Zero AI Slop

Zero AI Slop – a quality check for prose

What this is: a quality filter that catches the patterns that make writing sound generated, whoever wrote them. It is not an AI detector – it never claims a text was or wasn't written by AI; detection is unreliable and accusations are not this skill's job. It judges the writing, not the writer.

Core belief: slop is not a vocabulary problem, it's a specificity problem. The tell words matter, but the deeper tell is writing that gets vaguer as it gets louder – importance claimed, nothing shown. The fix is always the same direction: replace the inflation with a specific, or cut it.

The two laws (read before doing anything)

1. Preserve the voice. Rewrite in the author's register – their sentence habits, their formality, their quirks. Keep the meaning. Never delete substance to escape a pattern. A slop-free rewrite that also lost the author's personality is just different slop.
2. Flag a context-dependent pattern only when it repeats; never rewrite a sentence that already sounds human. One em dash is punctuation. Five in a paragraph is a tell. When a sentence reads naturally aloud, leave it alone even if it technically contains a listed pattern.

The three modes

Pick by what the user asked for:

An explicit ask always wins. If the user supplies a draft with no verb, ask whether they want it cleaned or just checked.

The pattern library

Three tiers. Tier 1 is rewritten on sight. Tier 2 is rewritten only when the pattern repeats within the piece (law 2). Tier 3 is judged over the whole piece, not per sentence.

Tier 1 – hard bans (rewrite on sight)

Vocabulary – these words in their figurative or abstract sense are tells. Replace with the plain word for what is actually meant:

delve · tapestry · testament · underscores · showcases · pivotal · crucial · foster · garner · vibrant · landscape (abstract) · interplay · intricate · seamless · robust · leverage (verb) · elevate · unlock · unleash · empower · supercharge · harness (figurative) · embark · journey (figurative) · realm · navigate (figurative) · boasts · nestled · groundbreaking (figurative) · game-changing · cutting-edge · transformative · revolutionize · holistic · synergy · myriad · plethora · ever-evolving · enduring · profound · invaluable · beacon · cornerstone · treasure trove · deep dive

Stock constructions – delete or state plainly:

• "Thrilled / excited / humbled to announce"
• "It's not just X – it's Y" (and "isn't just about", "more than just")
• "stands as / serves as a testament to"
• "plays a vital / crucial / key role in"
• "In today's fast-paced world / digital age / ever-evolving landscape"
• "In a world where…"
• "The best part?" · "Here's the kicker:" · "Let that sink in."
• "Let's dive in" · "Let's unpack this"
• Chatbot residue: "I hope this helps" · "Certainly!" · "Great question" · "Would you like me to…"
• Knowledge-cutoff hedges: "as of my last update", "while specific details are limited"

Tier 2 – context-dependent (rewrite only when repeated in the piece)

• Em dashes – one or two across a piece is punctuation; several per paragraph is a tell. When rewriting, restructure the sentence; don't just swap the dash for a comma.
• Rule of three – "fast, simple, and effective." Once is rhythm. As a habit, it's a metronome; break one of the triples into a sentence that says something.
• Negative parallelism – "It's not only X, but also Y."
• Connective sentence-openers – Additionally, Moreover, Furthermore, "In conclusion". Most sentences don't need an on-ramp.
• Elegant variation – rotating synonyms for a term the piece should just repeat. Locked terms stay locked.
• Bold overuse – bolding as decoration, or symmetric "Label: explanation" list items marching in identical shape.
• Rhetorical-question transitions – "So what does this mean?"
• False ranges – "from startups to Fortune 500s" pretending two examples are a spectrum.
• Title Case Headings – sentence case reads human.
• Emoji as bullet decoration.
• Hedging doublets – "could potentially", "may possibly."

Tier 3 – structural (judge the whole piece)

These are the deepest tells and the real work of this skill:

• Significance inflation. Importance claimed, nothing specific shown – "this marks a pivotal moment." Fix: state the concrete fact that makes it matter, or cut the claim.
• Generic-positive smoothing. An adjective standing where a specific belongs – "a powerful tool" (which tool? does what?). Fix: every claim carries a specific – a number, a name, a mechanism, an example – or it goes.
• Participial padding. Sentences trailing off in ", highlighting…", ", ensuring…", ", reflecting…", ", fostering…" – analysis-shaped clauses that add nothing. Fix: cut the clause. The sentence was done.
• Vague attribution. "Experts say", "studies show", "industry reports" with no named source. Fix: name the source, or mark [SOURCE NEEDED] and tell the user – never leave an unnamed authority standing, and never invent a citation.
• Empty contrast. The "not X" half of a sentence whose point already landed – "real work, not sitting through slides." Fix: cut the comparison unless the contrast IS the point.
• Symmetric scaffolding. Every paragraph the same length, every section a header plus three bullets, every list item the same grammatical shape. Fix: break the grid where the content doesn't need it.
• Hedged non-endings. "Despite these challenges, the future looks bright"; conclusions that restate the piece; both-sidesing that refuses to land on a position. Fix: end on the last real point. If the piece has a position, the ending states it.
• Mind-reading openers and manufactured reversals. "You've probably wondered…", "But then everything changed."
• Metronome rhythm. Every sentence the same length. Human writing is uneven; let it be.

Mode instructions

Draft mode

1. Write the piece the user asked for. Know the library while writing – the cheapest fix is the pattern you never type.
2. When the draft is done, run the check: pass over the full text against all three tiers. Rewrite every failure per the two laws.
3. Deliver the clean draft, ending with the receipt (below). The user never sees the unchecked version.

Clean mode

1. Read the whole piece first and infer the author's register from the parts that already sound human – sentence length habits, formality, dialect, vocabulary. That register is what you rewrite into. Never Americanize or normalize dialect, and never strip a stylistic quirk that is clearly the author's.
2. Pass over the text against all three tiers. Rewrite flagged passages in place, per the two laws.
3. Deliver the full cleaned text (copy-paste ready, never a list of patch instructions), ending with the receipt.

Audit mode

1. Pass over the text against all three tiers.
2. Report each finding: line reference · the quoted text · the pattern name · the suggested fix in one line.
3. Change nothing. End with the one-line verdict: how much of the piece is affected and whether the problems are word-level, structural, or both.

The receipt

Every Draft or Clean delivery ends with one line:

Slop check: caught and rewrote N patterns.

If the user asks what was caught, list the changes as before → after pairs with the pattern name. Don't volunteer the full list unprompted – the receipt is the proof, the detail is on request.

A checked draft that comes back with zero catches gets the receipt too ("caught 0"). An honest zero builds more trust than a padded count.

What this skill never does

• Never changes facts, claims, numbers or names while rewriting.
• Never shortens for its own sake – cutting slop shortens naturally; that's the only shortening allowed.
• Never invents a source to fix a vague attribution.
• Never flattens a human sentence to satisfy a rule (law 2).
• Never claims a text was AI-written.

Pattern sources: several tells here were first catalogued in Wikipedia's field guide "Signs of AI writing"; they are reworded and operationalized here, with additional patterns from editing practice. That page is worth reading in full.