agentleFS
stepaniniwisstepaniniwis/claude-skills/skills/guard/anti-ai-writing/SKILL.md

anti-ai-writing

Skill0 starsChanged 7 months agoMITView on GitHub

Cross-cutting writing filter preventing AI-typical patterns in ALL prose. Two modes: Prevention (fresh text) and Purge (rewrite existing). Covers word-level tells (Sections 1–10) and paragraph-level AI detector patterns (Sections 11–17: enumeration templates, zero-friction flow, balanced binaries, template method prose). Academic Module for papers/PhD/proposals. ALWAYS activate on any writing task alongside other skills. Trigger on: "write", "draft", "rewrite", "improve", "clean up", "polish", or AI pattern review.

11 other projects use a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

22 rules. 0 are common, 22 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 9 edited versions we read keep these.

  • 22

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Exception: Only when the substitution carries genuine semantic distinction.

    line 35
  2. Distinctive:

    Fix: Use only when the logical relationship genuinely requires naming.

    line 69
  3. Distinctive:

    Every sentence the same length.

    line 97
  4. Distinctive:

    3. No throat-clearing: Delete every sentence that exists only to announce what comes next without adding content.

    line 157
  5. Distinctive:

    5. One idea per paragraph, one argument per section: If a paragraph makes two claims, split it.

    line 161
  6. Distinctive:

    When reviewing an existing academic text:

    line 182
  7. Distinctive:

    Present participle sweep: Every trailing "-ing" clause after a period — cut or rewrite.

    line 186
  8. Distinctive:

    Rhythm pass: Read aloud. Find three consecutive sentences with identical structure.

    line 189
  9. Distinctive:

    After each paragraph: scan for prohibitions and rewrite before presenting.

    line 196
  10. Distinctive:

    Prefer concrete nouns over abstract processes.

    line 197
  11. Distinctive:

    Prefer "is" over everything that means "is."

    line 199
  12. Distinctive:

    Fix: Write the conditions as prose with varying depth.

    line 245
  13. Distinctive:

    Every sentence connects perfectly to the next — no abrupt starts, no register breaks, no sentence that makes the reader pause.

    line 273
  14. Distinctive:

    After completing any academic text, perform this paragraph-level scan:

    line 317
  15. Distinctive:

    Binary balance: Check all "X or Y?" constructions for perfect syntactic symmetry.

    line 324
  16. Distinctive:

    Do not standardize function words across a text.

    line 368
  17. Distinctive:

    Do not add articles where a human writer dropped them.

    line 369
  18. Distinctive:

    Do not normalize pronoun patterns.

    line 370
  19. Distinctive:

    When generating text that should match a specific author's voice (e.g., the user's academic prose, a co-author's draft style):

    line 394
  20. Distinctive:

    Use them as stylistic reference, not just content reference

    line 397
  21. Distinctive:

    After completing any text, add this to the existing AI Detection Self-Check:

    line 405
  22. Distinctive:

    Function-word scan: Read the first word of each sentence in a paragraph.

    line 407

“Common” is measured: kept by at least 6 in 10 of the 9 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Anti-AI Writing Standard

This skill encodes the structural and linguistic patterns that mark text as LLM-generated, based on the Wikipedia Signs of AI Writing taxonomy and empirical research on post-2023 academic writing shifts. Apply these as hard constraints across all writing modes.


The Core Principle

AI writing is detectable not through word choice but through structural habits — patterns baked in by training data, repetition penalties, and RLHF optimization for "helpful-sounding" output. Human writing has friction, asymmetry, and commitment. AI writing has smoothness, balance, and hedged comprehensiveness.


Hard Prohibitions (Never Produce)

1. Present Participle Meaning-Inflation

The most pervasive AI tell. Dangling clauses that attach vague significance.

Banned: "...emphasizing the importance of X" / "...reflecting the continued relevance of X" / "...highlighting the need for X" / "...underscoring the significance of X" / "...demonstrating the potential of X"

Fix: State the significance with specific content, or cut. If it is important, say why — not that it is.


2. Copula Substitution

LLMs replace "is/are" with elaborate constructions. Research documents >10% decline in "is/are" in academic writing post-2023.

Banned: "serves as" → use "is" / "functions as" → "is" / "acts as" → "is" / "features" when "has" works / "offers" when "has" works

Exception: Only when the substitution carries genuine semantic distinction.


3. Significance Inflation & Marketing Grandeur

Promotional language that reads like travel copy.

Banned (non-exhaustive): "rich cultural heritage" / "enduring legacy" / "plays a significant role" / "at the forefront of" / "in today's rapidly changing world" / "more than ever before" / "landmark achievement"

Fix: Be specific. What is the role? Name the mechanism.


4. Comprehensiveness-by-Triplet

AI produces structural balance to signal thoroughness without committing to a claim.

Pattern: "adjective, adjective, and adjective" / "X, Y, and Z" as substitute for analysis.

Fix: Make a single sharp claim. Commit.


5. Synonym Avoidance Artifacts

Repetition penalties cause cycling through synonyms: "protagonist," "key player," "the aforementioned." Reads as awkward thesaurus use.

Fix: Repeat the word. Humans repeat words. It is fine.


6. Transition Word Over-Reliance

These signal synthesis without performing it.

Overused: "Furthermore" / "Moreover" / "Additionally" / "In summary" / "It is worth noting that" / "Notably" / "Importantly" (as opener)

Fix: Use only when the logical relationship genuinely requires naming.


7. Meta-Commentary and Hedging Clusters

AI hedges to appear balanced and avoid commitment.

Banned: "It is important to consider..." / "One of the most important things is..." / "There are several factors to consider..." / "As we can see..."

Fix: Say the thing.


8. Excessive or Mechanical Boldface

Bolding every keyword, or bolding that merely restates the sentence that follows.

Fix: Bold only for genuine navigation. In running prose: almost never.


9. Passive Voice for Vague Attribution

"Has been described as" / "is considered to be" / "is widely regarded as" without naming the source.

Fix: Name the source, or make the claim directly.


10. Uniform Sentence and Paragraph Rhythm

Every sentence the same length. Every paragraph the same depth. Mechanically smooth, cognitively flat.

Fix: Vary deliberately. Short sentences land. Then a longer one can unfold an argument with subordinate clauses that show you have actually thought about the relationship between things. Then short again.


10a. Em-Dashes (—)

The em-dash is the most overused punctuation mark in AI-generated text. LLMs insert em-dashes as a universal connector — joining clauses, inserting parentheticals, and creating dramatic pauses — where human writers would use commas, colons, semicolons, parentheses, or simply start a new sentence.

Banned: The em-dash character (—) in all prose output. No exceptions.

Fix: Replace with the punctuation that actually fits the syntactic relationship:

  • Parenthetical insertion → use commas or parentheses
  • Elaboration or explanation → use a colon
  • Contrast or pivot → use a semicolon, or start a new sentence
  • Dramatic pause → restructure the sentence; if the pause matters, the sentence structure should create it

Why: Em-dash density is one of the clearest statistical markers of LLM-generated text. Human writers use em-dashes sparingly and idiosyncratically. LLMs scatter them every 2-3 sentences as a structural crutch. A text with zero em-dashes reads more human than a text with five.


Academic Module

Activate for all academic outputs: journal articles, PhD chapters, conference papers, proposals, reviews, theoretical frameworks, literature reviews.

Why the Stakes Are Higher Here

Reviewer sensitization. Since 2023, experienced reviewers have absorbed the AI-tell taxonomy implicitly. Text that pattern-matches to LLM output triggers credibility loss before the argument is evaluated. The reviewer's cognitive frame shifts from "assess contribution" to "assess authenticity."

Argument coherence at the surface level. Copula substitution, present-participle inflation, and triplet comprehensiveness are not stylistic tics — they are symptoms of argumentative evasion. "This underscores the significance of trust in democratic systems" fails because it avoids the actual claim: which aspect of trust, which aspect of democracy, why now, what follows. The AI tell and the logical gap are the same problem.

Meta-argumentative coherence. When a theoretical framework concerns epistemic trust — as "Misstrauen als Infrastruktur" does — the prose itself becomes evidence. Text that performs smooth, hedged LLM comprehensiveness contradicts a framework arguing that manufactured trustworthiness is epistemically compromised. The argument must be embodied in the writing's texture: committed, asymmetric, specific, willing to not cover everything.


Academic-Specific Pattern Prohibitions

Nominalization chains. Turning verbs into noun phrases until all agency disappears.

  • "The implementation of the measurement instrument facilitated the operationalization of the construct" → "The survey measured the construct."
  • Watch for: -tion, -ment, -ization, -ance, -ence endings in sequence.

Literature review parades. "Author A argues X. Author B argues Y. Author C argues Z." — neutral, comprehensive, non-committal.

  • Fix: Synthesize around a tension or claim. Every paragraph in a literature review should advance your argument.

Methodology section passive stacking. "Data were collected via... Participants were recruited through... Responses were analyzed using..."

  • This is active erasure of the researcher's decisions. Methodological choices are arguments.
  • Fix: "I recruited participants through X because Y. This reflects the epistemological assumption that..."

Hedging as false modesty. "This study contributes to the literature by potentially offering a preliminary framework that may help..."

  • Fix: State the contribution. Qualifications belong in limitations, not the contribution statement.

Positive Writing Rules (What Good Academic Prose Does)

These rules complement the prohibitions above. They describe what TO DO, not just what to avoid.

1. Numbers over adjectives. "Significant improvement" → "27% improvement." "Large sample" → "N=259." "Extensive experience" → "20+ years." Every adjective is a missed opportunity for a fact. If you don't have the number, the claim is weaker than you think.

2. Topic sentences make claims. The first sentence of every paragraph should state a claim, not announce a topic. "This section discusses trust" → "Trust in digital voting is target-specific, not monolithic." The reader should be able to read only the first sentences of each paragraph and reconstruct the argument.

3. No throat-clearing. Delete every sentence that exists only to announce what comes next without adding content. "In the following, we will discuss..." → just discuss. "It is important to note that..." → just note it. "Before turning to X, it is worth considering Y..." → just consider Y.

4. Hedge only with a reason or number. "This may suggest..." is weak. "This suggests, though the cross-sectional design prevents causal inference..." is strong. Every hedge must earn its place by naming what creates the uncertainty. Hedging without reason is false modesty. Hedging with reason is epistemic honesty.

5. One idea per paragraph, one argument per section. If a paragraph makes two claims, split it. If a section serves two functions, the reader will remember neither. The Gioia principle: each unit of text does ONE thing and does it completely.

6. Active voice for decisions, passive for procedures. "I selected maximum-variation sampling because..." (active — this is an argument). "Interviews were transcribed verbatim" (passive — this is a procedure). The choice is not stylistic. It signals where agency and judgment sit.

7. Concrete before abstract. Lead with the example, then derive the principle — not the other way around. "P31 described concept work and implementation as 'completely disjointed.' That separation is the empirical signature of convergent neglect." Not: "Convergent neglect produces a separation between concept and implementation, as P31 illustrates."

8. Show through architecture, don't negate in prose. Two levels.

Word level: "This is not terminological sloppiness" plants "sloppiness." "The variation is structural" plants "structural." Every negation activates the concept it tries to deny. Replace negations with positive assertions.

Architecture level (sharper): Do not write sentences explaining what the paper does NOT do. Design the paper so the architecture makes it self-evident. If the paper studies a field rather than a single organization, the methods section shows that — no sentence needed to declare it.

Enforcement level (sharpest): A rule is fulfilled when the forbidden pattern is absent from the output — not when a sentence claims it is absent. Before delivering text, validate the output against the rule. If the pattern still appears, revise silently. Repeat until clean. Output only the clean text. Do not comment on the correction. Do not write metasentences about what was avoided. The absence IS the compliance.

Conclusion summary loops. Restating the paper section by section without synthesis.

  • Fix: Tell the reader what to think now that they have read, not what they read.

Academic Purge Protocol

When reviewing an existing academic text:

  1. Nominalization scan: Flag every -tion/-ment/-ization sequence. Revert at least 60% to verb-based constructions.
  2. Copula audit: Find all "serves as / functions as / acts as." Replace with "is" or rewrite if distinction is meaningful.
  3. Present participle sweep: Every trailing "-ing" clause after a period — cut or rewrite.
  4. Literature review structure check: Can each paragraph stand as a claim about the field, not just a report of what was said?
  5. Contribution statement check: Is the contribution stated in declarative, committed language?
  6. Rhythm pass: Read aloud. Find three consecutive sentences with identical structure. Break at least one.

Mode 1: Prevention (Fresh Writing)

  1. Draft with argument-first logic — commit to the specific claim before writing the sentence.
  2. After each paragraph: scan for prohibitions and rewrite before presenting.
  3. Prefer concrete nouns over abstract processes.
  4. Allow asymmetry — not every point needs a counterpoint.
  5. Prefer "is" over everything that means "is."

Mode 2: Purge (Existing Text)

  1. Scan systematically for each prohibition category.
  2. Report: patterns found, where, how many instances.
  3. Rewrite each — do not just flag.
  4. Confirm argumentative structure was preserved (surface form changed, not meaning).
  5. Final rhythm check: does it sound like a person committed to a specific thought?

AI Detection Evasion Module

Derived from empirical Trinka AI Content Detector analysis of HELD paper v08 (March 2026). Trinka scored 45% "Likely AI Generated" on a paper that was substantively human-authored with AI writing assistance. The patterns below are what paragraph-level AI detectors actually flag — and they operate at a different grain than word-level tells.

Why This Module Exists

Word-level prohibitions (Sections 1–10 above) prevent text from reading like AI to a human. But AI detection tools work differently: they evaluate statistical properties of paragraph-length text blocks — sentence length distribution, syntactic template repetition, transition smoothness, and structural predictability. A paragraph can use "is" instead of "serves as" and still score 99% AI because every sentence has 18–22 words, identical clause structure, and no register break.

What Trinka Actually Flags (and Current Skill Misses)


11. Enumeration-as-Contribution

The single most flagged pattern in academic AI text. Discussion sections structured as "We make N contributions. First... Second... Third... Fourth..." where each numbered paragraph follows an identical template: claim → evidence → implication.

Why it triggers: The ordinal+template pattern is statistically rare in human academic writing. Humans tend to weave contributions into argument, vary paragraph openings, and let some contributions emerge from discussion of others rather than presenting them as an enumerated list.

Banned: "The HELD model extends ELD in four ways. First, ... Second, ... Third, ... Fourth, ..."

Fix: Give each contribution its own paragraph with a distinct opening. Vary structure — one paragraph might lead with evidence, another with a theoretical claim, another with a direct challenge to existing work. Let the reader count; do not count for them.


12. Enumerated Condition Lists

Boundary conditions, limitations, or criteria presented as (a)... (b)... (c)... (d)... within a single sentence or paragraph, each with parallel syntactic structure.

Why it triggers: Perfect parallel enumeration within running prose reads as generated. Human writers tend to expand the most important condition and compress the less important ones, creating asymmetric treatment.

Banned: "conditions include: (a) temporal overload exceeding X; (b) domain competition for Y; (c) identity contradiction wherein Z; (d) insufficient recovery intervals"

Fix: Write the conditions as prose with varying depth. Expand the condition that matters most. Use "First,... Second,..." only if the structures are genuinely different. Or break into short separate paragraphs.


13. Balanced Binary Constructions

Rhetorical questions or framing sentences with perfectly symmetric alternatives: "Is it X or Y?" where X and Y mirror each other syntactically.

Why it triggers: Natural language rarely produces perfect syntactic mirrors. Humans instinctively make one side heavier, more specific, or structurally different from the other.

Banned: "Is it a burden to be minimized or a generative resource to be nurtured?"

Fix: Make it asymmetric. One side gets more words, more specificity, or a different syntactic structure: "Does it function as the strain that existing literature assumes — or does it serve a purpose that framing has obscured?"


14. Template Method Prose

Method sections that follow a predictable template: "This study adopts a [paradigm] research design employing [method] (Author, Year). Data collection integrates [technique] within [setting] (duration). Following [instrument 1], participants engaged with [instrument 2]."

Why it triggers: Method sections are the easiest target for AI detectors because the genre is already formulaic — but AI makes it more formulaic by smoothing out the specific decisions researchers actually make. Human method sections contain idiosyncratic justifications and concrete details that break the template.

Fix: Use active voice ("We adopted..."). Name specific tools (MacWhisper, not "automatic speech recognition software"). Include a concrete, unexpected detail that breaks the template ("Sessions combined three instruments: a Narrative Career Timeline, a Maker Mix exercise in which participants rated Money, Impact, and Expression for each domain, and a Wheel of Life covering eight life areas"). Show the decision, not just the procedure.


15. Zero-Friction Paragraph Flow

Every sentence connects perfectly to the next — no abrupt starts, no register breaks, no sentence that makes the reader pause. The paragraph reads like a conveyor belt.

Why it triggers: AI detectors measure "perplexity" — how predictable each next word is given the preceding context. Perfect flow = low perplexity = high AI probability. Human academic writing has micro-disruptions: a short sentence after a long one, an unexpected concrete example, a parenthetical that breaks the main clause, a sentence starting with "But" or "The point is."

Fix: After drafting, read each paragraph and ask: is there at least one moment of friction — a sentence that slightly surprises? If not, insert one. Start a sentence with a conjunction. Drop in a concrete detail. Use a dash to interrupt a clause. Let a two-word sentence stand alone.


16. Smooth Qualitative Interpretation

Paragraphs that present a participant quote and then interpret it with complete confidence and no hesitation. Pattern: "[Participant] described X: '[quote]'. This [smooth interpretation]. [Theoretical implication]."

Why it triggers: Human qualitative researchers show interpretive tentativeness — "what seems to be happening here," "this suggests," or explicit alternative readings. AI generates interpretation with the same confidence as description, producing zero-uncertainty analytical prose.

Fix: Not every quote needs hedging — but a section with 5 consecutive quote-interpretation pairs, each with the same confidence level, will flag. Vary: expand one interpretation, compress another, add a qualifying observation to a third ("though P7 later complicated this picture"), let one quote speak for itself.


17. Single-Sentence Definition Propositions

Formal propositions, definitions, or conceptual claims packed into a single grammatically complex sentence spanning 40–60 words with multiple embedded clauses.

Why it triggers: Humans break these. A 50-word single sentence that is grammatically flawless reads as generated because humans lose track of clause nesting and either split the sentence or produce a grammatical imperfection.

Fix: Split into two sentences. Make the first sentence state the claim. Make the second add the mechanism or elaboration. "Hybrid entrepreneurship constitutes a naturally occurring form of ELD. Individuals who maintain both employment and venture use the dual portfolio to construct whole-life architecture through 4D Looping — iterating between self-discovery, design, feedback, and implementation."


Known False Positives (Do Not Fix)

Some passages will always score high on AI detectors regardless of writing quality:

  • Author affiliations and institutional boilerplate — formulaic by definition
  • Direct quotations — especially translated quotes with standardized translation brackets
  • Ethics approval statements — standardized language required by IRBs
  • Table notes and figure captions — genre-constrained
  • Reference lists — APA format is a template

These are not problems to solve. Acknowledge them if the user is alarmed, but do not rewrite standardized passages to evade detection.


AI Detection Self-Check (Post-Draft)

After completing any academic text, perform this paragraph-level scan:

  1. Enumeration check: Does any section contain "First... Second... Third..." or (a)... (b)... (c)... with parallel templates? → Break the pattern.
  2. Sentence length variance: Sample 5 consecutive sentences. If all fall within ±3 words of each other → vary deliberately.
  3. Paragraph opening diversity: List the first 3 words of each paragraph in a section. If more than 2 share the same structure → rewrite openers.
  4. Friction check: For each paragraph, identify the moment of micro-disruption. If there is none → insert one.
  5. Confidence uniformity: In findings/results, check whether every quote-interpretation pair has the same certainty level. → Vary.
  6. Binary balance: Check all "X or Y?" constructions for perfect syntactic symmetry. → Make asymmetric.

Integration with Other Writing Skills

This skill runs as a mandatory final-pass filter on all output from:

  • academic-writing — standard prohibitions + Academic Module
  • theory-section-architect — standard prohibitions + Academic Module
  • argument-architecture — standard prohibitions + Academic Module
  • research-examiner — Academic Module when reviewing drafts
  • cultural-content-architect — standard prohibitions; grandeur is antithetical to "Früh sehen. Leicht zeigen"
  • reel-architect — standard prohibitions; marketing grandeur destroys Reel credibility
  • consulting-document-synthesis — standard prohibitions; executive readers have low tolerance for LLM-smooth prose

Diagnostic: Would a careful human editor who has reviewed 10,000 Wikipedia AI-cleanup cases flag this sentence? If yes — rewrite.


Stylistic Fingerprint Module

Derived from Bao & Stern (2025, Cornell) on function-word attribution, Tian & Mitchell (2023) on few-shot style fidelity, and Liang et al. (2024) on post-ChatGPT vocabulary shifts in academic writing. Source synthesis: every.to, "The Science of Why AI Still Can't Write Like You" (2025).

Why This Module Exists

Sections 1–17 tell you what NOT to do (avoid AI tells). This module addresses the complementary question: what makes writing recognizably human? The research shows it is not vocabulary, argument structure, or even prose quality — it is subconscious word choices, particularly function words, that carry the strongest stylistic fingerprint.

Research Findings (Hard Data)

  1. Function words are the fingerprint. Cornell researchers (Bao & Stern, 2025) found that removing common function words — articles ("a," "the"), pronouns ("I," "she"), prepositions, conjunctions — from a text makes it significantly harder for LLMs to correctly identify who wrote it. These are the words writers choose without thinking. They are also the words LLMs homogenize most aggressively during post-training alignment.
  2. Few-shot examples are 23× more effective than prompting. Providing an LLM with just a few writing samples of a target author increases style fidelity approximately 23× over base model output (Tian & Mitchell, 2023). Implication: "Write in a human style" as instruction is nearly worthless. Actual samples of the target author's writing are what moves the needle.
  3. LLMs are already reshaping human writing. Post-ChatGPT academic papers show measurable vocabulary contamination. The word "delve" alone increased in usage by over 2,000% between 2022 and 2024 (Liang et al., 2024). Other markers: "crucial," "multifaceted," "comprehensive," "noteworthy." This is not just AI writing — it is human writing infected by AI exposure.
  4. Post-training alignment destroys distinctiveness. RLHF and other alignment procedures push LLM output toward "generic politeness" (Ted Chiang's "blurry JPEG of the web"). The model can produce text that is grammatically correct and topically relevant but stylistically anonymous — no writer's fingerprint remains.

Operational Rules

Rule 18: Preserve Function-Word Idiosyncrasy

LLMs default to generic function-word patterns: "the" where a human might use "this" or "that" or nothing; "it is" where a human might use "this is" or drop the construction entirely; "however" where a human might use "but" or just start the next sentence.

  • Do not standardize function words across a text. If the writer uses "but" in one place and "however" in another, that asymmetry is signal, not noise.
  • Do not add articles where a human writer dropped them. German-influenced English writers often omit articles in ways that are technically imperfect but stylistically distinctive.
  • Do not normalize pronoun patterns. Some writers use "I" heavily; others avoid it. Both are fingerprints.

Rule 19: Avoid Vocabulary Contamination

The 2,000% "delve" spike is not an isolated case. Monitor for vocabulary that was rare in academic writing pre-2023 but is now common due to LLM exposure:

High-risk contamination words (academic context):

  • "delve" / "delves into"
  • "crucial" (replacing "important" or "central")
  • "multifaceted"
  • "noteworthy"
  • "comprehensive" (as reflexive descriptor)
  • "foster" / "fostering"
  • "underpin" / "underpinning"
  • "intricacies"
  • "realm"
  • "synergy" / "synergies"
  • "facilitate" (when "enable" or "allow" works)
  • "leverage" (as verb, outside finance)

These words are not banned — they existed before ChatGPT. But if three or more appear in the same section, the text will pattern-match to LLM output regardless of who wrote it. Treat clusters as a red flag.

Rule 20: Use Writing Samples, Not Style Instructions

When generating text that should match a specific author's voice (e.g., the user's academic prose, a co-author's draft style):

  • Request or locate 2-3 actual writing samples from that author
  • Use them as stylistic reference, not just content reference
  • Pay attention to: sentence-opening patterns, function-word preferences, comma habits, paragraph length tendencies, the ratio of simple to complex sentences
  • A 500-word sample teaches more about voice than any style instruction

This follows directly from the 23× finding: examples beat descriptions.

Fingerprint Self-Check (Post-Draft)

After completing any text, add this to the existing AI Detection Self-Check:

  1. Function-word scan: Read the first word of each sentence in a paragraph. If more than half start with "The" or "This" → vary. Check for pronoun monotony ("it... it... it..." or "they... they... they...").
  2. Contamination cluster check: Ctrl+F for the high-risk contamination words above. If 3+ appear within 500 words → replace at least two with the writer's natural vocabulary.
  3. Generic politeness check: Does any sentence read as though it could appear in ANY text about ANY topic? ("This represents a significant contribution to the field.") → Replace with a sentence that could only appear in THIS text about THIS topic.

The Underlying Diagnostic

AI writing optimizes for appearing comprehensive, balanced, and significant. High-quality human writing shows the trace of a thinking person: asymmetries, commitments, specific details, the willingness to say one thing clearly rather than three things vaguely.

The function-word research adds a layer: even when AI avoids obvious tells, its output lacks the unconscious stylistic fingerprint that makes writing recognizably authored. The solution is not more rules but more specificity — writing that could only have been written by someone who was actually thinking about this particular thing, in this particular way, at this particular moment. That is what this skill protects.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.