Anyone can say they’re an AI expert.

---
name: remove-ai-tells
description: 'Draft, edit, audit or rewrite text that must not sound AI-generated: emails, LinkedIn posts, blogs, landing pages, founder, customer, investor, internal and social comms. Use whenever generic AI style would hurt trust. The user''s own voice always wins.'
---

# Remove AI Tells

A skill file by Alex Hawke (vu.co.uk). Free to use, share and adapt.

Drop this into Claude, ChatGPT, Cursor or any agent as instructions. In Claude Code, save it as `.claude/skills/remove-ai-tells/SKILL.md`.

This is not a generic "make it more human" prompt. It is an editing system for turning smooth, obedient, over-polished text into writing that feels like it came from a specific person, in a specific situation, with a real reason to write.

## The core idea

An AI tell is rarely one word. It is a pattern.

AI writing tends to:

- announce what it is about to say before saying it
- over-explain obvious context
- smooth every rough edge
- use formal synonyms where plain words work
- claim something is important instead of proving it
- force neat transitions between unrelated thoughts
- end with a tidy moral, summary or call to action
- sound like it was written for everyone, which makes it sound like it was written for no one

The fix is not to add random typos, slang or fake mess. The fix is to restore intent, context, pressure, specificity, rhythm and point of view.

A good human edit makes the writing more useful, not just less detectable.

## Priority order

1. Preserve the writer's actual meaning.
2. Match the writer's known style if samples exist.
3. Fit the channel and the recipient.
4. Remove AI-patterned structure and phrasing.
5. Keep the text as short as the job allows.
6. Never invent personal details, anecdotes, facts, sources or opinions.

## The tells

### 1. Throat-clearing

AI starts by describing the task instead of doing the task.

Cut openings like: "In today's fast-paced world...", "In this article, we'll explore...", "I wanted to reach out...", "I hope this message finds you well...", "It is worth noting that...", "When it comes to...", "At its core...".

Replace with the first real fact, question, opinion or action.

### 2. Announcing importance instead of showing it

AI says things are crucial, pivotal, significant, powerful or transformative. Humans show why the thing matters.

Bad:

> Data quality is crucial for AI success.

Better:

> If the CRM has three versions of the same customer, the model will confidently give you three different answers.

### 3. Meta-structure

AI writes like the outline is still showing: "First, let's understand...", "Next, we'll examine...", "There are three key reasons...", "The takeaway is...", "In conclusion...".

Use structure only when it genuinely helps the reader. Otherwise write the thing directly.

### 4. Manufactured contrast

The single most recognisable AI tell. Cut patterns like:

- "It's not X. It's Y."
- "The question isn't A, it's B."
- "This isn't just about X, it's about Y."
- "No X. No Y. Just Z."
- "Not a bug. Not a feature. A fundamental design flaw."

One in a piece might work. Ten is an insult to the reader. Prefer a plain sentence: state Y directly and drop the negation.

### 5. Fake-specific personalisation

AI tries too hard to prove it knows the recipient.

Bad:

> I enjoyed your recent podcast appearance, particularly your point about pragmatic founders leveraging existing models to accelerate product-market fit.

Better:

> Enjoyed your podcast yesterday.

If a human would not naturally quote the point back, do not quote it back.

### 6. Over-polished helpfulness

Cut lines like "I completely understand how frustrating this must be", "I'd be more than happy to assist", "Please don't hesitate to reach out", "I hope this helps".

Use simple acknowledgement and action:

> That's frustrating. I'll check what happened and come back to you today.

### 7. Formal synonym inflation

Prefer the plain version unless the formal word is technically precise.

- leverage, utilise -> use
- optimise -> improve, tune, make faster, make cheaper
- unlock -> get, create, make possible
- facilitate -> help, arrange
- streamline -> simplify, speed up, remove steps
- empower -> give, let, enable
- enhance, elevate -> improve
- foster -> encourage, build
- navigate -> deal with, handle
- robust -> reliable, solid, hard to break
- comprehensive -> complete, full, covers X
- seamless -> avoid unless it genuinely means no visible handoff

Also watch the "serves as" dodge: "serves as", "stands as", "marks", "represents" where a plain "is" works. AI avoids simple verbs because they feel repetitive to it. Humans just say "is".

### 8. Generic positivity

Cut "game-changing", "revolutionary", "groundbreaking", "cutting-edge", "transformative", "best-in-class", "world-class", "innovative solution", "powerful platform".

Replace hype with behaviour, proof, numbers, tradeoffs or a concrete example.

### 9. Magic adverbs

AI reaches for "quietly", "deeply", "fundamentally", "remarkably" and "arguably" to make mundane sentences feel significant: "quietly orchestrating workflows", "fundamentally reshaping how teams work". If the sentence needs an adverb to feel important, the sentence is hiding a missing fact.

### 10. Summary disease

AI repeats the same idea in the intro, the body and the conclusion. Before finishing, check: could the first and last paragraph both be deleted without losing meaning? If yes, delete or replace them.

A human ending says what to do next, states a final opinion, gives the unresolved tradeoff, or just stops when the point is made.

### 11. Formatting that screams model output

Avoid: emoji bullets in professional writing, excessive bolding, every bullet starting with a bolded phrase, too many headings, headings with a colon and a punchy subtitle, numbered lists when order does not matter, every paragraph the same length.

Also avoid unicode decoration: arrows like → and smart quotes where a person typing would produce -> and straight quotes.

Use formatting because the reader needs it, not because the model likes structure.

### 12. The universal paragraph

A paragraph reads as AI when it could be dropped into any company blog, any LinkedIn post, any investor update. Watch for changing landscapes, rising expectations, operational efficiency, meaningful outcomes, modern businesses, today's consumers.

Replace universal claims with local facts: who is affected, what broke, what changed, what the writer saw, what decision needs making.

### 13. False authority

Two flavours. Vague attributions: "experts argue", "industry reports suggest", "observers have noted". If you cannot name the expert, you do not have a source. And invented concept labels: "the supervision paradox", "workload creep", "the acceleration trap". Naming a thing is not the same as making the argument.

### 14. False confidence

AI writes as if it knows everything. Humans are more bounded. Use precise uncertainty when needed: "I think", "My read is", "Looks like", "I might be missing something". Do not over-hedge either.

## High-risk patterns

Not always forbidden, but each needs a strong reason.

**Phrases**: in today's fast-paced world; ever-evolving landscape; at its core; when it comes to; it's worth noting; let's dive in; delve; unlock potential; harness the power; game-changing; seamless experience; robust solution; key takeaway; the bottom line; here's the thing; here's the kicker; make no mistake; moving forward; thought leader; future-proof; supercharge; drive outcomes; foster collaboration; empower teams; shed light; bridge the gap; pave the way; embark on a journey.

**Structures**:

- "It's not X. It's Y." and every variant
- "X is no longer optional." / "X has never been more important."
- "In a world where..." / "Imagine a world where..."
- "Whether you're X or Y..."
- Self-posed questions answered immediately: "The result? Devastating."
- The same sentence opening repeated three times in a row
- Three back-to-back rule-of-three lists
- "From X to Y" where X and Y are not on any real scale
- "Despite these challenges, [optimistic conclusion]"
- A "-ing" clause bolted on to inject fake significance: "...highlighting its importance", "...reflecting broader trends"
- Giving inanimate things human verbs: "the decision emerges", "the data tells us", "the culture shifts". Name the person who acts.
- One metaphor beaten to death across the whole piece. Introduce it, use it once, move on.
- Historical analogy stacking: "Apple didn't build Uber. Facebook didn't build Spotify. Stripe didn't build Shopify."

**Formatting**:

- em dashes everywhere (a human uses two or three per piece, naturally; AI uses twenty)
- every sentence on its own line for drama
- bold-first bullets in every list
- a forced FAQ section
- a generic CTA at the end of every piece

**Words that cluster**: one "comprehensive" is not fatal. A paragraph with "comprehensive", "robust", "streamline", "unlock" and "seamless" is dead. When a high-risk word appears, ask: is it precise, would the writer naturally say it, is there a shorter word, is the sentence hiding a missing example?

## The rewrite process

1. **Find the real job.** Inform, ask, persuade, apologise, sell, update, challenge, get a decision. If a sentence does not help that job, cut it.
2. **Preserve the human material.** A specific detail, a constraint, a number, a tradeoff, a mild opinion, an awkward but honest line, a real deadline. These are usually the best parts. Do not polish them away.
3. **Delete the AI skeleton.** Generic intro, generic conclusion, transition padding, repeated claims, inflated adjectives, fake empathy, fake enthusiasm.
4. **Replace abstractions with evidence.** "Improves efficiency" -> "saves 3 hours a week". "Better visibility" -> "shows overdue renewals in one list". If there is no concrete meaning, cut it.
5. **Downshift the language.** Short words, active voice, contractions by default, ordinary transitions like "but", "so", "because". Write like a competent person talking to another competent person.
6. **Break the rhythm.** Mix short, medium and occasionally long sentences. Do not make every paragraph punchy; that also becomes a pattern.
7. **Add point of view.** A small honest opinion beats a big generic claim.
8. **Make the recipient real.** What do they already know? What would annoy them? What is the smallest useful ask? Do not explain context they already have.
9. **Final pass.** Scan for banned patterns, a generic first sentence, a generic last sentence, repeated rhythm, claims without evidence, em dashes. Rewrite any line that still sounds generated.

## Structure first, for long documents

Most AI tells that survive rewriting live in the document's structure, not its wording. A page built as hero, problem, five-part explainer, phases, FAQ, CTA will keep reading as AI even after every sentence is rewritten.

Before rewriting a long document, name the skeleton. If you find symmetrical sections with parallel headings, an "N things" framework, or a final section that restates the piece, the fix is structural. Rebuild from the writer's real material (real example, real constraint, real opinion) and let the structure emerge. A surface rewrite of a scaffolded document just regenerates the same scaffold with new words.

Emails, posts and short replies skip straight to the rewrite process.

## Context rules

- **Emails**: shorter than the model wants. Real opening, one thought per paragraph, one clear ask, simple sign-off.
- **Cold outreach**: loose genuine reference, proof fast, plain offer, small ask. No over-articulated compliments, no "I noticed..." hooks built from public info.
- **LinkedIn and social**: the tell is usually cadence. No one-sentence-per-line drama, no "Here are 7 lessons", no "Most people get this wrong". One real observation plus one specific example.
- **Blogs**: start with the thing, not why the topic matters. Concrete openings, functional headings, real examples, opinion mixed with instruction.
- **Landing pages**: say what it does, who it is for, what changes, with proof. Screenshots beat adjectives.
- **Support**: calm and specific. What happened, what you are doing, when they hear back.
- **Internal updates**: blunt is fine. What changed, why, what is blocked, next step, owner.
- **Formal, legal or academic**: do not force casual tone. Remove the tells through specificity, evidence and measured uncertainty instead.

## The specificity ladder

When a sentence feels generic, move it down this ladder until it becomes real.

1. Generic claim: "This improves efficiency."
2. Named function: "This reduces manual reporting."
3. Real workflow: "This stops the ops team copying numbers from Stripe into Sheets every Friday."
4. Measurable outcome: "This removes about 90 minutes of Friday reporting."
5. Human consequence: "Sarah sends the update before lunch instead of chasing numbers at 5pm."

Use the lowest true rung. Never invent details to reach a lower rung.

## Human texture without fake mistakes

Do not add typos. Do not make the writer look careless. Human texture comes from real constraints, asymmetric rhythm, specific examples, mild uncertainty, directness, a view not everyone would agree with, and leaving out obvious context.

Bad fake humanisation:

> Heyyy, this is kinda messy but lol I think we should maybe do this??

Good humanisation:

> I think we should do this, but not until the import issue is fixed. Otherwise we'll just create a second mess.

## Before and after

**Blog intro.** Before:

> In today's rapidly evolving business landscape, companies are increasingly looking for innovative ways to leverage AI to streamline operations and unlock new efficiencies.

After:

> Most AI projects don't fail because the model is bad. They fail because nobody gave it one clear job.

**Sales email.** Before:

> I wanted to reach out because I noticed your team is doing exciting work in the compliance space, and I believe our platform could help streamline your workflows.

After:

> Saw you're hiring in compliance ops.
>
> We help teams turn messy policy checks into repeatable workflows. Might be relevant if the new hires are going to be buried in manual reviews.

**Customer reply.** Before:

> Thank you for bringing this to our attention. We understand how important this is and are committed to resolving it as quickly as possible.

After:

> Thanks for flagging it. I can reproduce the issue on my side, so it's not just your account. I'll get a fix pushed today.

**LinkedIn post.** Before:

> The best founders don't just move fast. They move with clarity. Here are three lessons I've learned about building in the age of AI.

After:

> The best AI workflow I've found is still very manual: ask it to do the job, check where it failed, then save that as the instruction for next time. Not glamorous, but it works.

**Product copy.** Before:

> Unlock seamless collaboration across your organisation with a robust, AI-powered platform built for modern teams.

After:

> One place for the notes, decisions and follow-ups from every customer call.

## The final read-through

Before returning any rewritten text, ask:

- Would this sound normal if sent by a busy person?
- Could this have been written by anyone? If yes, make it more specific.
- Is the first sentence real, or just a setup?
- Is the last sentence useful, or just a tidy ending?
- Does every paragraph contain a fact, view, question, decision or next step?
- Are there claims that sound bigger than the evidence?
- Did I add fake warmth?
- Did I accidentally remove the writer's personality?

If it still reads as AI, cut more. The answer is usually not adding style. It is removing performance.

## Output rules

- Return the clean version first, without edit notes, unless asked.
- Keep formatting simple.
- Do not use em dashes.
- Prefer contractions unless the context is formal, legal or compliance-heavy.
- Preserve links, names, dates, prices and technical terms exactly.
- If a sentence is already good, leave it alone.
- Do not make the writing more expressive than the writer would be. The goal is not beautiful. The goal is believable, useful and specific.