The De-Slop Prompt Stack: Six Prompts That Stop Claude and ChatGPT From Sounding Like a Robot

Merriam-Webster's 2025 word of the year wasn't a tech term that crossed over. It was an insult. Slop: roughly, low-quality digital content churned out in volume by AI.

When the dictionary names the thing your content pipeline produces by default, that's worth twenty minutes of your attention. Here are the twenty minutes.

The trust penalty is measurable now

Canva's State of Marketing and AI 2026 report, a Harris Poll of 1,415 marketing leaders and 3,547 consumers across seven countries released in May, found that 97% of marketing leaders now use AI daily. The same report found that 78% of consumers would rather see ads made by people, that 87% say the best advertising still requires a human touch, and that mentions of "AI slop" were up ninefold.

Bynder's Human Touch survey of 2,000 UK and US consumers adds the twist. 52% said they get less engaged the moment they suspect copy is AI-written. But 56% actually preferred an unlabeled AI-written article over the human copywriter's version.

Read those two Bynder numbers together. People can't reliably detect good AI writing. They reliably punish detectable AI writing. The penalty isn't for using the tool. It's for the tells.

So the job is simple to state: remove the tells.

The tells, so you can delete them on sight

Before the prompts, the symptom list. Machine-default writing is recognizable because it always fails the same ways:

• Metronome rhythm. Every sentence runs 15 to 20 words, forever.
• The connector parade: furthermore, moreover, additionally, consequently.
• Giveaway vocabulary: delve, tapestry, landscape, elevate, crucial, robust, journey.
• "It's not just X. It's Y." Fine once. A tell when it's every third paragraph.
• Threes. Everything arrives in threes.
• Confident claims with no source attached.
• Em dashes every other sentence. We ban them at Promptway entirely, and the models adore them.

No single item proves anything. Stacked together, they're the smell. The six prompts below kill them at the source.

The stack

The first three are setup prompts. Run them before drafting, or better, save them permanently (instructions at the end). The last three are editing passes you run on a finished draft.

1. The rhythm and banned-words prompt

The two loudest tells are vocabulary and rhythm, so handle both in one standing instruction:

Apply these writing rules to everything in this conversation:

Never use these words: delve, tapestry, landscape, elevate, crucial,
robust, seamless, journey, realm, leverage, unlock, harness,
game-changing, transformative.

Never open with "In today's" or "In the world of."

No "furthermore," "moreover," "additionally," or "consequently."
Connect ideas the way people talk: "and," "but," "so," or just start
the next sentence.

No em dashes anywhere. Use a period, a comma, or parentheses.

Vary sentence length on purpose. Some sentences under six words.
Some over twenty. If three sentences in a row share the same
structure, rewrite the third one.

Maximum one exclamation point per piece. Prefer zero.

Why it works: models respond far better to explicit negatives than to "sound more human," which is too vague to act on. What goes wrong: in long conversations the rules drift. Repaste them, or save them permanently so you never have to.

2. The voice transplant

The default model voice is generic-competent, which means it belongs to nobody. Give it yours:

Below are samples of my actual writing, about 400 words total.
Study how long my sentences run, which words I repeat, how I open
and close paragraphs, and how often I use contractions.

[PASTE 300 TO 500 WORDS OF YOUR OWN WRITING]

Write the piece I describe next in that voice. Do not reuse my
sentences or my topics. Match the rhythm and the vocabulary level.
If you are not sure whether I would say a phrase, leave it out.

Use writing you produced without trying to impress anyone. Emails to colleagues work better than your polished blog posts. What goes wrong: a sample under 200 words produces a caricature instead of a voice. And the last line matters, because "leave it out" stops the model from filling gaps with its defaults.

3. The coffee test

Our writing guide's oldest rule is to read every draft out loud. This prompt builds the rule into the machine:

Rewrite this draft as if you were explaining it to a smart friend
over coffee. Contractions are fine. Sentence fragments are fine.
Keep every fact. Cut every sentence that sounds like a press
release. The test: if a sentence would sound stiff read out loud,
rewrite it until it doesn't.

This one earns its spot on tone alone. It converts announcements into explanations, which is most of what "sounding human" means.

4. The specificity pass

Vagueness is slop's natural register, and it's also where AI writing quietly lies. This pass fixes both:

Go through this draft and replace every vague claim with a concrete
one. "Saves time" becomes a number of minutes or hours. "Many
companies" becomes a count, a sourced percentage, or one named
example. "Recently" becomes a month and a year.

Important: if you do not have the real specific, do not invent one.
Write [NEED SPECIFIC] in its place so I can fill it in myself.

The [NEED SPECIFIC] line is the load-bearing part. Demand specifics without it and the model will manufacture them. A marked gap is useful. An invented number is a liability with your name on it.

5. The honest objection

Human writing argues with itself a little. Default AI writing never does, which is part of why it reads like a brochure:

Find the weakest claim in this draft. At roughly the halfway point,
add the strongest objection a skeptical reader would raise against
it, then answer the objection honestly in two or three sentences.
If the objection partly wins, concede that part. Do not strawman
the reader.

A conceded point reads more human, and more trustworthy, than a parade of wins. It also forces you to know which of your claims is the weak one, which is worth knowing before your readers find it for you.

6. The true-anecdote opener

Nothing reads more human than a real, specific moment. Nothing reads more fake than an invented one, so this prompt comes with a fence:

Here is something that actually happened: [DESCRIBE THE REAL MOMENT
IN TWO OR THREE SENTENCES: who, where, what went wrong or what
surprised you].

Open the piece with a short scene built only from those details.
Do not add details I did not give you. Keep the scene under 80
words, then move into the main point.

Ask a model for "an engaging opening story" and you get confident fiction. Feed it a real moment and restrict it to that moment, and you get the one thing slop can never fake: something that happened.

Save it once, stop repasting

Don't run this by hand every morning. Prompts 1 through 3 belong in your standing instructions: custom instructions or a Project in ChatGPT, and your project's custom instructions in Claude. Every new chat then starts with your rules and your voice already loaded. Prompts 4 through 6 work best as an editing pass you run on finished drafts before anything ships.

This is the same logic as putting your constraints at the top of the prompt: the rules you cannot afford to lose should be the first thing the model reads, every time, without you having to remember them.

Total setup time is about fifteen minutes, once.

What these prompts will not do

They fix tone. They do not fix truth.

A model can produce a beautifully rhythmic, perfectly voice-matched paragraph built around a wrong number. Verify every figure, name, date, and quote before you ship, every time. Slop that's also wrong is worse than slop.

And keep humans on the pieces where the reader needs to feel a person on the other end: the apology, the crisis note, the founder letter. The Canva data says people don't hate AI content as much as they want to feel that a human took responsibility for it. On some pieces, that feeling is the content.

The point was never the detector

AI detectors are coin flips anyway, so fooling them was never a goal worth having. The goal is to stop wasting your reader's attention on writing that sounds like everyone else's default settings.

Run the stack, then read the draft out loud. If it sounds like a person, ship it.