Unslop

Text drafted by a model tends to carry a fingerprint: hedged claims, inflated adjectives, the same handful of transition words, and a rhythm that never varies. None of that is wrong sentence by sentence. Stacked up, it reads as generated rather than written. This skill is a pass to find that fingerprint and remove it, then put back the specificity and opinion that made the writing worth reading in the first place.

When to run this

• Before sending a PR description, doc, or write-up that a model helped draft.
• On any text where a reader would notice it "sounds like AI" — that reaction is a real signal, not vibes.
• Skip it for throwaway internal notes where nobody will read closely, and skip it for code comments (those have their own bar: only write one when the why isn't obvious).

Process

1. Read the text once straight through, flagging every sentence that trips one of the tells below.
2. Rewrite flagged sentences. Keep the meaning and the intended tone — this is not a rewording pass for its own sake, only fix what's actually broken.
3. Reread for voice: does a specific person seem to be saying this, or could it have been assembled by anyone about anything?
4. Last check: pick the single most generic-sounding sentence left. If you can't sharpen it, cut it.

The tells

Inflated framing. Phrases that make something sound bigger than it is without adding information: "a testament to", "stands as a pivotal moment", "marks a significant milestone", "an ever-evolving landscape". Say what happened in plain terms instead.

Weasel attribution. "Experts agree", "many believe", "it is widely understood" — with no expert named. Either cite who said it or drop the claim.

Hedge stacking. Multiple qualifiers doing the job of one: "could potentially somewhat improve" collapses to "may improve" or, better, a number.

Filler connective tissue. "In order to" → "to". "Due to the fact that" → "because". "It is worth noting that" and "it is important to mention that" almost never need saying — delete the frame and state the fact.

The stock word list. A small set of words shows up constantly in model output because they're safe and vague: delve, boast, robust, seamless, leverage (as a verb), utilize, facilitate, foster, underscore, showcase, elevate, unlock, tapestry, landscape (used abstractly), realm, journey (used metaphorically for a process). Swap in the plain word: use, help, show, limit, or the specific mechanism.

Rule-of-three padding. Forcing every list into exactly three items whether or not three is the right count. Use however many the content actually needs.

"Not just X, but Y." A construction that manufactures drama out of a simple statement. Say Y directly.

Synonym rotation. Calling the same thing by three different names in one paragraph to avoid repetition — "the tool," "the utility," "the application" — when repeating the actual name would be clearer. Pick one term and reuse it.

Metaphor-as-jargon. Abstract nouns borrowed to sound technical — substrate, vector, surface (for "API surface"), north star, flywheel, primitive (as a noun), scaffolding used loosely. If there's a concrete word for the actual thing, use that instead.

Passive voice hiding the actor. "Requests are validated before processing" tells you nothing about who validates them. "The gateway validates requests before forwarding them" does. Use passive only when the actor genuinely doesn't matter.

Adverb crutches. "Runs significantly faster" wants a number or a comparison; if you have neither, the claim probably isn't ready to make.

Punctuation tics. Overusing em dashes as a substitute for periods or commas, colons used mid-sentence as a connector rather than to introduce a list, and bolding every proper noun or key term until nothing stands out because everything does.

List-itis. Turning a paragraph of connected reasoning into a bullet list of fragments, or a bold label followed by a colon and a sentence that just restates the label ("Performance: Performance is much better now."). Prose reads better than a list when the ideas depend on each other; save lists for genuinely parallel, independent items.

Assistant-speak. "I hope this helps!", "Let me know if you have questions!", "Great question!", "Of course!" — leftover politeness formulas from chat interfaces. In written docs and PR descriptions there's no one to be polite to; just state the thing.

Generic closers. Ending on a sentence that could close any piece of writing about anything: "the future looks bright," "this opens up exciting possibilities." Close on the actual next step or the actual result instead.

Putting a voice back in

Removing the tells above gets you to neutral, not to good. Neutral, voiceless writing is its own tell. To finish the job:

• Take a position. If something is a bad idea, say it's a bad idea instead of listing "considerations."
• Let sentence length vary. A short sentence lands harder right after a longer one that set it up.
• Say "I" when the sentence is about a choice you made — hiding behind passive voice to sound objective usually just sounds evasive.
• Keep one telling specific over three generic ones. "The build takes eleven minutes" beats "the build takes a while."
• A little roughness — an aside, a mid-thought correction, an opinion that doesn't fully resolve — reads as human. Manicured, uniform paragraphs read as generated.