Imported from spuvr/humanizer (SKILL.md). Install upstream with npx skills add spuvr/humanizer. Copyright stays with the author.

Humanizer

The goal

Take prose that sounds like AI wrote it and turn it into prose that sounds like a person wrote it. Not with a banned-word filter or a before-and-after template, but by understanding deeply what human writing actually feels like, so you catch the artificialness anywhere it shows up. Even in cases nobody specifically described to you.

Read the whole thing before you start. Let it sit. If you skim for highlights you'll end up running a checklist, and a checklist is what this skill exists to replace.

The one idea everything else flows from

Human writing trusts the reader. AI writing doesn't.

That's the whole thing. Every other pattern, every padded sentence, every rule-of-three cadence, every closing recap, all of it is a symptom of the same root cause. The writer (or the model) didn't trust the reader to follow along, so they over-explained, padded, hedged, restructured, etc... The reader could feel it. They got bored. They left.

If you internalize this one idea, the rest of the skill makes sense by itself. You don't need to memorize patterns because you can derive them.

A check that keeps you honest: every time you're about to edit, ask, "Am I trusting the reader, or babysitting them?" If the answer is babysitting, cut.

Another way of seeing the same idea. AI writing wears a suit and tie even when the conversation is at a kitchen table. Human writing matches the room. If the prose in front of you feels like it's wearing a suit at a backyard barbecue, that's the tell.

What human writing actually feels like

Before you can fix AI-ness, you have to know what you're aiming for.

Rhythm. Sentences are different lengths. Some short. Some longer, looping back on themselves. The variation isn't decoration, it's how reading feels alive. "The product is fast. It is reliable. It is innovative." reads dead. "The product is fast. Genuinely fast, the kind of fast you only notice once you go back to the old one." has a pulse.

Flow. One sentence leans into the next. Think about how a friend tells you about a car breakdown versus how a manual does. The friend moves. The manual plants its feet between every sentence and waits.

Space. Sets something up and trusts you to land. The most powerful sentence is often the one not written. AI is allergic to that space; it wants to close every loop. Don't.

Voice. Somebody is there. "This is impressive" is neutral. "This is impressive in a way that kind of unsettles me" has a person behind it.

Specificity. "3am" instead of "late at night." "The four people who actually open this every day" instead of "many users." "Since 2019" instead of "for a long time." Specifics cost something to write; that's what earns them trust.

Plain English. Use the everyday word, not the fancy one. "Use" instead of "utilize." "Help" instead of "facilitate." "Show" instead of "demonstrate." "Try" instead of "endeavor." Aim for the kind of English everyone understands, not the kind of English that wants to impress a committee. Fancier vocabulary feels smart but it slows the reader down, and it signals AI almost instantly. Plain words are not lower quality, they are higher trust. Write like the smartest person you know talking to a friend, not like an academic paper.

Mess. A half-thought. A tangent that earns its place. A sentence that ends somewhere you didn't expect. Don't sand off every edge that survived your first draft.

Stake. Something the writer cares about. They want you to see what they see. AI reports, balances, moves on. Real writing has a temperature.

Opens in a moment. Human writing tends to drop you into a specific time, place, or experience. "Last winter the pipes burst at 3am, and..." "The first time I tried this..." "Coffee starts with a goat." "On the train back from Berlin..." Concrete location before any claim. AI defaults to topic sentences or grand frames ("In a world where...," "At its core..."). The opener sets the temperature for the whole piece.

A little funny when it fits. Not joke-funny, but with angles. A turn of phrase that catches you. A moment that almost makes the reader smile because they discovered something.

How to spot AI-ness when you see it

You're not running a regex. You're reading like a person and listening for the moment where the prose stops sounding like a mind and starts sounding like an output. When that hits, slow down and ask what made it land that way.

What follows is a starter set of moves grouped by the shape of failure underneath them. The same shape shows up in a hundred costumes; your job is to recognize the shape, not the surface.

When the writer doesn't trust the reader

The over-explainer. Says the conclusion, then says what it means, then says why it matters. Fix: say it once and let it land. Before: "Our app is fast. By fast, we mean it loads quickly. This matters because users don't like waiting." After: "Our app loads fast. Nobody likes waiting."

The padder. Phrases that fill space without carrying meaning. "It's worth noting that," "additionally," "furthermore," "moreover," etc... Fix: delete and the sentence stands harder. Substitutions: "In order to" becomes "To." "Due to the fact that" becomes "Because." "At this point in time" becomes "Now." "Has the ability to" becomes "Can." "It is important to note that" usually just disappears, etc...

The closing recap. Section ends with a sentence summarizing what the section just said. Fix: trust the reader. Cut it.

The fragmented header. Heading followed by a one-line restatement of the heading before the real content begins. Fix: let the heading do its job. Start with the actual content.

The excessive hedge. "It could potentially possibly be argued that the policy might have some effect." Fix: one qualifier maximum. "The policy may affect outcomes." Sometimes drop the hedge entirely.

The knowledge-cutoff disclaimer. "While specific details are limited..." "As of my last update." Chatbot artifacts in the page. Fix: if you have the fact, state it. If you don't, write a sentence that doesn't need to mention what you don't have.

The fake balance. "On one hand, on the other hand," with no actual position. Performance instead of substance. Fix: have an opinion, or admit you're genuinely unsure.

When the rhythm gives it away

The rule of three. Three benefits, three considerations, three reasons, etc... Fix: list two, or four, or one. Before: "fast, flexible, and reliable" is the giveaway. After: "fast, and weirdly flexible once you stop fighting it."

The stop-start rhythm. Every sentence is its own island. Fix: rebuild the connections. Before: "The team launched the feature. It included three improvements. Users were excited." After: "The team launched the feature with three improvements, and users were on it within an hour."

The em-dash habit. AI uses "—" for everything. Fix: very rare. When you do use one, write a plain hyphen ("-"). Most of the time a comma, period, or restructure works better. A page should almost never need more than one or two.

The synonym cycle. Calling the same character "the protagonist," "the main character," "the central figure," "the hero" in four sentences. Fix: repeat the word. Repetition is less distracting than the synonym carousel.

The forced takeaway. Paragraph closes with a moralizing line that lifts an observation into a Life Lesson. Fix: let the observation be the observation. If a real lesson exists, the reader will arrive at it.

When abstraction replaces substance

The abstract noun parade. "Intricate landscape," "evolving tapestry," "pivotal moment," "rich heritage," etc... Fix: name the thing. If you can't, the sentence isn't ready.

The promotional voice. "Stunning," "vibrant," "breathtaking," "must-visit," etc... Brochure prose. Fix: describe instead of selling. Before: "Stunning beaches." After: "The kind of beach where the sand squeaks under your foot."

The pivot phrase. "It's not just X, it's Y." Fix: if Y is what you mean, just say Y.

The unearned weight. "In a world where...," "At its core...," "What truly matters is...," etc... Fix: start with the thing itself. Weight comes from what you say, not how grandly you frame it.

The false range. "From X to Y, from A to B" where X and Y aren't on a real scale. Fix: just list the things, or pick one and write about it for real.

The persuasive authority trope. "The real question is," "at its core," "what really matters is," etc... Fix: drop the framing and say the thing.

The vague authority. "Experts say," "studies have shown," "many believe," etc... Fix: name the source, drop the claim, or rephrase as your own view.

The present-participle tail. Sentences ending with an "-ing" clause for fake depth. Before: "The temple was built in 1842, reflecting the community's deep connection to the land." After: "The temple was built in 1842."

The copula avoidance. "Serves as," "stands as," "marks," "represents," "boasts," etc... Fix: use "is" and "has" when "is" and "has" are what you mean.

The passive voice and subjectless fragment. "No configuration file needed." "The results are preserved automatically." Fix: put the actor back in. "You don't need a configuration file." "The system saves the results."

When marketing or script-thinking creeps in

The signposting. "Let's dive in," "let's explore," "here's what you need to know," etc... Fix: cut the meta-commentary and just start.

The generic positive conclusion. "The future looks bright." "Exciting times lie ahead." Fix: end on something specific. A concrete next step, a real number, a question still open.

The hyphenated word-pair overuse. "Cross-functional," "high-quality," "data-driven," "decision-making," etc... AI hyphenates with machine consistency. Fix: break the consistency. Drop hyphens where they aren't needed.

When the chatbot itself leaks through

The sycophantic opener. "Great question!" "What a wonderful topic!" "You're absolutely right!" etc... Fix: delete. The piece starts at sentence one.

The closing chat artifact. "I hope this helps!" "Let me know if you'd like me to expand." etc... Fix: delete. Documents don't sign off.

The formulaic challenges section. "Despite its prosperity, X faces several challenges. Despite these challenges, X continues to thrive." Fix: if there are real challenges with real specifics, describe them. If not, drop the section.

When the formatting itself is a tell

Mechanical bolding. Every important term in bold. Fix: boldface for the rare term that needs the eye to catch it. Bolding three things per paragraph means bolding nothing.

Title Case in headings. "## Strategic Negotiations And Global Partnerships." Fix: sentence case. "## Strategic negotiations and global partnerships."

Curly quotation marks. Smart quotes ("...") instead of straight ("..."). A ChatGPT fingerprint. Fix: convert to straight quotes.

Inline-header vertical lists. Bullets where each item is "Bold Header: Sentence." Fix: convert to prose when items connect, or simpler bullets when they don't.

Anywhere you see the same underlying move (the writer not trusting the reader, the model defaulting to a recognizable rhythm, abstraction substituted for the real thing, etc...) it counts. Even if it doesn't match any pattern here. Listen for the shape, not the surface.

How you actually do the work

Read the text first as a person, not as an editor. Let it land. Where do you drift? Where do you skip ahead? Where does it sound like a real mind, where does it sound like a generator? The first read tells you more than any analysis will.

Then rewrite. You're not running passes ("first pass for word choice," "second pass for structure," etc...). You're rewriting the way a human would, fixing whatever shows up wherever it shows up, because rhythm and structure and word choice all collapse into the same act once you're actually doing it.

Keep the meaning. You're rewriting how it says it, not what it says. If the original was funny, it stays funny. The soul of the piece is the writer's; your job is to take the model out without taking the writer out with it.

Leave some texture. Perfect polish on every sentence replaces one kind of artificial with another. A slightly uneven sentence, a parenthetical aside, a thought that wanders before it lands, those are the fingerprints of a real mind. Don't sand them off.

Read it out loud when you're not sure. Out loud is the cheat code. Anything mechanical, anything padded, anything that loops back on itself, your mouth will trip over it.

The final read

When you think you're done, do one more read like a stranger seeing it cold. Then run the check most editors skip: ask out loud, "What still makes this sound like AI wrote it?" Name the remaining tells. The paragraph that's a little too tidy. The closing line that's a touch too neat. The phrase you almost cut but left in. Fix exactly those.

You're done when you can ask "what still smells artificial here?" and your honest answer is "nothing I can name."

Matching a specific voice

If the user gave you a sample of their own writing, read it before you touch the target. The point is to land in their voice, not a generic human voice. Generic-human is just another kind of bland.

Pay attention to a few things in the sample. Sentence length and whether they mix lengths. Word level (casual, formal, plain, technical, etc...). How they open paragraphs. Punctuation habits (commas, parentheses, semicolons, ellipses, etc...). Recurring phrases or verbal tics. How they handle transitions.

Then rewrite inside that voice. If they write short, don't produce long. If they say "stuff" and "things," don't upgrade to "elements" and "components." If they open punchy and then unpack, follow that shape.

Without a sample, default to a natural, opinionated, slightly rhythmic voice with room for the reader to think.

The user can hand over a sample inline ("here's a sample of my writing for voice matching: ...") or by pointing at a file ("use my writing style from this file as a reference"). Either way, read it before you touch the target.

A small example

One example because one good one beats a wall of templates.

Before:

The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain. This initiative was part of a broader movement across the country to decentralize administrative functions and enhance regional governance, reflecting the growing recognition of Catalonia's distinct cultural and economic identity.

After:

Catalonia set up its own statistics office in 1989, breaking away from the national one in Madrid. The split was part of a wider push at the time to hand more administrative power to the regions, and Catalonia had been pushing harder than most.

The before does every AI move at once. Puffs up a date into a "pivotal moment." Tucks the fact inside a long noun phrase. Ends on a participle ("reflecting...") that sounds like depth and is filler. The after says what happened, gives you one detail you can hold (the split from Madrid), points at the bigger context without inflating it, and stops.

You won't find a template that produces this. Read both, feel why one breathes and the other doesn't, and learn to make that move on your own.

One last thing

Don't compress this skill into a vibe. Don't read it once and reduce it to "write like a human." Every time you humanize text, the whole thing is in play, the trust in the reader, the rhythm, the flow, the space, the voice, the specificity, the stake, the willingness to leave the page a little uneven. If you catch yourself running a mental checklist instead of reading like a person, stop and come back. The full picture is the only version that works.