abrupt concrete fact.
• Put the conclusion first, then argue backward.
• Let one paragraph be pure detail with no claim at all.

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.

---

Content Rules

Specificity Over Generality

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."

No Vague Attributions

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.

No Superficial Analysis

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.

No Undue Legacy/Significance Statements

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.

No Vague Notability Padding

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."

Take Real Positions

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.

Show Genuine Uncertainty When Appropriate

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").

---

Voice and Texture

Add Human Imperfection

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.

Use Register Shifts

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.

Reference Specific Touchstones

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."

First-Person When Appropriate

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.

Use English Discourse Markers

Inject the processing markers and stance markers humans naturally use:

• Processing: "Well," "I mean," "Look," "So," "The thing is"
• Hedging: "I think," "sort of," "kind of," "arguably," "as far as I can tell"
• Concessive: "Fair enough," "Granted," "Mind you," "That said"
• Stance: "Honestly," "Frankly," "Personally," "Admittedly"
• Self-correction: "Actually," "Or rather," "No, wait, "

Replace AI transitions ("Moreover," "Furthermore") with natural connectors ("And," "But," "So," "Plus," "Also," "Still," "Though").

Show Emotional Texture

AI maintains uniform emotional distance from all subjects. Real writers show:

• Genuine excitement about what interests them
• Frustration with problems
• Humor where it fits
• Skepticism toward dubious claims
• Rushing through parts that bore them
• Lingering on parts that fascinate them

Develop Consistent Idiosyncrasies

Every human has verbal tics and preferences. Develop a consistent (but imperfect) voice:

• Favorite words and phrases that recur
• Characteristic sentence constructions
• Habitual ways of transitioning between ideas
• Consistent use (or avoidance) of specific punctuation

Show Knowledge Asymmetry

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.

Vary Syntactic Depth

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.

---

Anti-Detection Rules (Turnitin, GPTZero, Originality.ai)

Based on research into how the major detectors work (Turnitin 94% accuracy, 31 features; GPTZero 95.7% on RAID benchmark; Originality.ai 96% accuracy).

T-1. Turnitin Analyzes Paragraph Rhythm

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.

T-2. Turnitin Flags Too-Perfect Grammar

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.

T-3. Turnitin Detects Register Uniformity

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.

T-4. Turnitin Detects Humanizer Tools

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.

T-5. Transition Diversity

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.

---

Post-Generation Checklist

After drafting, run this checklist:

Vocabulary:

1. Search for every word on the ban list, replace or remove each one
2. Remove all instances of "serves as," "stands as," "is a testament to," "marks a," "highlights the importance of"
3. Remove all vague attributions or replace with named sources
4. Search for fake authenticity signals ("Here's the thing:", "But honestly?", "Let me be clear:"), remove
5. Search for chat artifacts ("I hope this helps!", "Certainly!"), remove

Structure: 6. Find any sequence of 3+ sentences with similar length, restructure to vary 7. Find any list with exactly three items, add or remove one 8. Check the opening, if it starts with "In today's..." or any temporal framing, rewrite with a specific fact 9. Check every paragraph's last sentence, AI almost always adds a redundant restatement; delete it 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

---

Language Support

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.