Universal Anti-Slop & Anti-Hype (85 Deep Stylometric & Cognitive Rules)

Why can experienced readers still "smell" AI writing or cheap marketing even after you delete obvious words like delve, tapestry, and seamless? Because Large Language Models have a deeper cognitive and syntactic signature: uniform information entropy per sentence, see-saw concession clauses, polite professorial neutrality, abstract noun chains, symmetrical paragraph geometry, and forced trend-jacking.

Apply all 85 rules across any human language (English, Vietnamese, Spanish, French, German, Japanese, Chinese, Korean) to strip both surface cliches and deep structural AI tells.

Dimension 1: Information Entropy & Discourse Geometry (Rules 1-10)

1. Kill Uniform Information Entropy because LLMs distribute information evenly so every sentence carries roughly the same lukewarm weight. Human writers compress high-surprise facts into dense bursts, followed by ultra-short punchlines.
2. Ban the 3-Sentence "Sandwich Paragraph" where LLMs follow the exact formula [Sentence 1: Abstract claim] -> [Sentence 2: Dual/triple example] -> [Sentence 3: Moralizing summary tail]. Break this mold: let a paragraph be a single 6-word verdict, or a 5-sentence causal narrative.
3. Vary Paragraph Visual Density by never writing four consecutive paragraphs of 3-4 lines each. Alternate short and long blocks so the page looks human-typeset before a single word is read.
4. Delete Throat-Clearing Macro Openers that announce the world is changing or a topic is important ("In today's...", "Trong bối cảnh...", "En el mundo actual..."). Start with a concrete actor, number, or event in Sentence 1.
5. Delete Summary Echo Closers starting with "In conclusion,", "Ultimately,", "All in all,", or "Tóm lại," that repeat the thesis. End at peak utility or on a concrete image.
6. Eliminate Heading-to-First-Sentence Echo by never repeating the words of an H2/H3 heading in the first sentence below it. Jump straight into the mechanism or story.
7. Ban Colon-Split Brochure Titles separated by a colon ("The Future of Caching: How Redis Transforms Latency", "Tối ưu hóa dữ liệu: Chìa khóa cho doanh nghiệp"). Write single-clause declarative or tension headings ("Why Redis evicted our session keys at 60% RAM").
8. Break the Rule-of-Three Reflex by restricting symmetrical triplets (A, B, and C) to at most once per 500 words. Use a single exact word, a 2-item contrast, or an uneven 4-item list.
9. Ban Bold-Lead Bullet Soup that formats explanations as vertical lists of - **Bold Label**: one-sentence generic definition. Write connected prose paragraphs for causal explanations and use compact tables for comparisons.
10. Kill Formulaic "Challenges and Future Outlook" Tail Sections that append unrequested closing sections titled "Challenges and Limitations" followed by "Future Prospects". Embed trade-offs directly inside the main workflow steps.

Dimension 2: Deep Syntactic & Clause Architecture (Rules 11-22)

11. Ban Negative Parallelisms ("Not Just X, It's Y") across all languages by cutting "It's not just X, it's Y", "More than merely X, it is Y", "Không chỉ X mà còn Y", and "No es solo X, sino Y". State directly what the thing does.
12. Kill See-Saw Concession Openers ("While X, Y") that start every third sentence with a balanced concession clause such as "While X offers significant benefits, Y remains...", "Although A is simple, B provides...", or "Mặc dù X rất quan trọng, nhưng Y..." (English gloss: "Although X is very important, Y..."). State the boundary condition directly ("X works up to 10k rows. Past 10k rows, use Y.").
13. Chop Superficial Participial / Gerund Commentary Tails attached after a comma to explain the "meaning" of the sentence ("..., highlighting the need for...", "..., ensuring seamless...", "..., reflecting a broader shift...", "..., qua đó khẳng định...", "..., góp phần nâng cao...").
14. Stop Copula Inflation (Use Plain "Is" and "Has") by replacing "serves as", "stands as", "functions as", "boasts a", "represents a", "đóng vai trò như một", and "được xem như là" with plain is, are, has, là, or có.
15. Ban Rhetorical Self-Q&A Micro-Clichés where LLMs fake conversational punch by asking 2-word questions ("The result? Faster builds.", "The catch? It costs more.", "The best part?", "Why does this matter?", "Here's the kicker:", "Kết quả là gì?"). Write normal, connected clauses instead.
16. Kill Nominalization Chains (Abstract Noun Subjects) by replacing "The utilization of batch processing facilitates the minimization of network overhead" or "Việc áp dụng quy trình mới mang lại sự cải thiện cho hiệu suất" (English gloss: "The application of the new process brings an improvement to performance") with active human syntax: "Batching requests cuts network overhead in half" / "Gộp request giúp giảm một nửa băng thông" (English gloss: "Batching requests cuts bandwidth in half").
17. Eliminate Passive Agency Erasure such as "Mistakes were made during the migration" or "A decision was reached to deprecate...". Name the actor ("We dropped the wrong table", "The core team deprecated...").
18. Prevent Translationese & Pronoun Stuffing in Pro-Drop Languages when writing in Vietnamese, Japanese, Korean, or Spanish. Never mechanically translate English passive voice ("được thiết kế để được sử dụng bởi...") or stuff pronouns ("bạn", "chúng tôi", "nó", "của bạn") into every clause where native speakers naturally omit them.
19. Ban False Universal Ranges ("From X to Y") such as "from solo developers to Fortune 500 enterprises", "from ideation to execution", or "từ người mới bắt đầu đến chuyên gia". Name the exact target cohort.
20. Strip Mechanical Transition Stacking by deleting sentence-initial "Furthermore,", "Moreover,", "Additionally,", "Consequently,", "In essence,", "At its core,", "Hơn thế nữa,", and "Bên cạnh đó,". Let logic carry the transition.
21. Avoid Synonym Cycling (Elegant Variation) where a writer calls a tool "the queue" in line 1, "this messaging backbone" in line 2, and "the asynchronous pipeline" in line 3. Repeat "the queue" or "it".
22. Cut Preparatory "It" & Cleft-Sentence Padding by replacing "It is important to recognize that it is the network layer that causes..." with "The network layer causes...".

Dimension 3: Stylometric Rhythm & Burstiness Math (Rules 23-30)

23. Enforce High Sentence-Length Burstiness (CV >= 0.35, SD >= 5.0 words) by never writing five consecutive medium-length sentences (15-22 words). Pair a 28-word sentence that traces a multi-step mechanism with a 3-to-5 word sentence that lands the consequence.
24. Vary Syntactic Tree Depth by alternating right-branching sentences (subject + verb + object + detail) with occasional direct imperatives ("Check the logs first.") or short conditional rules ("If p99 exceeds 200ms, drop the cache.").
25. Ban Consecutive Sentence-Opener Repetition by never starting two consecutive sentences with the exact same word (The... The..., This... This..., We... We...) unless using deliberate anaphora in a speech.
26. Strict Em-Dash Quota (< 2 per 1,000 words) because LLMs use em-dashes (— or --) to glue clauses together lazily. Use periods, parentheses, or colons instead.
27. Ban Semicolon Chains in Non-Academic Prose instead of joining balanced generic clauses; split them with a period.
28. Use Natural Contractions in English (it's, don't, won't, we've, didn't). Zero contractions in everyday prose sounds like a stiff translation.
29. Enforce List-Item Length Asymmetry when using bullet points so not every bullet is 18-22 words long. Let one item be 4 words and another be 32 words based on real substance.
30. Eliminate Triad Adjective Stacks by never placing three adjectives before a noun ("a fast, scalable, and intuitive CLI"). Use one verifiable qualifier ("a 4MB single-binary CLI").

Dimension 4: Epistemic & Pragmatic Voice - Killing the "Polite Professor" (Rules 31-42)

31. Ban Significance Inflation that announces something is "crucial", "pivotal", "vital", "a testament to", "transformative", or "then chốt". Show the concrete cost of ignoring it.
32. Ban Vague Weasel Attributions such as "Experts argue", "Studies show", "Industry leaders agree", or "Nhiều chuyên gia cho rằng". Name the author, paper, or dataset, or own the claim directly.
33. Zero Chatbot Meta-Frames & Sycophancy by stripping "Certainly!", "Great question!", "Here is a breakdown:", "Let's dive in", "I hope this helps", and "Without further ado".
34. Eliminate Micro-Hedging Stacks by cutting cowardly qualifiers like "tends to somewhat", "can often serve to", and "may potentially help". State the exact condition under which the statement holds true.
35. Kill Fence-Sitting Both-Sidesism that concludes a comparison with "Ultimately, the right choice depends on your unique business needs." Give a sharp decision rule: "Pick A if your team is under 5 people; pick B once write volume passes 1,000 req/sec."
36. Replace Unquantified Intensifiers with Numbers by banning "blazingly fast", "massively scalable", "dramatically reduced", and "vô cùng hiệu quả". State the before-and-after numbers.
37. Ban Self-Referential Artifact Narration such as "This article explores...", "In this section, we will discuss...", or "This commit adds...". Start with the subject itself.
38. Kill Toxic Positivity / Emotional Sanitization where LLMs turn every failure into a cheerful moral lesson ("While the outage was challenging, it provided an invaluable learning opportunity for our team"). Humans admit when something simply sucked ("The outage cost us $18,000 and ruined our weekend. We changed two rules on Monday so it wouldn't happen again.").
39. Ban Pedagogical Lecture Transitions like "To truly understand X, one must first examine Y" or "Before diving into X, let's take a step back and define Y". Assume the reader is a smart peer.
40. Ban Pompous Coined Frameworks that invent capitalized buzzword labels for ordinary actions (e.g., calling "checking your logs" "The Proactive Observability Paradigm").
41. Avoid Moralizing "Should / Must" Preaching by showing the consequence ("Leave port 6379 open without a password and a crypto-miner will find your Redis instance in under 12 minutes") instead of lecturing ("Developers must always prioritize security").
42. Preserve Verbatim Quotes Untouched by never paraphrasing or "cleaning up" text inside quotation marks or blockquotes.

Dimension 5: Abstract-to-Concrete Ratio (ACR) & Sensory Grounding (Rules 43-50)

43. The 35-Word Concrete Floor requires every 35 words of prose to contain at least one concrete anchor (a physical noun, an exact number/unit, a specific file/tool name, or a visible human action). If a 40-word span contains only abstract nouns (strategy, alignment, optimization, ecosystem, value), rewrite it immediately.
44. Ban "Ghost Actors" so the landscape, the ecosystem, modern organizations, or stakeholders never serve as the subject of a sentence. Name the actual person (the on-call engineer, your CFO, a first-time buyer).
45. Replace Cliché Metaphors with Mechanical Analogies by banning tapestry, beacon, north star, double-edged sword, tip of the iceberg, chìa khóa vàng, and kim chỉ nam. If you use an analogy, make it mechanical ("acts like a check valve that only lets traffic flow one way").
46. Ground Time References by replacing "recently", "in the past", and "nowadays" with exact temporal anchors ("in Q3", "after Postgres 16 shipped", "at 2:14 AM Tuesday").
47. Ground Failure Modes by replacing "performance may degrade" with "latency spikes from 40ms to 4.2s and the tab freezes".
48. Ground Human Behavior in Muscle Memory by describing the physical action ("saves you from scrolling horizontally across 19 columns to find the Delete button") instead of "improves user experience".
49. Ban Vaporware Verbs including empower, unlock, unleash, elevate, foster, spearhead, harness, revolutionize, catalyze, and streamline. Use physical or operational verbs (cut, block, merge, strip, route, cache, split).
50. Ban Empty Determiner Phrases such as "a wide range of", "a variety of", "numerous different", and "nhiều loại ... khác nhau". Name 2 specific examples or give the exact count.

Dimension 6: Multi-Language Lexical & Calque Blacklists (Rules 51-60)

51. English Tier-1 Instant-Ban Words include delve, tapestry, realm, pivotal, seamless, seamlessly, robust, game-changer, leverage (as verb), elevate, unlock, unleash, landscape (figurative), crucial, meticulous, multifaceted, holistic, synergy, paradigm, beacon, cornerstone, catalyst, plethora, myriad, and indelible.
52. English Tier-2 Phrase Bans include "in today's fast-paced", "it's important to note", "plays a vital/pivotal/crucial role", "stands/serves as a testament", "sits at the intersection of", "strikes a balance", "at the heart of", "paves the way for", "a deep dive into", and "in the wake of".
53. Vietnamese Tier-1 "AI Boilerplate" Bans include "trong bối cảnh" ("in the context of"), "trong kỷ nguyên số" ("in the digital era"), "không thể phủ nhận" ("it is undeniable"), "như chúng ta đã biết" ("as we all know"), "hãy cùng khám phá" ("let's explore"), "đóng vai trò quan trọng/then chốt" ("plays a vital/pivotal role"), "mang đến trải nghiệm mượt mà/toàn diện" ("delivers a seamless/holistic experience"), "chìa khóa vàng" ("golden key"), "kim chỉ nam" ("guiding compass / north star"), "mở ra cánh cửa" ("opens the door to"), "khẳng định vị thế" ("affirms the position"), "tạo tiền đề" ("creates the premise / paves the way"), "góp phần nâng cao" ("contributes to enhancing"), "hành trình khám phá" ("journey of discovery"), and "bức tranh toàn cảnh" ("the big picture / panorama").
54. Vietnamese Tier-2 "Machine-Translated Boilerplate" Bans include "Không chỉ ... mà còn ..." ("Not only ... but also ..."), "Một sự [Verb] đã được..." ("A [nominalized verb] was performed..."), "Việc thực hiện [Verb]..." ("The implementation of [Verb]..."), "Tóm lại, có thể thấy rằng..." ("In conclusion, it can be seen that..."), plus unnatural repetition of pronouns "bạn / của bạn / chúng tôi" ("you / your / we") in every clause where native Vietnamese omits the pronoun.
55. Spanish AI-Slop Bans include "en el vertiginoso mundo actual", "juega un papel fundamental/crucial", "no es solo X, sino Y", "un tapiz de", "es importante destacar que", and "en conclusión".
56. French AI-Slop Bans include "dans un monde en constante évolution", "joue un rôle crucial/clé", "ce n'est pas seulement X, c'est Y", "il est important de noter que", and "une pléthore de".
57. German AI-Slop Bans include "in der heutigen schnelllebigen Welt", "spielt eine entscheidende Rolle", "nicht nur X, sondern auch Y", and "es ist wichtig zu beachten".
58. CJK (Chinese / Japanese / Korean) AI-Slop Bans cover Chinese (在当今快节奏的, 不仅仅...而是..., 发挥着至关重要的作用, 总而言之, 让我们一起探索), Japanese (〜において重要な役割を果たしています, 単なる〜ではなく, 〜と言っても過言ではありません, 結論として), and Korean (오늘날 빠르게 변화하는, 단순히 ~일 뿐만 아니라, 중요한 역할을 합니다).
59. Code & Commit Message AI Tells in technical repos ban "This commit leverages/enhances...", "comprehensive error handling", "production-ready", "blazingly fast", and trailing ", ensuring consistency" in commit messages.
60. Emoji & Typographic Clutter bans rocket ships, sparkles, fire, lightbulbs, and green checkmarks in prose, headings, and professional copy.

Dimension 7: Positive Human Texture & Cognitive Authenticity (Rules 61-70)

To make writing indistinguishable from a sharp human practitioner, inject at least 3 of these 10 markers:

61. The Mid-Thought Self-Correction qualifies a thought the way a real person thinks in real time ("That sounds like a network bug - and for the first two hours we thought it was - until we checked the disk inode table.").
62. Candid Parenthetical Aside drops a dry, realistic detail in parentheses (mostly because nobody wanted to touch the 2019 billing cron job on a Friday afternoon).
63. Earned Understatement Over Hyperbole understates the adjectives when a result is impressive and lets the raw numbers do the work ("Cutting the payload from 1.8MB to 14KB made the mobile app noticeably less annoying to open on train Wi-Fi.").
64. Specific Annoyance / Friction Detail names the small, lived friction that proves you actually did the work ("having to paste a 64-character token into three different .env files").
65. First-Principles Causal Chain explains how A causes B mechanically ("Because Node uses a single event loop, parsing a 40MB JSON string blocks every incoming health-check ping for 600ms").
66. The 1-to-4 Word Verdict Sentence lands a complex paragraph with a blunt short sentence ("It didn't work.", "That fixed it.", "Dead simple.").
67. Direct Second-Person Muscle-Memory Callout references an exact micro-experience the reader recognizes ("You know that sinking feeling when git rebase stops on file 14 of 40...").
68. Dry Contrast Humor pairs a grand expectation with a mundane reality ("We architected the queue for 100,000 events per second. Our first paying customer sent 12 events a day, all at 9:00 AM, and still managed to break our parser with a semicolon in their company name.").
69. Un-Rounded Forensic Numbers prefer real, un-rounded measurements (14.2%, 38 minutes, 19 tables) over suspiciously round numbers (10%, 1 hour, 20 tables).
70. Abrupt High-Signal Closer stops writing the exact moment the final useful point lands, without adding a "warm sign-off" paragraph.

Dimension 8: Execution & Anti-Drift Protocol (Rules 71-75)

71. The 300-Word Block Audit (Anti-Drift Rule) prevents LLM stylistic decay after paragraph 4 on long texts (>400 words) by auditing and rewriting in 300-word chunks so paragraph 10 is just as sharp as paragraph 1.
72. Restructure Clauses, Never Just Swap Synonyms because swapping "leverage" for "use" inside a robotic sentence leaves the AI sentence skeleton intact. When a sentence triggers slop_scan.py, rewrite the subject and verb from scratch.
73. The First-and-Last Sentence Compression Test inspects the first sentence and the last sentence of the draft; if deleting either one loses zero factual information, delete them both.
74. The Coffee-Shop Read-Aloud Test reads every line aloud; if you would feel embarrassed saying a phrase to a respected colleague across a coffee table, cut it.
75. Mandatory Automated Verification (tools/slop_scan.py) runs python3 tools/slop_scan.py <file> --fix-hints whenever shell execution is available, revising until the score is <= 15/100 (CLEAN).

Dimension 9: Forced Trend-Jacking & Scam/Hype Marketing Tells (Rules 76-85)

76. Ban Forced News-Peg Openers such as "With the recent rise/explosion of [Trend]...", "As we saw in the news this week...", "In light of recent events...", "Trước làn sóng..." ("Facing the wave of..."), and "Với sự bùng nổ của..." ("With the explosion of..."). If a market shift matters (skills/market-intelligence), state its direct operational effect on the user's workflow without announcing the trend.
77. Ban Unrelated Metaphor Trend-Jacking that compares a mundane tool or topic to a trending celebrity, viral meme, or macro headline without a direct causal link (Organic Fit Gate).
78. Ban Scam ROI Multipliers such as "10x your productivity/revenue", "guaranteed ROI", "skyrocket your sales", "explode your growth", "x10 doanh thu" ("10x revenue"), and "cam kết lợi nhuận" ("guaranteed profit").
79. Ban "Secret Blueprint" Guru Framing including "the secret blueprint", "what the gurus won't tell you", "the underground method", "bí mật triệu đô" ("million-dollar secret"), and "công thức bí truyền" ("secret formula").
80. Ban Manufactured Scarcity Phrases such as "spots are filling up fast", "act now before it's too late", "once-in-a-lifetime opportunity", "chỉ còn duy nhất X suất" ("only X spots left"), and "đăng ký ngay kẻo lỡ" ("register now before you miss out").
81. Ban Manipulative Decline Buttons (Confirmshaming) like "No thanks, I prefer wasting time".
82. Ban Fake Authority Name-Dropping such as "Used by top engineers at Google, Meta, and Netflix" unless citing a verifiable enterprise agreement or public repo usage.
83. Ban Income / Lifestyle Bait including "financial freedom", "passive income machine", "work from the beach", and "tự do tài chính" ("financial freedom").
84. Ban Exclamation-Mark Inflation by keeping 0 exclamation marks in B2B/technical copy and at most 1 per 1,000 words in consumer copy. Never end a headline with !.
85. Ban All-Caps Urgency Prefixes such as "URGENT:", "BREAKING:", "WARNING:", or "LAST CHANCE:".