Anti AI slop writing

Applies to any prose before it ships. Produces copy with the recognisable AI patterns removed, in two disciplined passes.

The core idea: AI writing fails in two different ways, and each needs its own pass. Some patterns are rhetorical moves with no human upside - they are wrong every time they appear. Others are ordinary words and structures that read fine once but scream machine when they cluster. Scan for each separately.

This skill is the subtractive half of a two-pass writing standard. It takes tells out. Its companion, the human-mannerisms pass, does the opposite job - it puts voice in, so clean copy does not read flat. They are deliberately kept as two skills, not one, because the two jobs do not always both apply. Every piece that ships should have its AI tells removed; not every piece should be made to sound casual. A formal announcement, a compliance note, or a technical one-pager needs to be clean, not chatty - run this pass and stop. A LinkedIn post, a founder's thought leadership, or a personal email needs voice on top - run this pass first, then the mannerisms pass. Bundling them would force a casual register onto writing that should stay plain. Keeping them separate lets you apply the floor everywhere and the texture only where it fits.

The two passes

Run both, in order, before returning any prose.

Pass one - zero-instance patterns. Scan for rhetorical moves that are a tell every time. Any hit gets rewritten, not softened. The full catalogue is in references/zero-instance-patterns.md; the ones that slip through most often:

1. Negative parallelism - asserting what something IS by first saying what it ISN'T about the same subject: "It's not X, it's Y", "Not just X, Y", "We don't just X, we Y". The test: remove the negated half. If the point still stands, the negation was scaffolding - cut it and state the claim directly.
2. False-suspense transitions - "Here's the thing", "Here's what nobody tells you", "Here's what changed". The whole class, not set phrases. Does it name the revelation in the same beat, or just tease one? If it teases, cut it.
3. Self-answered rhetorical questions - "The result? Devastating."
4. The colon setup - the "abstract noun, colon, payoff" cadence ("The reality:", "The takeaway:"). Banned as a device; state the point as a plain sentence.
5. Formulaic openers and closers - "In today's fast-paced world", "At its core", humblebrag announcement openers ("Thrilled to share"), and "In conclusion", "Ultimately", "At the end of the day".
6. Patronising analogy - "Think of it like...", "It's like...". Even one forced analogy is too many; name the thing directly.
7. Manufactured stakes - grandiose inflation ("this reshapes everything") and phantom-future projection ("a year from now you'll wish you'd..."). Stakes come from a present cost the reader already pays.
8. Invented specificity - numbers, dates, percentages, or named moments not grounded in the source. If the source has no number, use no number.

Pass two - clustering patterns. Scan for vocabulary and structure that read fine once but signal machine when stacked. The full lists are in references/clustering-and-formatting.md. The shape of the rule: a single tricolon, one em dash, one strong adjective is human; three tricolons, an em dash in every sentence, and a wall of grandiose nouns is not. Watch grandiose nouns (tapestry, landscape, realm, ecosystem), pompous verbs (delve, leverage, unlock, foster, elevate), magic adverbs (deeply, seamlessly, fundamentally), anaphora (same opener three-plus times), uniform sentence and paragraph length, bold-first bullets, and emoji or decorative-unicode sprinkling.

Tune it per brand

Every team adds its own hard rejects on top. A starter set worth adopting: no em dashes in shipped prose; a house banned-word list (transformative, game-changing, seamless, robust, best-in-class, "excited to announce"); and no sales-pitch pivot, meaning a thought-leadership piece never closes by circling back to how early the company was or how the market is catching up to it. See references/clustering-and-formatting.md for how to build and apply the house list.

What good looks like

• What the best editor notices first: the single most common tell by far is negative parallelism, and it hides in dozens of shapes ("less X, more Y", "it feels like A but it's actually B"). Learn the semantic move, not a phrase list, and it stops slipping through.
• The common mistake: over-correcting into stilted prose. Real humans use tricolons, analogies, and the occasional em dash; AI overuses them. Stripping every one leaves copy that is clean but robotic in a new way. Pass two is "avoid clustering", not "remove all instances". Do not sand the writing into something no person would say.
• How you know it worked: the pub test - would two colleagues say this line to each other over lunch? And the ship test - would the author post this without feeling they need to de-AI it first? If either fails, the fix is almost always to name the specific thing a vague phrase is gesturing at.

Rules

• MUST run both passes. Pass one is zero-tolerance; pass two is anti-clustering.
• MUST rewrite every zero-instance pattern, never soften it.
• MUST NOT invent specifics to replace vague copy - cut the number, do not fabricate one.
• NEVER over-correct into stilted, voiceless prose; a clean-but-flat draft is not finished, it is ready for a human-texture pass.