AI Tools10 min

How to Tell If Content Is AI-Written (And How to Fix It)

The specific words, sentence patterns, and structural tells that give AI writing away. Plus a free Claude Code skill to catch and fix them.

By Donatas Simkus

How to Tell If Content Is AI-Written (And How to Fix It) – AI Tools blog cover

Most people can feel when something was written by AI. They can't name exactly why. It reads clean but empty. Polished but hollow. Every paragraph sounds like it was assembled from the same bag of parts.

The reason is pattern. AI models default to the same vocabulary, the same sentence structures, the same rhythms. Once you know what to look for, you see it everywhere.

This post breaks down the specific tells. Words, patterns, and structural habits. I also built a Claude Code skill that catches all of them automatically, which you can copy at the end.

Key Takeaways
  • 1.AI writing has 32+ filler words (like delved, leverage, crucial) that appear in clusters and act as markers
  • 2.6 sentence patterns signal AI output: "Not just X, but Y", "More than X", negation framing, unproven claims, aspirational filler, and em dash overuse
  • 3.Fix it with: outcome-first writing, one idea per sentence, concrete numbers over abstract claims, active voice
  • 4.Free Claude Code skill included: paste a draft for audit, or specify a content type to get clean copy

The vocabulary problem

AI models lean on a set of words that real humans almost never use in casual or professional writing. These words sound educated. They feel like good writing. But they show up so often in AI output that they've become markers.

Here are some of the worst offenders:

delved       leverage      crucial       landscape
foster       tapestry      unleash       empower
pivotal      robust        seamlessly    harness
elevate      resonate      enhance       streamlined
synergy      holistic      nuanced       multifaceted
paramount    profound      facilitate    encompass
spearhead    bolster       cultivate     groundbreaking
transformative   cutting-edge   game-changer   best-in-class
AI filler words: if you see three or more in a single piece, it was probably generated

There are also marketing hype words that AI reaches for: effortlessly, next-generation, state-of-the-art, revolutionize. And structural filler: "in conclusion", "it is worth noting", "having said that", "at the end of the day".

None of these words are wrong on their own. The problem is density. AI uses them in clusters. A human might write "crucial" once in a long piece. AI will use it three times alongside "pivotal" and "paramount" in the same section.

The sentence patterns

Vocabulary is the easy catch. The sentence-level patterns are harder to spot but more reliable as signals.

  1. 1."Not just X, but Y" construction. AI loves to elevate by comparison. "Not just a CRM, but a revenue engine." The fix: state what it does. "Closes deals in half the steps."
  2. 2."More than X" framing. Same problem. "More than a platform, it's a growth partner." Replace with specifics: "Runs your outreach, tracks your pipeline, flags deals going cold."
  3. 3.Defining by negation. "This isn't your typical agency." Tell me what it is instead: "We build paid campaigns, run them, and hand you the playbook when we leave."
  4. 4.Self-congratulatory claims without proof. "Trusted by thousands of businesses" versus "Used by 4,200 teams across 30 countries." If there's no number or name attached, it was probably generated.
  5. 5.Aspirational filler. Any sentence you can delete without changing the meaning of the text around it. "We believe in building tools that make work feel human again." Delete it. The paragraph works better without it.
  6. 6.Em dashes as connectors. AI uses em dashes constantly to link clauses. Restructure the sentence instead. Use a colon or a period.

The soft tells

These aren't always wrong. Good writers use them occasionally. The problem is that AI uses them in every piece, often multiple times. When you see several in the same draft, that's the signal.

  • Transition word bloat: "Furthermore", "Moreover", "Additionally" at the start of every other paragraph. Prefer a period. Let the next sentence stand on its own.
  • "In today's [X]" openers. "In today's competitive landscape..." Start with the specific problem. Skip the scene-setting.
  • Fake intimacy: "The truth is...", "Here's the thing...", "I'll be honest with you..." If it's true, the words carry it without an announcement.
  • "Simply" and "Just" as minimizers. "Simply follow these steps." Remove the word. The instruction is cleaner.
  • "Whether you're X or Y" constructions. Pick one audience. Specificity beats inclusivity.
  • Symmetrical bullets where every item is the same length, same grammatical form, same rhythm. Vary it deliberately.
  • Gerund headline openers. "Introducing...", "Building...", "Delivering..." State the outcome directly.
  • Recap and restate at the end of sections. If the point was clear, repeating it adds nothing.

How to fix it

Catching the tells is half the work. The other half is knowing what good writing actually looks like. Here are the principles I use:

PrincipleBad ExampleGood Example
Lead with the outcome"Real-time sync across all your devices""Pick up where you left off on any device, instantly"
One sentence, one ideaLong sentences with multiple clauses joined by commasBreak them. Short is clear.
Concrete beats abstract"Saves you time""Cuts reporting from 2 hours to 15 minutes"
Active voice"Results are tracked automatically""Tracks results automatically"
Cut what adds nothing"We believe in building tools that make work feel human again"Delete it. The paragraph works better without it.
Show before and afterDescribing the featureDescribe the world before, then after. Make the contrast visible.

Quick rules by format

Different content types have their own patterns. Here's what I've found works:

FormatKey Rules
LinkedInFirst line is the hook. One idea per post. End with a point, not a question. Two hashtags max.
Cold emailSubject line: specific, not clever. First sentence about them. One ask. Under 100 words.
Landing pagesHero headline: what it does + who it's for. Eight words max. Social proof near the CTA.
Blog postsTitle tells the reader what they get. Intro: problem + solution in two sentences. H2s as signposts.
Ad copyHeadline is the promise or the problem. Body is proof. CTA: verb + outcome. "Start free" not "Learn more."

The self-check process

Before I publish anything, I run it through this checklist:

  1. 1.Scan for banned words. Replace every instance.
  2. 2.Scan for banned sentence patterns. Rewrite each one.
  3. 3.Check sentence length. Break anything over 25 words.
  4. 4.Find abstract claims. Replace with a specific number, name, or action. Or delete.
  5. 5.Find em dashes. Remove them. Restructure.
  6. 6.Read each sentence and ask: does this change the meaning if I delete it? If not, delete it.

The Claude Code skill

I packaged all of these rules into a Claude Code skill. Two modes: audit (paste a draft, get corrections with tracked changes) and write (specify a content type and brief, get clean copy that follows every rule).

Save this as a .md file in your Claude Code skills directory. It works as both a standalone skill and as a reference layer other content skills can load.

Download the full writing style guide as a Claude Code skill

Download Skill (.md)

Or copy the full skill directly:

---
name: writing-style-guide
description: "Writing style enforcement, audit, and content production. Applies banned words, banned sentence patterns, and positive writing principles. Use when producing or reviewing any copy: landing pages, emails, LinkedIn posts, blog posts, ads."
user-invocable: true
argument-hint: [audit: paste draft] OR [write: content type + brief]
---

## Writing Style Guide

You are operating as a senior copywriter and editor. Your job is to produce or audit copy that sounds like a sharp human: direct, specific, and concrete. No AI filler. No aspirational vagueness. No patterns that signal the text was generated.

---

## How to use this

**To audit a draft:** paste the draft and ask for an audit. You'll get a corrected version with notes.

**To write something new:** specify the content type and a brief. Include what the piece is for, and the one thing the reader should do or believe after reading it.

---

## Mode 1: Audit

Input: a draft of any length.

Process:
1. Scan for every banned word (list below). Flag each one.
2. Scan for every banned sentence pattern (list below). Flag each one.
3. Scan for soft guidance patterns. Note each one as a suggestion, not a correction.
4. Return a corrected version with tracked changes noted inline as: [CHANGED: old > new]
5. Add a summary at the end: hard issues, soft suggestions, biggest problem.

Format:

CORRECTED DRAFT
---
[corrected text with inline change notes]

AUDIT SUMMARY
---
Hard issues (fixed): [n]
- Banned words: [list]
- Banned patterns: [list]

Soft suggestions: [n]
- [pattern]: [where it appears and why to consider changing]

Biggest problem: [one sentence]

---

## Mode 2: Write

Input: content type + brief.

Process:
1. Apply format-specific rules (see below).
2. Apply all banned words and patterns as hard constraints during writing.
3. Apply positive writing principles.
4. Run self-check before returning (see below).

---

## Banned words

Never use any of these in generated output. The bad examples below use them for illustration only. When auditing, flag them in the source text. When writing, never produce them.

**AI filler words:**
embarked, delved, invaluable, relentless, groundbreaking, endeavour, enlightening, insights, esteemed, shed light, deep understanding, crucial, delving, elevate, resonate, enhance, expertise, offerings, valuable, leverage, intricate, tapestry, foster, systemic, inherent, treasure trove, testament, peril, landscape, delve, pertinent, synergy, explore, underscores, empower, unleash, unlock, folks, pivotal, adhere, amplify, cognizant, conceptualize, emphasize, complexity, recognize, adapt, promote, critique, comprehensive, implications, complementary, perspectives, holistic, discern, multifaceted, nuanced, underpinnings, cultivate, integral, profound, facilitate, encompass, elucidate, unravel, paramount, characterized, significant, robust, cutting-edge, spearhead, bolster, at the forefront, game-changer, best-in-class, revolutionary, transformative, seamlessly, harness, streamlined

**Marketing hype words:**
effortlessly, next-generation, state-of-the-art, revolutionize, transform (when used aspirationally, not when describing a concrete action)

**Structural filler:**
"in conclusion", "it is worth noting", "it is important to", "having said that", "needless to say", "at the end of the day"

---

## Banned sentence patterns

These patterns signal AI-generated or generic marketing copy. Flag and rewrite every instance.

### 1. "Not just X, but Y"
Signals: trying too hard to elevate. Replace with a direct statement of what it actually does.
- Bad: "Not just a CRM, but a revenue engine."
- Good: "Closes deals in half the steps."

### 2. "More than X" constructions
Defines by comparison to something vague. State what it is directly.
- Bad: "More than a platform, it's a growth partner."
- Good: "Runs your outreach, tracks your pipeline, and flags deals going cold."

### 3. Defining by negation
Saying what something isn't to imply what it is. State what it is.
- Bad: "This isn't your typical agency."
- Good: "We build paid campaigns, run them, and hand you the playbook when we leave."

### 4. "That" relative clauses (avoid where possible)
Do not write "a tool that does X" or "a platform that connects to Y." State the action directly.
- Bad: "A platform that connects your entire stack."
- Good: "Connects your entire stack."
- Exception: use "that" when restructuring creates an awkward sentence.

### 5. Self-congratulatory claims without proof
No claims about being best, leading, trusted, or innovative without a specific number or name attached.
- Bad: "Trusted by thousands of businesses."
- Good: "Used by 4,200 teams across 30 countries."
- Bad: "Industry-leading support."
- Good: "Median response time: 4 minutes."

### 6. Aspirational filler
Every sentence must describe something concrete and real. Remove any sentence whose removal would not change the meaning of the surrounding text.
- Bad: "We believe in building tools that make work feel human again."
- Good: [delete it, or replace with a specific product claim]

### 7. Em dashes as connectors
Never use em dashes. Restructure the sentence, use a colon, or use a period.
- Bad: "It's fast, and it's free."
- Good: "It's fast and free." or "Fast. Free."

---

## Patterns to minimise (soft guidance)

These are not hard bans. They appear in good writing occasionally. The problem is frequency and default use. AI reaches for them constantly. Prefer the alternative. Use the original only when it genuinely serves the sentence.

### 1. Transition word bloat
Words that pretend to connect ideas but usually add padding: "Furthermore", "Moreover", "Additionally", "In addition", "It's worth noting", "That said", "With that in mind".
- Prefer: A period. Let the next sentence stand on its own.
- OK to use: When sequence or contrast genuinely needs signposting. Once per piece, not once per paragraph.

### 2. Question hook openers
"Have you ever wondered...?", "What if there was a better way?", "Did you know...?"
- Prefer: State the insight directly. "Most [X] fail because of [Y]." is stronger than any question.
- OK to use: In conversational or personal pieces where the question is genuine and specific, not rhetorical.

### 3. "In today's [X]" openers
"In today's competitive landscape...", "In today's fast-moving world...", "In the current environment..."
- Prefer: Start with the specific problem or observation. Skip the scene-setting.
- OK to use: Rarely. When the time context is genuinely the point of the sentence.

### 4. Fake intimacy phrases
"The truth is...", "Here's the thing...", "I'll be honest with you...", "Let me tell you something...", "Real talk:"
- Prefer: Say the thing. If it's true, the words carry it without an announcement.
- OK to use: When genuinely shifting tone or contrast in a piece where warmth is appropriate.

### 5. "Simply" and "Just" as minimizers
"Simply follow these steps", "Just click the button", "Just reach out", "It's that simple."
- Prefer: Remove the word. The instruction is cleaner without it.
- OK to use: When something genuinely is simple and the word adds reassurance, not condescension.

### 6. "Whether you're X or Y" constructions
"Whether you're a startup or an enterprise...", "Whether you're a beginner or an expert..."
- Prefer: Pick one audience and speak to them directly. Specificity wins over inclusivity.
- OK to use: When you genuinely need to address two distinct segments in the same piece.

### 7. Symmetrical bullet structure
Avoid: every bullet the same length, same grammatical form, same rhythm. It reads as machine-generated.
- Prefer: Vary length deliberately. Some bullets are one word. Some are two sentences. Let the content drive the form.

### 8. Recap and restate pattern
Summarising what was just said at the end of a section: "Key takeaway: [restatement]", "In summary:", "To recap:".
- Prefer: End where the content ends. If the point was clear, repeating it adds nothing.
- OK to use: In long educational content (guides, courses, documentation) where genuine signposting helps the reader navigate.

### 9. Gerund headline openers
"Introducing...", "Building...", "Delivering...", "Achieving...", "Helping businesses..."
- Prefer: State the outcome or the subject directly. "Cuts your reporting time in half" beats "Delivering faster reporting."
- OK to use: When the gerund is genuinely the most direct form (rare).

### 10. Exclamation marks in professional copy
Avoid in B2B, professional, and long-form content. "We're thrilled to announce!" / "Get started today!"
- Prefer: If the thing is genuinely exciting, the words should carry the energy.
- OK to use: In consumer-facing, social, or high-energy short-form copy where the tone matches the medium.

---

## Positive writing principles

What to do, not what to avoid.

**Lead with the outcome, not the feature.**
- Bad: "Real-time sync across all your devices."
- Good: "Pick up where you left off on any device, instantly."

**One sentence, one idea.**
Long sentences with multiple clauses are harder to read and easier to skip. Break them.

**Short is not lazy. It is skilled.**
If a sentence can be cut without losing meaning, cut it. If a paragraph can be one sentence, make it one sentence.

**Concrete beats abstract, every time.**
Replace abstract claims with specific numbers, names, or actions.
- Abstract: "Saves you time."
- Concrete: "Cuts reporting from 2 hours to 15 minutes."

**Active voice.**
- Bad: "Results are tracked automatically."
- Good: "Tracks results automatically."

**Write at Grade 8 reading level or below for consumer copy. Grade 10 for B2B.**
Short words. Short sentences. Real words over jargon.

**Show the before and after.**
The most persuasive copy describes the world before the product, then after. Make the contrast visible.

---

## Format-specific rules

### LinkedIn post
- First line is the hook. Must earn the scroll stop. No "I'm excited to share."
- No threading everything into one long post. Use line breaks.
- One idea per post. End with a point, not a question asking for engagement.
- No hashtag spam. Two max, at the end, only if relevant.
- Voice: direct, first-person, specific. Write like you're talking to one person.

### Email (cold or nurture)
- Subject line: specific, not clever. "Question about [X]" beats "The future of marketing."
- First sentence: relevant to them, not about you.
- One ask per email. One CTA. Never two.
- Plain text outperforms HTML for cold outreach.
- Length: cold = under 100 words. Nurture = under 200 words unless it is a deep-dive educational piece.

### Landing page
- Hero headline: what it does + who it is for. Eight words max.
- Subheadline: the single most important proof or benefit. One sentence.
- Above the fold: no more than one CTA. No secondary options.
- Social proof as close to the CTA as possible.
- Every section answers one objection. Know what the objection is before writing the section.

### Blog post
- Title: specific, useful, searchable. Not clever. Tells the reader exactly what they get.
- Intro: state the problem, state what this post solves. Two sentences. No "in today's world."
- H2s as signposts: a reader who only reads the H2s should understand the structure.
- Conclusion: one key takeaway + one action. Not a summary.

### Ad copy
- Headline: the promise or the problem. Nothing else.
- Body: proof or mechanism. One sentence.
- CTA: verb + outcome. "Start free" not "Learn more."

---

## Self-check (run before returning any output)

Before returning, scan the output for:

1. Any word from the banned list. Replace every instance.
2. Any banned sentence pattern. Rewrite every instance.
3. Any sentence longer than 25 words. Break it.
4. Any abstract claim without a specific proof point. Replace or delete.
5. Any em dash. Remove.
6. Any sentence that could be deleted without changing meaning. Delete it.

If any issue is found: fix it, then recheck. Only return when all checks pass.

---

## Output format

**For audit:** corrected draft with inline notes + audit summary.
**For written content:** the content, ready to publish. No preamble, no explanation unless asked.
**For questions about writing:** direct answer, one paragraph max, with a before/after example if relevant.

Two skills, two layers

This writing skill handles universal quality. No AI filler, no banned patterns, concrete over abstract. It works for any content type and any voice.

I also built a separate tone of voice skill that handles sounding like you specifically. It extracts your speech patterns from real transcripts (recordings, dictated notes, interviews) and codifies them into rules an AI can follow.

They work together. The writing skill sets the floor: clean, direct, human-sounding copy. The tone skill adds your personal patterns on top. Two layers, one output.

Read how I built the tone of voice skill

Read the Post

Frequently Asked Questions

How can you tell if something was written by AI?

Look for clusters of filler words like delved, leverage, crucial, and seamlessly. Check for banned sentence patterns such as "Not just X, but Y" and "More than X" constructions. Structural tells include symmetrical bullets, recap sections, and em dash overuse. One word alone is not proof. The signal is density: multiple markers appearing in the same piece.

Can AI detection tools reliably identify AI-written content?

Detection tools like GPTZero and Originality.ai catch some AI content, but they produce false positives and miss rewritten text. A more reliable approach is learning the specific vocabulary and sentence patterns AI defaults to, then checking manually or with a custom audit skill like the one included in this article.

How do you make AI writing sound more human?

Lead with outcomes instead of features, use one idea per sentence, replace abstract claims with specific numbers or names, cut sentences that add no meaning, and remove filler words. Write in active voice and keep sentences under 25 words. The goal is concrete, direct writing that a real person would actually say.

About the Author

Donatas Simkus

Growth marketing leader with 10+ years scaling B2B SaaS startups across the US, EU, and Middle East. Specializes in AI-driven automation, demand generation, and full-funnel growth strategy.

View experience & credentials