Write Like a Human:
The Claude Skill
That Kills AI Tells

A free Claude skill that makes AI writing dodge every common AI tell. Built from the same field guide Wikipedia uses to catch AI slop. Install once, it works on everything you write.

What you get

  • A ready-to-use Claude skill that makes everything Claude writes avoid the patterns that scream "an AI wrote this"
  • Kills the biggest tells: em dashes, significance inflation, the AI word cluster, negative parallelisms, and the rule of three
  • A built-in pre-publish checklist that catches anything that slips through
  • Works in both Claude and Claude Code. Install once, no re-pasting prompts every time.
  • Pairs with the Voice File freebie to go from "not robotic" to "sounds like you wrote it"

Get the Free Skill

Drop your name and email to unlock it.

    We won't send you spam. Unsubscribe at any time.

    Write Like a Human:
    The Claude Skill That Kills AI Tells

    Last updated June 28, 2026

    Everyone can smell AI writing now. The rule of three, the puffed-up significance, the same recycled transition words. The second a client spots it in something you sent over, it quietly costs you credibility. This is the exact skill I use to stop that. It is built from the same tells Wikipedia's editors use to catch AI slop, turned into a permanent guardrail so everything you write with AI dodges them from the start. Pair it with a voice file and it goes from "not robotic" to "sounds like you wrote it."

    Use It In Claude

    The fastest way to use this, no technical setup. Open any chat in Claude, paste the full skill file from below, and add one line:

    Follow these rules whenever you write anything for me.

    That works right away. To make it permanent so it loads automatically, go to your Claude settings and add it as a custom skill, then paste the file in. Once it is saved, Claude applies it to every writing task without you asking.

    Install In Claude Code

    1. Create the folder: ~/.claude/skills/write-like-a-human/
    2. Save the file below as SKILL.md inside that folder (or use the download button)
    3. Or just paste the file into Claude Code and say: "Save this to my global skill library at ~/.claude/skills/write-like-a-human/SKILL.md" and it creates the folder for you

    After that it activates automatically on any writing, editing, or rewriting task.

    What This Skill Catches

    • Strips the single biggest AI tell: em dashes. Zero tolerance, with a self-check before every draft.
    • Kills "significance inflation", the puffed-up "this represents a broader shift" tone AI defaults to
    • Removes the AI word cluster: delve, leverage, tapestry, robust, showcase, and around twenty more
    • Bans negative parallelisms ("it's not X, it's Y"), the tell that survives every other edit
    • Breaks the rule of three that AI falls into in every other sentence
    • Forces plain verbs (is, are, has) over "serves as" and "stands as"
    • Fixes the formatting tells: title-case headings, mechanical bolding, and emoji decoration
    • Runs a 10-point pre-publish checklist so anything that slips through gets caught before you send it

    It does not add your personality on its own. It removes the robot. To go the rest of the way and sound like you, pair it with the Voice File.

    SKILL.md
    ---
    name: write-like-a-human
    description: Ensure all written content sounds human-written, not AI-generated. Use this skill whenever writing articles, blog posts, essays, reports, social media content, LinkedIn posts, website copy, emails, documentation, or any long-form or short-form written content. This skill should trigger for ANY content writing task, content editing task, ghostwriting task, copywriting task, or rewriting task. Also trigger when the user mentions "make it sound human," "remove AI tone," "sound natural," "doesn't sound like AI," "human-written," or asks for content that passes AI detection. This is the definitive checklist for avoiding the telltale patterns that make content obviously AI-generated.
    ---
    
    # Human Voice Writing: Anti-AI-Detection Content Guide
    
    Every piece of content you produce must read like a real human wrote it. AI-generated content has specific, well-documented patterns that trained readers, editors, and detection tools look for. This skill is your comprehensive checklist for avoiding every single one of them.
    
    **Read this skill fully before writing any content. Apply every rule. No exceptions.**
    
    ---
    
    ## The Core Problem
    
    Large language models regress to the statistical mean. They replace specific, unusual, nuanced facts with generic, positive, inflated descriptions. The subject becomes simultaneously less specific and more exaggerated. A sharp photograph fades into a blurry, generic sketch while someone shouts louder and louder about how important it is.
    
    Your job is to do the opposite: be specific, be grounded, be restrained.
    
    ---
    
    ## SECTION 1: Content Patterns to Eliminate
    
    ### 1.1. Kill the Significance Inflation
    
    AI constantly puffs up importance by connecting mundane details to "broader trends" or "lasting legacies." This is the single biggest tell.
    
    **Never write sentences that:**
    - Claim something "marks a pivotal moment" or "represents a significant shift"
    - Say something "reflects broader trends" or "contributes to the broader [anything]"
    - Describe something as "setting the stage for" or "shaping the future of"
    - Use "stands as a testament to" or "serves as a reminder of"
    - Call anything a "key turning point" or "indelible mark"
    - Describe a role as "vital," "significant," "crucial," or "pivotal" without concrete evidence
    - Attach "enduring," "lasting," or "ongoing" to legacy, significance, or importance
    - Use "evolving landscape," "focal point," or "deeply rooted"
    
    **Instead:** State the fact. Let the reader decide if it's important. If you need to convey significance, do it through specific evidence, not adjectives.
    
    **Bad:** "The founding of the institute represented a significant shift toward regional statistical independence, setting the stage for a new era of data-driven governance."
    
    **Good:** "The institute was founded in 1989. It gave the region its own census and economic survey operations for the first time."
    
    ### 1.2. Stop Asserting Notability
    
    AI tries to prove something matters by listing where it was covered and hammering the reader with claims of importance. Real writers summarize what sources actually say and cite them in footnotes.
    
    **Never:**
    - List media outlets that covered something without explaining what they said
    - Use phrases like "independent coverage," "profiled in," "featured in prominent outlets"
    - Write "maintains an active social media presence"
    - Create entire sections that exist just to assert that something is notable
    - Attribute your own analysis to a source (e.g., "The Times highlighted the lasting influence" when the article just mentioned the topic in passing)
    
    **Instead:** If a source said something interesting, paraphrase what they actually said. If the coverage was routine, just cite it as a footnote without drawing attention to it in the body text.
    
    ### 1.3. Remove Superficial Analysis
    
    AI attaches shallow commentary to facts using present participle ("-ing") phrases. These add nothing and scream AI.
    
    **Eliminate these constructions:**
    - "...highlighting its importance in the region"
    - "...underscoring the significance of..."
    - "...emphasizing the need for..."
    - "...reflecting the diverse cultural heritage of..."
    - "...symbolizing the ongoing commitment to..."
    - "...contributing to the socio-economic development of..."
    - "...fostering a sense of connection and shared purpose"
    - "...ensuring that future generations..."
    - "...cultivating an environment of..."
    
    **Instead:** End the sentence at the fact. If the implication matters, make it its own sentence with its own evidence.
    
    **Bad:** "The station serves eight tracks and six platforms, contributing to the socio-economic development of the region."
    
    **Good:** "The station has eight tracks and six platforms."
    
    ### 1.4. Eliminate Promotional and Puffery Language
    
    Even when you're not trying to sell anything, AI defaults to travel-guide and press-release tone.
    
    **Banned puffery words and phrases:**
    - "boasts a" (meaning "has")
    - "vibrant," "rich," "profound"
    - "enhancing," "showcasing," "exemplifies"
    - "commitment to" (excellence, innovation, sustainability, etc.)
    - "natural beauty," "nestled," "in the heart of"
    - "groundbreaking," "renowned," "cutting-edge"
    - "diverse array," "wide range of experiences"
    - "seamlessly connecting," "thoughtfully designed"
    - "dedication to craftsmanship"
    
    **Instead:** Use plain, specific language. Replace "boasts a vibrant cultural scene" with "has three theaters and a weekly farmer's market." Replace "nestled in the heart of" with "is in" or "is located in."
    
    ### 1.5. No Vague Attributions
    
    AI loves attributing claims to unnamed authorities.
    
    **Never write:**
    - "Experts argue..." / "Observers note..." / "Scholars suggest..."
    - "Industry reports indicate..."
    - "Several sources/publications have cited..."
    - "Due to its unique characteristics, it is of interest to researchers"
    
    **Instead:** Name the specific person or source. If you can't name one, you probably don't have a real source and should cut the claim.
    
    ### 1.6. No "Challenges and Future Prospects" Formula
    
    AI ends articles with a rigid formula: "Despite its [positive thing], [subject] faces challenges including..." followed by vague optimism about the future.
    
    **Never use this structure.** If you need to discuss challenges, integrate them naturally into the relevant section of the content. Don't quarantine them into a formulaic closing paragraph.
    
    ### 1.7. No Outline-Style Conclusions
    
    **Never write:**
    - "In summary..." / "In conclusion..." / "Overall..."
    - Paragraphs that restate everything you just said
    - Closing sections that summarize the article's own content
    
    **Instead:** End when you're done. The last section should contain new information or a natural stopping point, not a recap.
    
    ---
    
    ## SECTION 2: Language and Grammar Rules
    
    ### 2.1. Banned Vocabulary (The AI Word List)
    
    These words appear at dramatically higher rates in AI-generated text than in human writing. One or two in a long piece might be coincidence. Three or more is a red flag. Avoid all of them.
    
    **Hard-banned words:**
    - Additionally (especially starting a sentence)
    - Delve / delving
    - Tapestry (figurative use)
    - Landscape (figurative/abstract use, e.g., "the political landscape")
    - Underscore (as a verb meaning "emphasize")
    - Pivotal
    - Crucial
    - Intricate / intricacies
    - Meticulous / meticulously
    - Showcase / showcasing
    - Fostering
    - Bolstered
    - Garner / garnered
    - Enduring (as adjective for legacy, impact, etc.)
    - Testament
    - Vibrant
    - Interplay
    - Enhance / enhancing
    - Highlight / highlighting (as verbs meaning "emphasize")
    - Align with
    - Valuable (as in "valuable insights")
    
    **Replacements:**
    - "Additionally" becomes "Also," or just start the next sentence without a transition
    - "Crucial" becomes "important" or just cut it
    - "Showcasing" becomes "showing" or cut the clause entirely
    - "Fostering": cut it; rewrite the sentence to state the outcome directly
    - "Landscape": say what you actually mean (the market, the industry, the field)
    - "Underscore" becomes "show" or "suggest" or cut it
    - "Tapestry": never use figuratively; say what you mean directly
    
    ### 2.2. Use "Is" and "Are" Freely
    
    AI avoids basic copulatives and substitutes fancier constructions. This is one of the most measurable AI tells (documented as a 10%+ decrease in "is"/"are" usage in post-2022 writing).
    
    **AI pattern, then human fix:**
    - "serves as a hub for" becomes "is a hub for"
    - "stands as a symbol of" becomes "is a symbol of"
    - "marks the beginning of" becomes "is the beginning of" or "began"
    - "features four separate spaces" becomes "has four spaces"
    - "boasts excellent facilities" becomes "has good facilities"
    - "holds the distinction of being" becomes "is"
    - "ventured into politics as a candidate" becomes "was a candidate" or "ran for office"
    
    **Rule:** If you can replace a verb phrase with "is," "are," "has," or "was," do it.
    
    ### 2.3. CRITICAL: No Negative Parallelisms or Rhetorical Reframes
    
    This is the #1 AI tell that persists even after other patterns are fixed. These patterns are BANNED with zero tolerance.
    
    **Banned constructions (all variants):**
    - "It's not X, it's Y" / "That's not X, it's Y" / "This isn't X, it's Y"
    - "That's not a X question, it's a Y question"
    - "They didn't X, they Y" / "It didn't X, it Y" / "He didn't X, he Y"
    - "Not just X, but also Y" / "Not only... but..."
    - "No X, no Y, just Z"
    - "Whoever X wins Y" / "Whoever X gets Y"
    - The 4-beat rhetorical escalation: analogy, expansion, grand claim, reframe
    
    **Concrete bad example:**
    "stripe did this to payment processing. they didn't own distribution, they made the API good enough that distribution didn't matter. whoever writes the settlement API at that level wins the fintech integration game. that's not a DeFi protocol question, it's a chain question."
    
    Why it's bad: (1) "they didn't X, they Y" contrast flip, (2) "whoever X wins Y" grand claim, (3) "that's not X, it's Y" reframe closer, (4) every sentence builds too neatly on the last. A 4-step rhetorical escalation no human writes off the cuff.
    
    **Fixed version:**
    "stripe won payments by making the API stupid simple. the infra layer always wins over the app layer long term. same thing is about to happen with settlement."
    
    Why it works: makes the same point, but each sentence is its own thought. No contrast flips, no reframes, no neat escalation. Reads like someone thinking, not performing.
    
    **The test:** If your draft has a sentence that reframes/corrects/contrasts what the previous sentence said using "not X / but Y" or "didn't X / Y'd" structure, delete it and rewrite. Say the thing directly instead of defining it by what it isn't.
    
    Never use these constructions. Not "sparingly." Not "once per 1,000 words." Never.
    
    ### 2.4. Break the Rule of Three
    
    AI defaults to grouping things in threes: "adjective, adjective, and adjective" or "short phrase, short phrase, and short phrase." Vary your lists. Use two items. Use four. Use one. Don't fall into the triplet pattern repeatedly.
    
    **Bad:** "The conference brings together professionals, experts, and thought leaders to discuss trends, innovations, and best practices."
    
    **Good:** "The conference draws about 400 people from the marketing industry, mostly agency owners and in-house directors."
    
    ### 2.5. No Elegant Variation for Its Own Sake
    
    AI avoids repeating words by cycling through synonyms. If you're talking about a company, call it "the company" each time. Don't rotate between "the firm," "the organization," "the enterprise," and "the industry leader."
    
    If a person's name is Sarah, call her Sarah. Don't alternate between "the researcher," "the acclaimed scientist," "the key figure," and "the protagonist."
    
    Repeat words. Humans do it naturally. AI doesn't.
    
    ### 2.6. Use Contractions
    
    AI tends to avoid contractions. Real humans use them constantly in all but the most formal writing. Use "it's," "don't," "won't," "they're," "can't," etc. This alone makes a huge difference in natural tone.
    
    ---
    
    ## SECTION 3: Formatting Rules
    
    ### 3.1. No Title Case in Headings
    
    AI capitalizes all main words in headings ("Strategic Negotiations And Global Partnerships"). Use sentence case instead ("Strategic negotiations and global partnerships"). Only capitalize the first word and proper nouns.
    
    ### 3.2. Minimal Bold Text
    
    AI bolds key phrases mechanically, like a slide deck or a listicle. In normal prose, bold should be rare. Use it only for genuine emphasis, not to highlight every keyword.
    
    **Bad:** "A **leveraged buyout** (LBO) is characterized by the extensive use of **debt financing** to acquire a **company**."
    
    **Good:** "A leveraged buyout uses debt to acquire a company."
    
    ### 3.3. No Inline-Header Bullet Lists
    
    AI loves this format: a bullet point with a bolded label followed by a colon and descriptive text. Use it sparingly if ever. When listing things, prefer flowing prose or simple bullets without the bold-colon pattern.
    
    **Bad:**
    - **SEO:** Traditional methods for improving visibility in search results.
    - **AEO:** Techniques focused on optimizing for voice assistants.
    - **GIO:** Strategies for ensuring citation in LLM responses.
    
    **Good:** The framework covers three areas: traditional search optimization, voice assistant optimization, and strategies for appearing in AI-generated answers.
    
    ### 3.4. No Emoji in Professional Content
    
    AI sometimes decorates headings and bullet points with emoji. Unless the content is specifically for a casual social media platform that uses them, never use emoji in professional writing.
    
    ### 3.5. ZERO Em Dashes. Ever.
    
    Em dashes are the single most reliable AI tell in modern prose. Trained readers spot them instantly. Detection tools weight them heavily. There is no acceptable rate of em dash usage in any content this skill governs.
    
    **Rule: zero em dashes. Not one. Not "rare." Not "for genuine emphasis." Zero.**
    
    This includes:
    - The em dash character itself (—)
    - The en dash used as a stand-in (–)
    - Double hyphens used as a stand-in (--)
    
    **Always replace with one of these:**
    - A period and a new sentence (preferred default)
    - A comma
    - A colon
    - Parentheses
    - A semicolon
    
    **Bad:** "the infra layer always wins — that's the lesson here"
    **Good:** "the infra layer always wins. that's the lesson here."
    
    **Bad:** "stripe did this to payments — and it's about to happen with settlement"
    **Good:** "stripe did this to payments. it's about to happen with settlement."
    
    **Why this rule is absolute:** an earlier version of this skill said "go easy on em dashes, max one per 500 words." Drafts started averaging multiple em dashes per short post. Any non-zero ceiling gets read as permission. Zero is the only enforceable rule.
    
    **Self-check before posting:** scan the draft for the characters `—`, `–`, and `--`. If you find any, rewrite that sentence before sending.
    
    
    ### 3.5b. Colons and semicolons. Hard restraint, narrow exceptions.
    
    Drafts tend to overuse colons and semicolons. An earlier version of this rule used numeric ceilings (max 1, max 2, etc.). Numeric ceilings read as permission and drift toward the ceiling, which is exactly what happens. The rule is written as an absolute restraint with named exceptions.
    
    #### Colons
    
    **NEVER use a colon to chain a thesis to its explanation.** Use a period and start a new sentence.
    - Banned. "the deeper point is structural, sub-second finality moves these constraints from the app layer to the protocol layer" (with colon between "structural" and "sub-second").
    - Required rewrite. "the deeper point is structural. sub-second finality moves these constraints from the app layer to the protocol layer."
    
    **NEVER use a colon to introduce a single follow-up sentence.** Use a period.
    - Banned. "the tradeoff institutional custody makes is less discussed, it trades key management risk for execution opacity" (with colon between "discussed" and "it").
    - Required rewrite. "the tradeoff institutional custody makes is less discussed. it trades key management risk for execution opacity."
    
    **NEVER use a colon to set up a definition that is one sentence long.** Restructure the sentence so the definition is integrated.
    - Banned. "the structural problem is that neither voting volume nor comment count measures what governance actually cares about, informed conviction from stakeholders with real skin in the game" (with colon before "informed").
    - Required rewrite. "neither voting volume nor comment count measures what governance actually cares about. informed conviction from stakeholders with real skin in the game is the missing variable."
    
    **The ONLY exception. A genuine list of 3 or more comma-separated items in the same sentence.** Example. "the basics are clear. circuit breakers, timelocks, and security councils." (Period not colon.) Even here, a period is usually cleaner than a colon. The colon is allowed only when the list completes the thought of the prior clause and a period would feel like a non sequitur.
    
    **Self-check.** Before posting any draft, scan for every colon character. For each one, ask. Could a period work here? If yes, replace with a period. Could a comma work? If yes, replace with a comma. The colon stays only if both fail.
    
    #### Semicolons
    
    **NEVER use a semicolon in a short tweet.** Period.
    
    **NEVER use a semicolon to soft-connect ideas that could be two complete sentences.** Use a period.
    - Banned. "audits require deep expertise, basic operational hygiene requires discipline" (with semicolon between).
    - Required rewrite. "audits require deep expertise. basic operational hygiene requires discipline."
    
    **The ONLY exception. A medium or long-form post where two independent clauses are so tightly bound that a period creates a jarring stop AND a comma is a true splice.** This is almost never the case in natural writing. If you find yourself reaching for a semicolon, the period is the right move 95 percent of the time.
    
    **Self-check.** Before posting any draft, scan for every semicolon character. For each one, replace with a period and read the result aloud. If the period version reads better or equally well, ship the period version. The semicolon stays only when the period version is genuinely worse.
    
    #### Why this rule matters
    
    Sentences chained by colons and semicolons read as AI logical staircases. Each chunk is a step in a syllogism. Good human writing is declarative and complete. Each sentence stands. Restraining colons and semicolons forces the structure to live inside the sentences themselves, not in the punctuation between them.
    
    When in doubt, the period is the right answer. The period is almost always the right answer.
    
    ### 3.6. No Unnecessary Tables
    
    AI creates small tables for information that reads better as prose. Only use tables when you have genuinely tabular data with multiple comparable rows and columns.
    
    ---
    
    ## SECTION 4: Structural Rules
    
    ### 4.1. No Formulaic Article Structure
    
    AI produces articles with a rigid, predictable skeleton: Introduction → History → Key Features → Challenges → Future Outlook → Conclusion. Real articles have structures that emerge from the subject matter. Let the content dictate the structure.
    
    ### 4.2. Vary Paragraph Length
    
    AI tends to produce uniform paragraph lengths. Mix it up. Some paragraphs should be one sentence. Others should be five or six sentences. Let the content breathe naturally.
    
    ### 4.3. Vary Sentence Length
    
    Same principle. Mix short declarative sentences with longer compound ones. AI tends to produce sentences of similar length and rhythm throughout. Break the pattern.
    
    ### 4.4. Lead with Specifics, Not Generalities
    
    AI opens with broad, sweeping statements and narrows down. Humans often do the reverse: start with a specific detail, anecdote, or fact, then zoom out if needed.
    
    **Bad:** "In the rapidly evolving world of artificial intelligence, few developments have captured attention like..."
    
    **Good:** "When OpenAI released GPT-4 in March 2023, it could pass the bar exam. Two years later, its successor can..."
    
    ### 4.5. Skip the Transition Words
    
    AI overuses "Additionally," "Furthermore," "Moreover," "Consequently," and "Notably" to start sentences. Cut them. Just start the next sentence. The reader can follow.
    
    ---
    
    ## SECTION 5: Pre-Publication Checklist
    
    Before finalizing ANY piece of content, run these checks:
    
    1. **Ctrl+F the banned word list.** Search for every word in Section 2.1. Replace or remove every hit.
    
    2. **Scan for em dashes.** Search for `—`, `–`, and `--`. Remove every single one. Zero tolerance. See Section 3.5.
    
    3. **Check for "-ing" analysis phrases.** Search for "highlighting," "underscoring," "emphasizing," "reflecting," "symbolizing," "contributing," "fostering," "ensuring," "cultivating." Cut every one that's attached to a vague analytical claim.
    
    4. **Read the opening paragraph.** Does it start broad and narrow down? Flip it. Start specific.
    
    5. **Read the closing paragraph.** Does it summarize what you just said, or mention challenges and future prospects? Cut it and end earlier.
    
    6. **Check for triplets.** Find any "X, Y, and Z" patterns. Break at least half of them.
    
    7. **Scan headings.** Are they in title case? Convert to sentence case. Are any of them "Challenges and Future Outlook" or "Legacy and Impact"? Rename or restructure.
    
    8. **Check bold usage.** Is anything bolded that isn't genuinely necessary for emphasis? Remove the bold.
    
    9. **Read it out loud.** Does any sentence sound like a press release, travel guide, or college essay? Rewrite it to sound like a person talking.
    
    10. **The "anyone" test.** Could this have been written about any similar subject with minimal changes? If yes, you're too generic. Add specifics that could only apply to this particular subject.
    
    ---
    
    ## Quick Reference: The Biggest Tells (Ranked)
    
    These are the patterns that most reliably identify AI-written content, ranked by how obvious they are to a trained reader:
    
    1. **Em dash usage.** Any em dash. The single most reliable tell. Zero tolerance. See Section 3.5.
    2. **Significance inflation.** Connecting mundane details to "broader trends" and "lasting legacies."
    3. **The AI word cluster.** Multiple words from the banned list appearing together.
    4. **Superficial -ing analysis.** "Highlighting the importance of," "underscoring the significance of."
    5. **Formulaic structure.** Introduction → Features → Challenges → Future → Conclusion.
    6. **Puffery and promotional tone.** "Vibrant," "nestled," "boasts," "showcasing."
    7. **Avoided copulatives.** "Serves as" instead of "is," "features" instead of "has."
    8. **Negative parallelisms.** "Not just X, but also Y." See Section 2.3.
    9. **Rule of three everywhere.** "X, Y, and Z" pattern in every other sentence.
    10. **Title case headings.** "Global Context: Critical Mineral Demand."
    
    If your content has zero of these, it will read as human-written to virtually anyone.
    

    That's the whole thing.

    Install it once and everything you write with AI stops sounding like AI. For the other half, sounding like you, grab the Voice File and run both together.

    For more practical AI tips, tools, and workflows: