1. The three tiers
Not every "AI word" is the same problem. We sort them into three buckets, because merging them into one list blocked ordinary human phrasing.
| Tier | What it is | What to do |
|---|---|---|
| Machine register | Words a model overproduces and a person rarely types | Never ship it |
| Cliche | Empty corporate filler, human but weak | Fine in email, banned in a resume bullet |
| Formula | Stiff or dated register | Your call, it is a style note |
2. Machine register: block these
delve, foster, leverage, utilize, facilitate, empower, streamline, robust, comprehensive, seamless, holistic, ensures, calibrate, timely, exceptional, genuinely, significant, resonates, synergy, keen, thrilled, delighted, honored, spearheaded, cutting-edge, paradigm shift, game changer, tapestry, realm, beacon, multifaceted, meticulous, intricate, paramount, transformative, embark, supercharge, ever-evolving, testament, underscore, showcase, garner, interplay, pivotal, groundbreaking, renowned, nestled, breathtaking.
Also the inflections. "Leveraging" and "fosters" are the same word.
Fancy ways to write "is"
serves as, stands as, boasts, features. Write "is" or "has".
Weasel attribution
experts agree, studies show, industry reports suggest, many argue, widely regarded as. Name the source or cut the claim.
Chatbot leftovers
I hope this helps, great question, of course, certainly.
3. The tells that are shapes, not words
A word list will never catch these. Every word in them is ordinary.
The mirrored contrast
"It's not a tool, it's a teammate." Also "not just X, but Y", "X isn't about A, it's about B", and "less about X, more about Y".
A model reaches for this to sound insightful when it has one flat point.
FixDelete the negated half. "It's a teammate."
The trailing -ing clause
"The launch adds file search, underscoring the team's commitment to better workflows."
The clause explains nothing.
FixWrite the consequence. "...so users find old drafts without leaving the editor."
Negative listing
"Not a vendor. Not a contractor. A partner."
FixWrite the claim once.
The dropped subject
A model writes "Glad to connect." A person writes "I'm glad we connected."
The brisk register is the tell.
FixPut the subject back. Mid-sentence it is fine. "Either way, glad to be connected" reads human.
The colon reveal
"The detail that makes it work: a separate agent grades it."
Colons are for lists and labels, not drama.
FixWrite it as a sentence.
Faux-insight setups
"Here's the thing." "What most people get wrong." "Here's what nobody tells you."
These flatter the writer.
FixCut the setup, make the claim stand alone.
Summary endings
"In conclusion." "Ultimately." "Overall."
The reader was just there. End on the last concrete point.
Fake-profound kickers
The final deep line that turns the point into an aphorism. Delete it. Do not write a better metaphor, delete it.
4. Punctuation and formatting
- No em dashes. This is the single fastest tell. Use a period or a comma. Reaching for parentheses instead just trades one tell for another.
- No curly quotes. Straight quotes.
- Sentence case headings. Not Title Case.
- No emoji in headings.
- No bold sprinkled mid-sentence.
- No bullet list where two sentences of prose read better.
5. The two tests that catch the rest
The portability test
If a sentence could move unchanged to another person, company, product or country, it is filler. "The integration improved efficiency" fits anywhere. "The integration cut deploy time from 40 minutes to 4" fits one place.
The cadence test
Read the longest sentence aloud. If it needs a breath in the middle, split it. Then check whether three sentences in a row are the same length. Metronomic rhythm reads as machine output even when every word is clean.
6. What not to overcorrect
Stripping the patterns is half the job. Voiceless, sanded-down prose is just as obvious as slop.
- Keep your own opinions. React to a fact instead of listing both sides.
- Keep "I think" and "to be honest" when they carry real uncertainty.
- Keep the blunt line, the joke, the digression.
- Vary sentence length on purpose.
- Over-editing strong human writing is a failure. Sometimes the right edit is no edit.
7. The five skills this draws on
We read all five and did not install any. The rules above are the union of what they agree on, plus what we learned writing recruiter email.
| # | Skill | Author | Repo |
|---|---|---|---|
| 1 | stop-slop | Hardik Pandya | github.com/hardikpandya/stop-slop |
| 2 | no-ai-slop | Peter Yang | github.com/petergyang/no-ai-slop |
| 3 | humanizer | Siqi Chen | github.com/blader/humanizer |
| 4 | unslop | Cursor | github.com/cursor/plugins |
| 5 | slopkit | Rasty Turek | github.com/ehmo/slopkit |
Install any of them with the Skills CLI:
npx skills add hardikpandya/stop-slop
8. One last check before you send
Ask the draft one question: what makes this obviously AI generated? Then fix what you find. That single pass catches more than any list.