How to remove AI slop from your writing

What is AI slop?

AI slop is any content that is generated that looks or sounds like it was puked out of a data center. In writing, most AI tools reach for the same small set of words and build their sentences the same way every time, so content starts to feel generic no matter what it is about or who is publishing it. Your readers might not be able to point at the exact problem, but they can feel that something is off. It makes your words sound like everyone else's and for many, instantly reduces trust.

For a maker or a small business, your success is built off of trust in your brand (among other things). Your listings, emails, and posts are how customers hear your voice, and that voice is a big part of why they pick you over the next option. When the writing sounds generic, you start to sound generic too, and the trust you worked hard to build is vaporized.

Why it is worth fixing

The good news is that you do not have to become a better writer to fix this. You only have to learn which words and patterns give the game away, and then get into the habit of cutting them. Using AI to help your draft your work is fine. Publishing it without giving it a second glance these days is effectively taking a shortcut in your business and it will show. The same phrases that make a draft feel polished are usually the ones that make it feel fake, so taking them out is the quickest way to sound like a person again.

Everything below I would encourage you to start using on your very next email.

The words and phrases to cut

The lists below are grouped by the job they are performing. Once you see them, they will start to become obvious. Whenever one of these shows up in your writing, pause and ask what you were actually trying to say, and then consider writing a plainer version in its place.

The first group of phrases represent claims that lean on a source that isn't cited, which is a fast way to lose a reader's trust:

studies show          research shows        research suggests
experts say           experts agree         scientists say
data shows            statistics show       it's well known
everyone knows        many believe          some say
people say            sources say           countless studies

The next group tells the reader that a point is a big deal, instead of just making the point and letting them decide for themselves:

it's worth noting     worth remembering     important to note
important to remember the part that matters which is exactly why
cannot be overstated  and that matters      the whole point

These are the filler lines that order the reader to pay attention, instead of earning attention with something that is actually worth reading:

let that sink in      read that again       make no mistake
let me be clear       let's be honest       let's be real
truth be told         needless to say       buckle up
you're not alone      look no further       we've got you covered

Next we have fake-casual openers and empty transitions that pad a paragraph out without adding anything of substance to it:

here's the thing      here's the kicker     the truth is
the reality is        what most people miss what nobody tells you
that being said       at the end of the day the bottom line
first and foremost    last but not least    in conclusion

This group is the corporate and hype vocabulary that makes a small shop sound like a big faceless brand:

game-changer          paradigm shift        the future of
cutting-edge          state-of-the-art      best-in-class
unlock the power      supercharge           revolutionize
transformative        move the needle       think outside the box
low-hanging fruit     circle back           boils down to

And these are the single words that most readers now spot right away as a sign that a machine did the writing instead of you:

delve      seamless     seamlessly    robust      leverage
streamline holistic     multifaceted  underscore  underscores
resonate   tapestry     a myriad of   a plethora  treasure trove
at the heart of         in the realm of         shed light on
ever-evolving           pave the way            a double-edged sword

Two more shapes tend to sneak in, and they can sometimes be a little harder to spot:

it's not just about X, it's Y     (the "not just" escalation)
honest truth, in all honesty      (using "honest" or "real" to prop up a claim)

The shapes a word list cannot catch

A few of these habits live in the shape of your sentences rather than in the words, so no list of banned phrases will ever flag them. You have to catch these ones with your own eye.

Start with the em dash, which is the long dash that some tools love to sprinkle everywhere. Take it out and use a comma, or start a fresh sentence with a period. You often lose nothing by dropping it.

Next, watch for patterns where you name something, add a colon, and then follow it with a neat little list. Doing that once is fine, but a whole page built that way feels machine-made even when no single line looks wrong, so keep mixing up the way you build your sentences.

Another habit is writing as if the reader already saw whatever you are responding to - as if they have awareness of research that the AI tool did for you. They obviously did not see it, so spell out the point in your own words first, and only then react to it. If a sentence only makes sense to someone who read the same research sources you did, then your content is flawed and it will stand out like a sore thumb.

Last, cut any line that is only there to sound deep. A tidy, clever-sounding sentence that flips at the end can feel wise while saying almost nothing. This is incredibly common in newsletters and social media posts (LinkedIn is full of them). So, if a line is putting on a performance instead of making a point, delete it and move on.

How do I check my own writing?

Before you send anything, spend two minutes and run this quick review over your draft.

1. Read the entire draft out loud, and if you would never say a sentence to a customer's face, either cut it or rewrite it.
2. Scan for the words in the lists above, and replace each one with the point you were making.
3. Remove every em dash, and undo any label-and-list arrangement created purely for effect.
4. Confirm that each claim stands on its own, without requiring the reader to have seen your source.
5. Read it a final time for flow.

Build a simple helper that flags slop for you

You do not have to carry this entire list around in your head. You can teach your AI tool to run the review for you a single time, and afterward it will search for the same problems across everything you write.

A helper like this is simply a saved set of instructions that your tool follows every time, and most tools already include a place to store one. In ChatGPT it lives under custom instructions or a saved project, in Claude it lives inside a project, and in most writing apps it is a saved prompt. You do not need any coding knowledge to arrange this, because you are only pasting the rules in one time.

Copy the block below into whichever spot your tool provides.

You are my editor, and your one job is to catch AI slop and fix it.
 
AI slop is writing that leans on a small set of tired phrases and shapes instead of saying the plain thing. Whenever you spot any of these, rewrite the sentence to say what it was reaching for:
 
- Borrowed authority: studies show, experts say, research suggests, data shows. Attribute the claim to a source you can name, or cut it.
- Rating a point instead of making it: it's worth noting, important to note, the part that matters, cannot be overstated.
- Filler that orders the reader to care: here's the thing, the reality is, at the end of the day, the bottom line, make no mistake.
- Hype vocabulary: game-changer, paradigm shift, unlock the power, move the needle, think outside the box.
- Single-word tells: delve, seamless, robust, leverage, streamline, holistic, underscore, resonate, at the heart of.
- The "not just" escalation: it's not just about X, it's Y.
 
Treat anything in the same family as these examples as slop too.
 
Then fix the shapes a word list misses:
- Use a comma or a period in place of any long dash.
- Turn a label, a colon, and a tidy list into plain sentences, and keep your sentence shapes varied.
- State every claim in full and in your own words, so a reader who never saw my source still follows it.
- Keep the language plain, so a busy person reads it without slowing down.
 
When you finish, return the cleaned draft, then a checklist that names every phrase you changed and every claim you attributed or cut. Keep going until every one is accounted for.

From then on, you paste your draft in and let the tool take the first pass. It catches the obvious slop for you, and you still make the final call on how you want to sound.

Turn it into a reusable skill with Claude

If you use Claude, you can go one step further and package these rules into a skill. A skill is a small bundle of instructions that Claude loads on its own whenever a task calls for it, so you never have to remember to paste anything again.

You do not have to build that bundle by hand. Anthropic provides a skill builder that does the assembly for you. You describe what you want the skill to do, hand it the list from this guide, and it writes and saves the skill for you. The whole process takes a few minutes and needs no code.

To start, open Claude and tell it you want to create a skill that removes AI slop from your writing. Paste in the rules from the block above as the material, and let the builder walk you through the rest. The one piece to get right is the description, which is the line that tells Claude when to reach for the skill, so write it to name the trigger in plain words, something close to "Use when I ask to check, clean, or edit a draft for AI slop." With that line in place, you can simply ask Claude to look over a draft, and it opens your slop skill on its own. Anthropic's guide to building skills covers each step if you want to follow along.

The best version you can build is a deterministic checker

An AI helper is a strong first pass. However, an AI model runs a little differently every time, so it will catch every slop word on one draft and miss two on the next - it may even insert worse slop than what you started with. A deterministic tool/checker does not guess at all. It compares your draft against a fixed list and flags the exact same words every single time, which is the only way to know for sure that a page is clean.

I am finalizing that checker now, as a small tool you will be able to download and run on your own writing. The Mountain Labs newsletter is where I will announce it first, so sign up below and you will get it the day it ships, along with the weekly newsletter.

And if this guide helped you, please send it to someone else who writes and fights this same battle. Most people have never seen this list, they know they struggle with it, but they don't have a great solution. A single share can save a friend hours of second-guessing their own words and we can, together, remove a little more slop from the internet one post, newsletter, or email at a time.