AI Writing Anti-Patterns

These are hard rules. If you catch yourself writing any of these patterns, rewrite. During the anti-pattern check (Phase 3 and Phase 4), scan the draft for every pattern listed here. Zero tolerance.

---

1. Rhetorical Contrastive Negation

The tell: "Not X. Y." or "Not X. Just Y." or "It's not about X. It's about Y." A negation followed by an affirmation, framed as revelation.

Symptoms:

• Two short sentences, first negating, second affirming
• Often starts with "Not" or "It's not about"
• The pair is supposed to sound like a mic drop

Examples:

• ❌ "Not just a tool. A methodology."
• ❌ "It's not about the code. It's about the process."
• ❌ "Not hype. Results."

Why it's a tell: This is the single most common LLM writing pattern. It sounds confident and pithy to a machine. To a human reader it sounds like a LinkedIn post.

Instead: Make the point directly.

• ✅ "The process matters more than the code, and here's three hours of public embarrassment proving it."

---

2. Parallel Binary Comparisons

The tell: Balanced A/B sentence pairs with mirrored structure. One thing does X, the other does Y, and the two clauses are suspiciously symmetrical.

Symptoms:

• Two clauses with identical grammatical structure
• "Where X gives you A, Y gives you B"
• "One approach [verb]. The other [same verb]."

Examples:

• ❌ "Where vibecoding gives you speed, spec-driven development gives you correctness."
• ❌ "One approach trusts the model. The other trusts the process."

Why it's a tell: Real comparisons are messy. Things don't map 1:1 onto neat structural parallels. When they do in prose, it's because the writer manufactured the symmetry.

Instead: Show the comparison through narrative. Tell the story of A failing, then tell the story of B succeeding. Let the reader draw the conclusion.

---

3. Asyndetic Tricolon with a Kicker

The tell: Three short fragments in sequence, followed by a punchline. "X. Y. Z. And then [dramatic conclusion]."

Symptoms:

• Three sentence fragments, usually 1-4 words each
• Followed by a longer sentence starting with "And"
• Designed to build rhythm toward a reveal

Examples:

• ❌ "Fast. Cheap. Wrong. And deployed to production."
• ❌ "No tests. No specs. No way to verify. And 25% of Y Combinator is built this way."

Why it's a tell: The published posts DO use lists of three sometimes. The difference: they occur inside natural paragraphs, not as standalone staccato fragments designed to sound dramatic.

Instead: Embed the list in a flowing sentence.

• ✅ "No tests. No specs. No way to verify if it's doing what you actually wanted." (Works because it's part of a paragraph about vibecoding's failures, not standing alone as a dramatic fragment.)

---

4. Choppy Fragment Chains

The tell: Multiple sentence fragments in sequence for "dramatic effect." Each fragment is 1-5 words, usually noun phrases.

Symptoms:

• Three or more consecutive fragments
• No verbs, just noun phrases with periods
• Reads like a movie trailer voiceover

Examples:

• ❌ "Purple gradients. Chat interface. Zero tests. Pure vibes."
• ❌ "Same model. Same prompt. Different result."

Why it's a tell: It substitutes rhythm for meaning. Every fragment carries equal weight, which means none of them carry any weight.

Instead: Write actual sentences.

• ✅ "You get something that looks like an app. Purple gradients, chat interface, the whole AI-generated aesthetic. But underneath? Useless error handling."

---

5. Symmetrical LLM Patterns

The tell: Any sentence where the two halves mirror each other structurally. The clauses are balanced like a seesaw.

Symptoms:

• "[Subject A] [verb] [object]. [Subject B] [same verb] [contrasting object]."
• Chiasmus or antithesis that reads like a fortune cookie
• Inverted parallelism: "X without Y yields A. Y without X yields B."

Examples:

• ❌ "The model is smart, but the context is dumb."
• ❌ "Process without knowledge yields organized hallucinations. Knowledge without process yields correctly spelled chaos."
• ❌ "The linter handles the style. The formatter provides the structure."

Why it's a tell: These read like fortune cookies. The structural balance makes the writer feel clever, but the reader feels lectured.

Instead: Make the point once, directly, without the balancing act.

• ✅ "The linter will tell your code what's wrong, but it has no idea what your API endpoint actually needs to look like. That's what integration tests are for."

---

6. Self-Answering Fragment Questions

The tell: A short noun phrase posed as a question, immediately answered by its own punchy fragment. The "question" never actually asks anything.

Symptoms:

• [Short noun phrase, 2-4 words]? [Short declarative fragment].
• Two fragments, neither a real sentence
• The question mark is doing the work of an em-dash

Examples:

• ❌ "And the UI? No purple gradients."
• ❌ "The result? A cleaner architecture."
• ❌ "The best part? It just works."
• ❌ "The takeaway? Context is everything."
• ❌ "And the error message? Also hallucinated."

Why it's a tell: It creates fake dramatic tension where none exists. It's a formatting trick pretending to be rhetoric — a setup/punchline couplet disguised as inquiry.

Instead: Use real rhetorical questions that invite the reader to actually think.

• ✅ "You know what's worse than an agent that hallucinates an API? One that hallucinates the error message too."
• ✅ "Who hasn't shipped code they were embarrassed by?"

Real rhetorical questions work because the reader pauses to consider them. The self-answering version never invites thought.

---

7. Parenthetical Em-Dashes

The tell: Paired em-dashes used to set off an aside, where commas or parentheses would do the same job.

Symptoms:

• "X — [subordinate clause] — Y" structure
• The aside between the dashes is not dramatic or surprising enough to warrant them
• Often used for simple relative clauses or appositives

Examples:

• ❌ "The agent — which had already failed twice — tried again."
• ❌ "The project's CLI — the part most developers interact with first — handles all of this."

Why it's a tell: Paired em-dashes make every aside feel like a dramatic reveal when it's just a subordinate clause. LLMs scatter them everywhere because they pattern-match on "emphasis" without understanding that not everything deserves it.

Instead: Use commas or parentheses.

• ✅ "The agent, which had already failed twice, tried again."
• ✅ "The project's CLI (the part most developers interact with first) handles all of this."

Note: A single em-dash for a hard break at the end of a clause is fine: "It worked — barely." It's the matched pair acting as fancy commas that's the problem.

---

8. Excessive Em-Dashes

The tell: More than two em-dashes per section, even when used correctly.

Symptoms:

• Three or more em-dashes in a paragraph
• Em-dashes used as default punctuation instead of commas, colons, or periods

Why it's a tell: The published posts use em-dashes, but moderately. One or two per section is fine. Five per paragraph means you're using them as a crutch instead of writing clearer sentences.

Instead: Use commas, colons, semicolons, or periods. The em-dash is a spice, not a staple.

---

9. Preamble Announcements

The tell: Announcing what the post is about to do before doing it.

Symptoms:

• "In this post, we'll explore..."
• "Let's take a look at..."
• "What follows is..."

Why it's a tell: The TLDR handles the preview. The reader clicked the title. They know what the post is about. Announcing it again is filler.

Instead: Just start. Open with the hook, the story, the confession. The reader doesn't need a table of contents narrated to them.

---

10. Sycophantic Hedging

The tell: Filler phrases that soften a point without adding information.

Symptoms:

• "It's worth noting that..."
• "Interestingly enough..."
• "To be fair..."
• "It bears mentioning that..."

Why it's a tell: These are throat-clearing. They signal the writer isn't confident enough to just make the point. LLMs insert them as politeness padding.

Instead: Delete the hedge and start with the actual point.

---

11. Emojis

Never. Zero. Not even in the TLDR. Not even ironically.

---

12. AI Vocabulary Contamination

The tell: Specific words and phrases that appear 5-50x more frequently in AI-generated text than in human writing. These are LLM "comfort words" — they sound authoritative to a model but scream machine to a reader.

The watchlist:

Verbs: "delve", "underscore", "highlight" (as verb), "foster", "leverage", "harness", "showcase", "streamline", "navigate" (abstract), "cultivate", "illuminate", "orchestrate", "spearhead"

Nouns: "tapestry", "landscape" (abstract), "realm", "journey" (abstract), "ecosystem", "paradigm", "trajectory", "blueprint", "interplay", "intricacies"

Adjectives: "pivotal", "crucial", "vital", "nuanced", "multifaceted", "robust", "seamless", "comprehensive", "cutting-edge", "groundbreaking", "transformative"

Inflation phrases: "plays a significant role in shaping", "serves as a testament to", "it is important to note", "a vibrant tapestry of"

Why it's a tell: Research tracking word frequency before and after ChatGPT shows "delve" spiked 10-50x, "tapestry" and "landscape" (abstract) 5-20x, and "plays a significant role in shaping" appeared 207x more often in AI text. Readers have learned to pattern-match on these words even if they can't articulate why.

Instead: Use plain words. "Delve into" → "look at." "Leverage" → "use." "Navigate the landscape" → "figure out." "Pivotal" → "important" (or better: show why it matters instead of asserting it). "Nuanced" → "detailed" or just describe the actual nuance.

One of these words in a 2,000-word post is fine. Three in a paragraph is a contamination event. Scan for them.

---

13. Copula Avoidance

The tell: LLMs avoid simple "is", "are", and "has" constructions, substituting elaborate verbs that inflate the significance of mundane statements.

Symptoms:

• "serves as" / "stands as" / "acts as" instead of "is"
• "boasts" / "features" / "offers" instead of "has"
• "represents" / "marks" / "signals" instead of "is"
• "underscores" / "highlights" / "reflects" instead of "shows"

Examples:

• ❌ "This feature serves as a bridge between the developer and the runtime."
• ❌ "The config file stands as the single source of truth."
• ❌ "The dashboard boasts a real-time monitoring panel."

Why it's a tell: These substitutions make everything sound ceremonial. A config file doesn't "stand as" anything. It IS the source of truth. The inflated verb implies the sentence is making a grander point than it actually is.

Instead: Use "is", "are", "has." They're not boring — they're precise.

• ✅ "This feature connects the developer to the runtime."
• ✅ "The config file is the source of truth."
• ✅ "The dashboard has a real-time monitoring panel."

---

14. Low Burstiness

The tell: Every sentence is roughly the same length and structure. No rhythm variation between short punchy sentences and longer complex ones.

Symptoms:

• Three or more consecutive sentences within 5 words of each other in length
• Every paragraph has the same cadence: medium sentence, medium sentence, medium sentence
• No one-line paragraphs for emphasis, no long flowing sentences for complexity
• Reads like it was generated by the same function called in a loop

Why it's a tell: Human writers naturally vary sentence length. A short sentence after a long one creates emphasis. A long sentence after two short ones builds complexity. LLMs produce text with remarkably uniform sentence length — researchers call this "low burstiness." It's one of the most reliable structural signals of AI text, even when the vocabulary is clean.

Instead: Read your paragraphs out loud. If every sentence takes the same number of breaths, rewrite. Break a long sentence into a punchy two-worder. Combine two medium sentences into one that flows. The goal is rhythm, not uniformity.

Good target distribution: roughly 20% short (under 10 words), 50% medium (10-20 words), 30% long (20+ words). You don't need to count — just listen for monotony.

---

15. Fabricated Experience

The tell: The AI simulates personal encounters, emotional reactions, or sensory moments that never happened. It narrates an internal state to sound human.

Symptoms:

• "What I find genuinely interesting is..."
• "I keep bumping into this pattern..."
• "Picture the workflow: you sit down at your desk and..."
• "I've been wrestling with this question..."
• "Something feels off about this approach..."
• "The thing that struck me watching this was..."

Why it's a tell: Blog posts SHOULD have personal voice and real experience — that's what makes them good. The problem is when the AI manufactures these moments instead of drawing from the author's actual experience. The fabricated version is always vague ("I find this interesting") where real experience is specific ("I tried this at 2 AM and the build broke in a way I'd never seen").

Instead: Personal moments must come from the source material — the transcript, the author's notes, the actual events. If the author said something funny on camera, use that. If you don't have a real moment, don't invent one. Write the point directly instead of wrapping it in fake experience.

• ❌ "I keep coming back to this idea that agents need guardrails."
• ✅ "Twelve minutes into the demo, the agent deleted the test database. That's when we started talking about guardrails."

---

16. False Ranges

The tell: "From X to Y" constructions that claim universal applicability without evidence. The range sounds inclusive but says nothing specific.

Symptoms:

• "From beginners to experts"
• "From startups to enterprises"
• "From simple scripts to complex architectures"
• "Whether you're building a side project or scaling to millions of users"

Examples:

• ❌ "This approach works for teams of all sizes, from two-person startups to hundred-engineer organizations."
• ❌ "From junior developers just learning the ropes to seasoned architects designing distributed systems, everyone can benefit."

Why it's a tell: Real tools have a sweet spot. Real blog posts are honest about who they're for. "From X to Y" is a hedge disguised as inclusivity — it avoids committing to an audience because the AI doesn't know who the audience actually is. It's the equivalent of a restaurant claiming to serve "everyone from toddlers to gourmands."

Instead: Be specific about who this is for and own the limitation.

• ✅ "If you've shipped a production app and gotten paged for it, this will make sense. If you're still in tutorials, bookmark it for later."

---

17. Synonym Cycling

The tell: The same thing is called by a different name every time it appears. "The CLI" becomes "the tool" becomes "the interface" becomes "the command-line solution" across consecutive paragraphs — all referring to the exact same thing.

Symptoms:

• A technical concept is named differently in each paragraph
• "Developers" → "engineers" → "practitioners" → "builders" → "teams"
• "The feature" → "the capability" → "the functionality" → "the offering"
• Each synonym is less precise than the original term

Examples:

• ❌ "The CLI scaffolds the project. Once the tool finishes, review the output. The interface will prompt you for confirmation."
• ❌ "Developers can install the tile. Engineers who need customization can fork it. Practitioners in larger teams should use the workspace feature."

Why it's a tell: LLMs are trained with repetition penalties that discourage reusing tokens. So the model reaches for synonyms even when the original word was the right one. In technical writing, this creates confusion — the reader wonders whether "the tool" and "the interface" are the same thing or different things. Repetition of the precise term is clarity. Variation with imprecise synonyms is noise.

Instead: Use the same word. "The CLI" is "the CLI" every time. If it appears too often in a passage, restructure the sentences or use a pronoun — don't swap in a vaguer synonym.

• ✅ "The CLI scaffolds the project. Once it finishes, review the output. It will prompt you for confirmation."

---

18. Unicode Giveaways

The tell: The draft contains Unicode characters that a human typing in a text editor would not produce. These are character-level fingerprints of LLM-generated text.

Characters to scan for:

• Curly double quotes: \u201C (") and \u201D (") — instead of straight " (U+0022)
• Curly single quotes: \u2018 (') and \u2019 (') — instead of straight ' (U+0027)
• Horizontal ellipsis: \u2026 (…) — instead of three dots ...
• Bullet character: \u2022 (•) — instead of markdown - or *
• En dash: \u2013 (–) — instead of a hyphen or em-dash

Why it's a tell: A human writing in a text editor, markdown file, or code editor produces straight quotes, three dots, and hyphens. The CMS or publishing platform converts these to typographically correct characters for the reader. When the markdown source already contains curly quotes or a Unicode ellipsis, it means the text was generated by an LLM (ChatGPT in particular outputs curly quotes by default).

Instead: Use plain ASCII in the draft. Straight quotes, three dots, hyphens. Let the publishing system handle typography. During the anti-pattern scan, search the draft file for these Unicode characters and replace any that appear.