Kill the Slop: The Patterns That Give AI Writing Away
Constraint, not personality. Judgment, not tics.
Slop is not a vibe. It is a set of nameable, repeatable patterns. You don't fix it by asking for "more human" writing — you fix it by forbidding the patterns the model reaches for by default. This guide governs what gets generated, held to a stricter bar than your own keyboard. A pattern doesn't earn a pass just because you write that way sometimes. And removing slop must never strip the genuine connection to the reader. Clean but cold is its own failure.
Connection earns the sell
Every public-facing piece moves through these four stages in order. Skipping ahead — selling before the reader feels understood — is the most common failure.
Connection
Lead with a specific, real pain the reader recognises in themselves.
Awareness
Show the burden is structural — they don't have to carry it alone.
Desire
Desire from relief, clarity, agency — not hype or slogans.
Sell
The product appears as the guide, after the reader is the hero.
Slop has fingerprints
Each one is a structural fingerprint of generated text. The example shows the move so it can be recognised and avoided. Any of these requires a rewrite before publishing.
Synthetic scene stack
One concrete detail is allowed only if it proves the problem. Two is risky. Three fails.
Fragment cadence
Fragment-style emphasis may appear once per piece, total — never in the opening.
Fake contrast pairing
Contrast belongs in a complete sentence, not a fragment pair.
Rhetorical question chain
One question, only at a true turn. Consecutive questions fail.
Explainer symmetry
Category comparison uses complete-thought paragraphs, except on a landing page.
Polished universal claim
Open with observed, specific pressure — not universal emotional certainty.
Transition tissue
If the paragraph still works after you delete the transition, delete it.
Tidy recap ending
End on the final useful point. No recap paragraph.
What every piece is judged on
In order. Pattern control overrides personal style whenever style creates AI-pattern risk.
Pattern control
Free of the AI fingerprints. Overrides style when style reads as generated.
Argument pressure
Every paragraph carries a claim, evidence, consequence, or turn — never mood.
Specificity
Real nouns, real numbers, real constraints. No "operational friction."
Judgment
Reflects how you think — directness, skepticism — not a performance of personality.
Words that announce a machine
A ready-to-use starter set. Do not use these unless directly quoting a source — and add your own industry offenders.
- In today's fast-paced world
- It's important to note
- At its core
- X isn't just about Y, it's about Z
- That being said
- To summarize
- Key takeaways
Shape, not just words
Default to paragraphs of 2–3 complete sentences. At most one emphasis move in the whole piece. Mark section turns with subheads, quotes, or stats — never with fragment cadence or white space.
Answer before publishing
Any failed answer means a rewrite before a human ever reads it.
| Question | Must be |
|---|---|
| Does the opening connect to a real, specific pain before it teaches or sells? | yes |
| Does the piece move through the journey model in order? | yes |
| Is the reader the hero and the product the guide? | yes |
| Does it use more than one emphasis move? | no |
| More than two one-sentence paragraphs in the opening? | no |
| Does any paragraph exist only for rhythm, mood, or transition? | no |
| Does the ending recap instead of landing? | no |
| Did the rules get applied so hard they removed the connection? | no |
"Copy your tics, get a caricature. Copy your judgment, get your voice."
Run this automatically: brand-voice-noai plugin
The eight failure patterns, the banned-word list, and the review gate on this page — automated. A Claude Code plugin that ships a kill-the-slop skill, an on-demand /slop-check linter, an enforcement hook that runs on every .md write, and a judgment-level reviewer agent for the patterns a regex can't catch.
Now put it into a working file
You know the patterns to forbid. The companion piece is the procedure: a Claude Project, two files, and a prompt — the actual workflow for turning this catalog into a constraint file your team uses on every generation.
