Stop Sounding Like AI: 205 tells, 3 Claude skills and 1 checker that make your AI writing, decks and designs read like a pro made them

The kit I use so clients read the work, not the tool. Free.

Last verified: October 2026.

What's in the folder

Start here: your 5-minute win

1. Double-click tools/slop-checker.html. It opens in your browser. No install, no account, no internet needed.
2. Paste the last thing you wrote with AI and were about to send. An email, a LinkedIn post, a proposal intro. (Or press Load sample text to see how it works.)
3. Look at the score and the colored highlights. Tap any highlight to see the fix.
4. Press "Copy fix-it prompt". It copies a rewrite prompt that includes the exact tells found in your text.
5. Paste it into Claude, ChatGPT or Gemini. You get a rewrite plus a short list of [ADD: ...] gaps where a real number or name would make it believable.
6. Fill the gaps with real facts, paste the result back into the checker. Aim for 80+.

That's it. The client email in Chapter 3 went from 38 to 95 this way, and the gap list shows you where a draft had nothing real to say.

Chapter 1: Why people can tell (it's not one word)

You've seen the jokes about "delve". It's real: a 2025 study in Science Advances (Kobak et al.) looked at over 15 million biomedical abstracts and found words like "delves", "showcasing" and "underscores" jumped in frequency after ChatGPT arrived. The authors estimate at least 13.5% of 2024 abstracts were processed with an LLM. Wikipedia's editors keep a public field guide, "Signs of AI writing", for spotting machine-written articles. And Merriam-Webster picked "slop" as its 2025 Word of the Year, defined as low-quality digital content produced in quantity by AI.

So your reader has been trained. They spot AI text in three layers:

Layer 1: Words. The vocabulary that spiked after 2022. Delve, tapestry, landscape, seamless, robust, leverage, pivotal, testament. One is fine. Three in a paragraph is a signature.

Layer 2: Shapes. You can delete every word above and still sound like a model. "It's not just X, it's Y." "The result? ..." Lists of three adjectives. An em dash every other sentence. Every sentence about 15 words long. A moral at the end of each paragraph.

Layer 3: Emptiness. This is the one that loses clients. No numbers, no names, no dates, no opinion. Text that could be sent to anyone, so it feels sent to no one. The Reddit post that started this kit said it best: a client told the writer their strategy deck "sounded like every other strategy deck." The words were fine. There was nothing in it that belonged to that client.

Most "humanizer" tools only fix layer 1. They swap synonyms and the text still reads like AI, because layers 2 and 3 are untouched. This kit fixes all three, in that order of difficulty:

Chapter 2: The Slop Checker, and how to read the score

Open tools/slop-checker.html. It works offline, on a phone too.

What it detects (198 detectors):

• 187 phrase rules across 9 categories: AI vocabulary, throat-clearing openers, robotic transitions, hedging, significance inflation, empty intensifiers, chatbot residue, generic CTAs, deck/corporate clichés.
• 11 structural patterns: "not just X, it's Y", "This isn't X. It's Y.", "not only... but also", fake suspense questions ("The result?"), colon reveals ("Here's why:"), "From X to Y" openers, -ing openers ("Leveraging AI, we..."), triplet lists, emoji bullets, exclamation marks, em dashes.
• Rhythm: how much your sentence lengths vary.
• Specificity: whether there's a single number in the text.

How the score works. It starts at 100. Each tell costs points based on how strongly readers link it to AI ("as an AI language model" costs far more than "very") and on density: one "moreover" in 600 words barely moves the score; four in 100 words tank it. Em dashes only cost real points when you use more than about one per 200 words. Monotone rhythm costs up to 15 points. Zero numbers in 80+ words costs 6.

The rhythm chart. Each bar is a sentence, colored by length. Human writing looks like a city skyline: tall, short, medium, very short, long. AI writing looks like a fence: the same height again and again. The "variation" number is the standard deviation of sentence length divided by the average. Under 40% reads monotone. Over 55% reads natural.

Honest limits of the checker.

• It's a pattern matcher, not a judge. A human who loves "moreover" will get flagged. A model told to avoid every pattern on the list will pass. The score tells you how your text reads, not who wrote it.
• It's English only.
• It's not an AI detector and won't help anyone pass one. That's not the goal. The goal is a reader who trusts you.

Chapter 3: De-slop any text in 2 minutes

The de-slop skill rewrites text using 12 rules. The full rules, a self-check and 4 before/after examples are in skills/de-slop/SKILL.md. Here's the core:

1. Lead with the point.
2. Concrete nouns over abstractions.
3. Numbers, names, dates, and never invented. Missing specifics become [ADD: ...] gaps.
4. Plain words.
5. Vary the rhythm.
6. One claim per sentence.
7. Take a position.
8. No banned phrases (and no fancy synonyms for them).
9. Kill the AI shapes.
10. Write to one person.
11. Earned emphasis only.
12. End on the action.

Rule 3 is the one that matters most. Ask any model to "make this sound human" and it invents details: a fake client name, a made-up 47% result. That's worse than slop, because it's false. Both the skill and the prompt forbid it and use [ADD] gaps instead. A rewrite with five gaps is the kit telling you the draft had no substance yet.

Before and after: a client email

Before (raw ChatGPT):

Hi Sarah, I hope this email finds you well. In today's competitive landscape, having a robust online presence is crucial. Our agency specializes in crafting tailored solutions that empower businesses like yours to unlock their full potential. Would you be open to a quick call to explore how we can help?

Slop Checker: 38/100. Seven tells in 51 words, zero numbers, flat rhythm.

After (de-slop + 3 real facts from the sender):

Hi Sarah, I tried to book a cleaning on your site last night. The form asks for 11 fields before it shows a time slot. I gave up at the insurance question. Most of your visitors probably do too.

I fix this kind of page for dental clinics. Usually it's a 2-day job: move the time slot first, ask for details after.

Want me to send a 2-minute screen recording of what I'd change? No call needed.

Slop Checker: 95/100. Nothing clever happened. A real observation replaced flattery, and the ask got smaller.

Install it

• Claude (web/desktop app): zip the skills/de-slop folder (the folder itself, not just the files inside), then upload it under Settings → Capabilities → Skills. Skills need a paid plan with code execution turned on. Menu names move; if you can't find it, search Claude's help center for "skills".
• Claude Code: copy the folder to ~/.claude/skills/de-slop/. Then just say "de-slop this" with your text.
• ChatGPT, Gemini, anything else: open skills/de-slop/PROMPT.md, copy the box, paste your text at the bottom.

Same steps for voice-clone and human-design.

Chapter 4: Clone your own voice (15 minutes, once)

De-slop gets you to "sounds like a sharp human." Voice clone gets you to "sounds like you." That matters most for anyone whose clients or audience already know how they write.

What you need: 3–5 things you wrote without AI, 800–3,000 words total, ideally in 2+ formats. Old client emails and Slack messages work better than polished posts, because they're how you actually talk.

What you get: a VOICE GUIDE with 8 sections: your sound in one line, rhythm numbers (average sentence length, % of short sentences), signature words, words you never use, punctuation rules (counted, not guessed), how you open and close, your attitude with a quoted example, and 10 do/don't rules a ghostwriter would need. It ends with a calibration paragraph you correct once.

How to use it every day: save the guide in a Claude or ChatGPT Project's instructions (or a Gemini Gem). From then on, every draft in that project starts in your voice. Run de-slop with the guide attached and the guide wins wherever they conflict. If you love em dashes, you keep your em dashes.

Full steps: skills/voice-clone/SKILL.md (Claude) or skills/voice-clone/PROMPT.md (any model).

A warning that saves people a week: don't feed it AI-written samples. If your last 20 LinkedIn posts were drafted by ChatGPT, the guide will clone ChatGPT. The skill flags samples that look AI-written, but pick old stuff to be safe.

Chapter 5: Quick recipes by format

Use these with the de-slop prompt, or as a 60-second manual edit.

Emails

• Delete your first line. It's almost always "I hope this finds you well" in some form.
• One ask per email, in the last line, small enough to say yes to in 10 seconds.
• Subject line = the specific thing, not the category. "Your booking page loses people at step 3", not "Website improvement opportunity".
• Follow-ups: never "just circling back." Add one new fact or one new reason.

LinkedIn / X / newsletters

• Open with a scene or a number, not a thesis. "I stopped answering Slack before 11am in March." beats "Productivity isn't about working harder."
• One idea per post. If you wrote "3 tips", cut 2 and go deeper on the best one.
• Admit something. One honest limit or mistake makes the rest believable.
• End with a question someone can answer in one line. Not "What are your thoughts?"

Landing pages

• Headline = outcome + number for one reader. "Close your books in 2 days, not 9."
• Subhead = the mechanism. How it works, in one sentence.
• Proof = real or [ADD]. Never "Join thousands of happy customers" unless you can count them.
• Then run the page design through Chapter 7.

Proposals and client reports

• Restate their problem in their words, using a phrase from their email or the call.
• Show one thing you noticed that they didn't say. That single line proves a human read their stuff.
• Price with a reason. "2 days of work because X" beats "competitive pricing".
• Take a recommendation. Option A, B, C is fine, but say which you'd pick and why.

Chapter 6: Decks that don't sound like every other deck

This is where AI hurts most, because decks run on clichés anyway and the model writes the average of all of them.

The swap test: replace your client's name with their competitor's on any slide. If the slide still works, it's generic.

The SNEF framework: every slide that matters needs four things.

(The numbers above are illustrations from a fictional client.)

The kill condition ("if not X by date, we stop") is the most human thing you can put in a deck. Models almost never write one unless asked, and clients remember the strategist who did.

decks/ANTI-GENERIC-DECKS.md has the 10-slide structure, the prompt, and three full before/after slides: the "strategic pillars" slide, the "market opportunity" slide and the "recommendations" slide.

Chapter 7: Designs that don't look AI-built

Ask any AI to "build a landing page" with no direction and you'll recognize the result before it loads: purple-blue gradient, a blurred orb, a centered gradient headline that says "Revolutionize your workflow", two pill buttons, fake logos, three identical cards with emoji icons. Anthropic's own frontend-design skill exists partly to steer Claude away from exactly this: generic fonts, purple gradients, predictable layouts.

The reason is simple. With no direction, the model picks the most common option at every step. The fix is to make the decisions before it writes code. The human-design skill forces that in 5 steps:

1. Decide 6 things first: exact reader, one action, a reference from outside SaaS templates ("like a Swiss train timetable"), a 2-word personality with one unexpected word, real content, constraints.
2. Lock a system as CSS variables: max 2 fonts, one type scale (4–5 sizes), 8px spacing, 60/30/10 color with one accent, one radius, one icon set.
3. Ban 17 clichés (gradient blobs, glassmorphism, gradient text, sparkle icons, three identical cards, fake logos and stats...).
4. Write real copy: headline = outcome, subhead = mechanism, buttons = verb + object.
5. Audit against a 28-point checklist before showing anything.

The skill includes three before/after descriptions (a landing page hero, an AI-built dashboard, a strategy slide). If you use v0, Lovable, Bolt or Cursor, paste skills/human-design/PROMPT.md as your first message. There's also a second prompt to fix a page that already looks generic.

Chapter 8: The full workflow (what I actually do)

1. Notes first, AI second. Before prompting, write 3–5 bullets of real facts: numbers, names, what you saw, what you think. 2 minutes. This is the step everyone skips, and it's why their output is empty.
2. Draft with your voice guide attached.
3. Paste into the Slop Checker. Under 80? Copy the fix-it prompt and run it.
4. Fill every [ADD] gap with real facts. If you can't fill one, cut that claim.
5. Read it out loud. Anything you wouldn't say to a friend, rewrite by hand.
6. For decks: swap test + SNEF. For pages: the 28-point checklist.

Total extra time: 3–6 minutes per piece. Compare that to a client who quietly decides your work is "just ChatGPT."

Going further

• Build a "never" list for your niche. Every industry has its own slop. Real estate has "nestled" and "boasts". SaaS has "seamless" and "all-in-one". Coaching has "unlock" and "journey". Add your 20 to the bottom of banned-phrases.md and the prompt.
• Add your own rules to the checker. Open slop-checker.html in any text editor, find the list that starts with var R = [, and add a line in the same format: ["vocab","\\byour phrase\\b","What to write instead.",2],. Save, reopen.
• Put the rules where the AI writes. Paste the 12 de-slop rules into your Claude Project or ChatGPT custom instructions. Prevention beats editing.
• Keep a "facts bank". One note with your real numbers: clients served, results, prices, timelines, quotes from happy customers. Paste the relevant part into every prompt. This is the cheapest way to make every output specific.
• Second-opinion pass. Ask a different model than the one that drafted: "You're the client. Which sentence made you trust this less?" Fix that sentence.
• Calibrate the score to your writing. Paste 3 things you wrote before 2022. That's your human baseline. If they score 85, don't chase 100 on new work.

Limits (read this, it'll save you from overpromising)

• This won't beat AI detectors and isn't built to. Detectors are unreliable in both directions. If someone requires "no AI", the honest answer is to not use AI, not to disguise it.
• It can't add knowledge you don't have. If your draft is full of [ADD] gaps, the problem is research, not wording.
• The banned list is about density. Humans use "crucial" and "moreover". Don't strip every instance from writing that's already yours.
• The checker is English-only and pattern-based. It can miss new tells and flag false positives. Trust your ear over the number.
• Voice clone needs real samples. With fewer than 3 or under ~500 words, expect a generic guide.
• Design skill ≠ brand. It makes AI-built pages look deliberate and professional. A distinctive brand still needs a reference and taste from you.

Want this done automatically?

Right now this is the do-it-yourself version: open the checker, copy the prompt, paste it back.

I'm building De-slop: a browser extension + web app that checks your text as you type in Gmail, LinkedIn, Notion, Docs and Slack, underlines the AI tells, and rewrites them in your own voice (from your voice guide) in one click. Planned price: $9/month.

If you'd use it, reply YES to the DM and I'll put you on the early-access list. If enough people say yes, I build it.

Follow @danshipped for the next drop.

Sources (checked October 2026)

• Kobak, D. et al. "Delving into LLM-assisted writing in biomedical publications through excess vocabulary." Science Advances 11, eadt3813 (2025). Preprint: arxiv.org/abs/2406.07016
• Wikipedia: "Signs of AI writing" (WikiProject AI Cleanup). en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
• Merriam-Webster 2025 Word of the Year: "slop" (announced December 2025; reported by PBS NewsHour and TechCrunch).
• Anthropic frontend-design skill (anthropics/claude-code repository), guidance against "generic AI slop aesthetics".
• Claude Skills: Claude Code reads personal skills from ~/.claude/skills/; Claude apps accept skill uploads as zip files on paid plans with code execution enabled (Anthropic docs).

Made by Dan Shipped — @danshipped