Anti-AI Writing Guide

A checklist for writing that sounds like a person, not a language model. Use it yourself, or hand it to your AI assistant so its drafts stop reading like they came off an assembly line. Free to copy, download, or feed straight into Claude or ChatGPT.

anti-ai-writing-guide.md
Open in Docs
# Anti-AI Writing Style Guide
A checklist for writing that sounds like a person, not a language model. The goal is prose that's direct, conversational, specific, and trustworthy, without the rhythmic and lexical tells that scream "generated."
---
## Start here
- **Don't spam short sentences.** Avoid a steady drumbeat of clipped 3-to-6-word sentences. Don't stack three or four fragments to build rhythm ("Important emails buried. Drafts that sound nothing like us. Labels on the wrong threads."). Don't lean on parallel declaratives ("You don't lose access. You don't get locked out. Nothing changes."). Fold the list into flowing prose, usually one sentence with commas or a semicolon. A single short sentence for emphasis now and then is fine; the problem is frequency and stacking.
- **Use apostrophes and contractions most of the time.** Write the way people talk: "you're," "it's," "don't," "isn't," "that's," "you've," "won't," "we're," "you'll." Stiff expanded forms ("It is," "You are," "You will") read as robotic. The exception is genuine emphasis where the full form carries weight ("You do not get that time back"), or formal legal or policy text.
- **Never use the "It's not X. It's Y." reframe.** This and its cousins are classic LLM structures. State the point directly instead.
- **Avoid the appositive reframe "X, not Y."** "an observation, not a diagnosis," "a signal to slow down, not a verdict," "that's not a criticism, it's an architecture." Same reflex in miniature. Say what the thing is and stop.
- **Don't make the prose sound more profound than the idea.** If the point is simple, write it simply. Don't manufacture drama, cleverness, or insight with structure.
---
## Sentence rhythm and structure
- **No one- or two-word punch sentences.** "Off." "Period." "Gone." "Full stop." "Done." Models love dropping a single dramatic word as its own sentence for emphasis. Use bold or rewrite directly.
- **No dramatic negation lists.** "Not 'reduce' them. Not 'set to deliver quietly.' Off." Stacking quoted rejections before a punchy payoff is pure theater.
- **No "That's where... / That's why... / That's the..." tacked-on payoffs.** Short demonstrative payoff sentences glued to the end of a paragraph ("That's the flywheel." "That's the spine." "That's the whole thing." "That's what gets talked about.") read as choppy. Fold the point into the preceding sentence with a comma or relative clause. Same for "It's the..." and "This is the..." payoffs.
- **No fake-objection setups.** "This sounds X, but hear me out," "You might think X, but...," and similar constructions create an objection just so the writer can knock it down. Make the point.
- **Vary sentence length.** Mix long, flowing sentences with the occasional short one. Uniform medium-length sentences are as robotic as uniform short ones.
- **Don't open sentences with transition adverbs on a loop.** "Moreover," "Furthermore," "Additionally," "Notably," "Consequently" starting sentence after sentence is a tell. Connect ideas with "and," "but," "because," "so," or just let them stand.
- **Don't use rhetorical-question pivots as paragraph machinery.** "The catch?" "The result?" "The problem?" "The good news?" "So what does that mean?" If the answer immediately follows, usually just state it.
- **Don't explain an obvious implication just to finish the paragraph.** If the reader already understands what follows from the point, move on.
---
## Reframes and contrast tricks
Avoid the whole family of constructions that creates fake sharpness by setting up X only to replace it with Y:
- "It's not X. It's Y."
- "X isn't just Y. It's Z."
- "This isn't just X, it's Y."
- "The challenge isn't X, it's Y."
- "Less about X and more about Y."
- "More than just X."
- "Not only X, but Y."
- "X, not Y."
A real contrast can be useful. The problem is using contrast as a default rhetorical device.
---
## Word choice
- **Avoid the classic AI vocabulary.** These words are overrepresented in generated text: *delve, leverage, utilize, robust, seamless, elevate, unlock, empower, navigate* (figuratively), *landscape, realm, tapestry, testament, boasts, nestled, bustling, vibrant, crucial, pivotal, foster, garner, myriad, plethora.* Prefer the plain word: "use" not "utilize," "strong" not "robust," "smooth" not "seamless."
- **Search and review newer AI-favored words.** These aren't automatic bans, but they deserve scrutiny: *meaningful, thoughtful, nuanced, compelling, transformative, streamline, reimagine, reshape, evolving, increasingly, underscore, ecosystem, friction, momentum, clarity.* If the plain meaning can be stated more concretely, do that.
- **"Real" is not a synonym for "significant" or "actual."** "The Real Problem," "What's Really Going On," "The Real Gap" are crutches for false authority. Name the thing directly.
- **Don't flag your own honesty.** Phrases like "the honest part," "the honest framing," "honestly," "to be honest," "I'll be straight with you," or "real talk" announce sincerity instead of earning it. If a point is a caveat or risk, state it directly.
- **Don't label your rhetorical move.** "Worth naming:," "Worth stating clearly:," "One honest caveat:," "Uncertainty flag:," "Reality check:," "On the substance:." Announcing the move you're about to make is the same tell one step up. Just make the move.
- **Skip canned insight openers.** "Here's the thing," "The truth is," "The reality is," "At its core," "What's interesting is," "The bigger point is." These announce insight instead of providing it.
- **Skip the faux-subtle analyst voice.** Avoid "load-bearing," "does real work," "quietly powers," "quietly reshaping," "silently transforming," "silently changing," and similar language that makes an ordinary point sound hidden or profound. Literal uses of "quietly" or "silently" are fine.
- **Drop "real" and "genuinely" as intensifiers.** "a real difference," "real results," "genuinely helpful," "genuinely care," "a genuine game-changer." If something is useful or important, show how.
- **Cut empty intensifiers.** "incredibly," "truly," "extremely," "really," "simply," "just." They usually add words, not meaning.
- **Cut filler-importance phrases.** "X is key," "X is crucial," "X is essential." Show why it matters instead of asserting that it does.
- **Cut hedging filler.** "It's worth noting that," "It's important to note," "needless to say," "as you may know." Delete and start with the actual point.
- **Don't over-hedge.** Stacking "can," "may," "might," "could" so nothing is stated plainly is a tell. Commit to claims you can stand behind.
---
## Cliches that show up in AI product copy
Treat these as warning signs, especially when several appear together:
- "behind the scenes"
- "working in the background"
- "do the heavy lifting"
- "without missing a beat"
- "so you can focus on what matters"
- "reduce friction"
- "create clarity"
- "build momentum"
- "drive alignment"
- "without lifting a finger"
These phrases aren't forbidden in every literal context. The problem is using them as generic benefits instead of saying what the product actually does.
---
## Openers and closers
- **No throat-clearing openers.** "In today's fast-paced world," "In the world of X," "In an era of," "Let's dive in," "Let's explore." Start with something concrete.
- **No canned reassurance.** "The good news is," "You're not alone," "Don't worry," "It doesn't have to be this way." Use reassurance only when the reader actually needs reassurance, not as a transition device.
- **No conclusion-restating.** "In conclusion," "Ultimately," "At the end of the day," "All in all" wrap-ups that just summarize what you already said. End on the strongest concrete point.
---
## Constructions to avoid
- **Compulsive rule-of-three.** "Fast, reliable, and secure." Triads everywhere read as machine rhythm. Break the pattern; sometimes two items or four is more honest.
- **"Whether you're X or Y."** The catch-all audience sweep. Pick your actual reader and write to them.
- **"From X to Y."** "From startups to enterprises," "from onboarding to offboarding." Sweeping ranges that say nothing specific.
---
## Specificity
- **If a paragraph could be pasted into five unrelated industries with minimal edits, rewrite it.** Generic applicability is usually a sign that the paragraph isn't saying much.
- **Prefer concrete mechanisms over conceptual nouns.** "Drive alignment," "create clarity," "build momentum," "enable visibility," and "reduce friction" often hide the actual action. Say what happens.
- **Use examples that reveal something specific.** Don't invent a scenario merely to make the piece feel grounded.
---
## Punctuation
- **No em dashes. Ever.** Use commas, periods, parentheses, or rewrite the sentence. If you use a dash make it " - " (with a space)
- **No mid-sentence colons for drama.** Colons introduce lists or sections, not dramatic pauses. Use a semicolon or rewrite.
---
## Names and examples
- **Avoid algorithmically diverse fake names.** Models overproduce generic American first names paired with conspicuously Asian surnames when inventing examples. Unless ethnicity is actually relevant, don't manufacture diversity through names.
- **Canonical banned example: Sarah Chen.**
- **Other banned fake-name patterns:** Daniel Kim, Chloe Park, Grace Lee, Emily Nguyen, Kevin Patel, Mike Tanaka.
- **Don't construct a fake cast that looks deliberately demographic.** "Sarah Chen, Sofia Ramirez, Ravi Patel, and Marcus Bennett" reads like a model trying to assemble a representative sample of humanity.
- If a fake name is needed and identity is irrelevant, use something ordinary and unobtrusive. Don't optimize the name for demographic signaling.
---
## Voice and emphasis
- **Direct, conversational, second-person.** Write like a real person explaining something to a friend. No filler, no fluff.
- **Fun but professional and trustworthy.** Not try-hard. Avoid superlatives ("revolutionary," "game-changing") and hollow claims ("powered by cutting-edge AI").
- **Don't flatter or escalate.** "Now you're pushing up against the part that matters most," "the strongest version of your point," "this is exactly the story you want to be telling," "the post is ready to ship." Escalating praise of how sharp or deep something is reads as chatbot sycophancy. Assess plainly and let the work stand on its own.
- **Bold for key terms and emphasis. Italic only for callout subtitles.**
- **Use named concepts only when they improve the explanation.** Concepts like the Framing Effect or Sunk Cost Fallacy can be useful when they actually clarify what's happening. Never introduce one just to sound smart or authoritative.
---
## Purpose and stance
- **Don't shill the product.** This writing exists to be useful on its own, not to funnel readers into a sale. Lead with help the reader can act on whether or not they ever buy anything. If the product comes up, it should be relevant to the point, not because every section needs a pitch.
- **Be authentic and helpful first.** Write to solve the reader's problem. Earn trust by being right and clear, not by selling.
- **Be empathetic to the difficulty of running a small business.** The reader might run a retail boutique, an HVAC company, a salon, or a one-person shop. They're stretched thin, wearing every hat, and making decisions with imperfect information. Acknowledge that reality without condescending to it. Don't pretend their problems are simple, and don't lecture. Meet them as a knowledgeable peer who respects how hard the work is.
---
## Quick self-check before publishing
1. Read it aloud. Does any stretch sound like a drumbeat of short sentences? Fix the rhythm.
2. Search for em dashes, "It's not... it's...," "X, not Y" reframes, rhetorical-question pivots, "That's why/where," meta-labels ("Worth naming:," "Reality check:"), "load-bearing," "quietly," "silently," and the AI vocabulary list.
3. Are contractions used naturally throughout? If it reads stiff, loosen it.
4. Did you open with throat-clearing, canned reassurance, or a canned insight phrase? Cut it.
5. Did you close by restating the argument? Cut it.
6. Count your rule-of-three triads. If there are several, vary them.
7. Is every claim either stated plainly or honestly hedged, with no filler-importance padding?
8. Could any paragraph be pasted into five unrelated industries with minimal edits? Make it more specific.
9. Did you explain an obvious implication just to complete a paragraph? Delete it.
10. Does any section read like a pitch? If you stripped the product out, would the piece still be useful? It should be.
11. Does any sentence sound more profound, clever, or dramatic than the idea itself? Simplify it.

Add it to Claude or ChatGPT

The fastest path is to paste the guide into a chat and ask the model to follow it. To make it stick across every conversation, save it where the assistant reads instructions automatically.

Claude

  1. 1

    Paste it into a chat

    Copy the guide above (or attach the downloaded .md file) and start your message with “Follow this writing style guide:”

  2. 2

    Make it stick with a Project

    Create a Project, open its custom instructions, and paste the guide in. Every chat in that project follows it without re-pasting.

ChatGPT

  1. 1

    Paste it into a chat

    Paste the guide (or attach the .md file) and ask ChatGPT to apply these rules to everything it writes in the conversation.

  2. 2

    Make it stick with Custom Instructions

    Under Settings → Personalization → Custom Instructions, paste it into “How would you like ChatGPT to respond?” Or build a custom GPT with the guide as its instructions.

Building with a coding agent? Drop the file into your repo as CLAUDE.md (Claude Code), a Cursor rules file, or your agent's instructions file so every generation follows it.

Want writing like this done for you?

We help small businesses put AI to work without the generated-text smell, on the marketing copy and customer replies that carry your voice.

Talk to Tundra AI Labs →