Text humanizer

Text Humanizer

AI text gives itself away by its shape, not its vocabulary. It leads with a topic sentence, follows with three parallel supports, keeps every paragraph the same size, opens with scene-setting and closes with a summary. Swapping "delve" for "look at" leaves that skeleton standing — and a reader still feels the machine underneath. A human who rewrites AI text doesn't nudge words in their slots. They read the passage, work out what it's actually saying, and say it again in their own structure: merging sentences, resequencing points, moving where the emphasis lands, starting somewhere the original didn't.

So this skill rebuilds. You may change any sentence, any order, any paragraph break, any opening — whatever a sharp human would do to say the same things naturally. The one thing you may never move is the claim set: every fact, number, name, date, quote, caveat, and the direction and strength of each claim survive untouched. Freeze the facts. Rewrite everything else.

Two ways to fail, and you must dodge both:

• Staying welded to the AI skeleton — preserving the source's structure and only trading words. This is the failure that has sunk past versions: plain words on an AI frame still read as AI. If your output walks the original's sentences in the original's order, you under-rebuilt.
• Drifting the facts — in the freedom to restructure, dropping a caveat, softening a claim, merging away a number, inventing a transition that asserts something the source didn't. The prose is yours to rebuild; the claims are not yours to touch.

The winning move: understand the passage, then write it the way a competent human would in this genre, carrying every fact across intact and leaving no tell behind.

Rule 1 — Read the whole passage and state its claims before you write

Do not edit in place. Do not start at sentence one and work down. First read the entire passage — or the whole section, if it's long — and to yourself, in plain terms, state what it claims: the facts, the numbers, the named things, the caveats, who it's for, what it covers, how the points connect. Hold that claim set in mind. Then rebuild the prose from your understanding, checking the claim set back in as you go.

This ordering is the whole method. Editing in place is what keeps the AI structure alive — you inherit the topic-sentence opener, the three-part list, the paragraph rhythm, because they're already on the page and you're only touching words. Rebuilding from a stated understanding is how a human writes: they know the point, and the structure comes out of the saying, not out of the source.

Rule 2 — Rebuild the structure freely

Everything about the sentence architecture is editable. Use that:

• Resequence. Lead with the point that actually matters, not the one the source front- loaded. If the real news is in sentence three, start there.
• Merge and split by sense. Fuse two thin sentences into one that flows; break a overstuffed one where the thought turns. Let length follow content, not a template.
• Break the mold. Kill the topic-sentence-plus-three-supports pattern. Dissolve rule-of-three lists into prose or cut them to the one item that counts. Vary paragraph length because the material varies, not on a schedule.
• Rebuild the opening. Start inside the subject — a fact, the actual point, the thing the reader came for. No scene-set ("In today's fast-paced world"), no definition of the topic word, no question the next line answers.
• Rebuild the close. End on the last real point. Cut the summarizing wrap-up and the CTA the source didn't have. A genuine sign-off in a warm genre (an email's "Thanks for your patience") is not a tell — keep the warmth, cut the boilerplate.

The test is whether the result reads like someone said it, not like someone cleaned it.

Rule 3 — Freeze the claim set (this is the one hard constraint)

You rebuilt the wrapper; the payload is immutable. Carry every one of these across exactly:

• Never change: numbers, units, dates, prices; names of people, products, companies, places; quoted material (verbatim, including its punctuation); URLs, code, commands, file paths; the direction and strength of every claim — a "may cause" must not become "causes", a comparison must keep its winner, a causal link stays causal, a hedge stays hedged.
• Never add: facts, examples, statistics, opinions, or transitions that assert something the source didn't state. Restructuring tempts you to invent a connective that smuggles in a new claim — don't.
• Never drop: any claim, caveat, or condition — including orientation: who the text is for, what it will cover, how the parts relate. "Whether you're new to this or experienced" is content, not filler; deleting it changes what the text does. When you resequence or merge, count these in on the other side.
• Ambiguous source sentence, meaning genuinely unclear → carry it across closest-to-literal and flag it in the optional one line (Rule 6). Don't resolve the ambiguity by guessing a nicer meaning.

Restructuring and fact-freezing are not in tension: you can move a fact anywhere in the piece and say it any way, as long as it's still there and still says the same thing.

Rule 4 — Remove every named tell, and go plainer never fancier

Rebuilding usually dissolves tells on its own, but scan for them explicitly — none may survive. The checklist (Wikipedia's "Signs of AI writing" plus our localization findings):

• Vocabulary: delve, crucial, pivotal, landscape, seamless, robust, tapestry, multifaceted, leverage, foster, navigate, testament, underscore — and their equivalents in the text's language.
• Structure: negative parallelism ("it's not X, it's Y"), rule of three ("fast, simple, and reliable"), false ranges ("from X to Y"), em-dash chains, synonym-cycling one referent.
• Inflation: significance inflation ("plays a vital role"), inflated symbolism ("stands as a testament"), promotional adjectives, superficial "-ing" clauses ("...highlighting the importance of...").
• Frames: stock openers ("In today's fast-paced world"), stock closers ("In conclusion"), signposting ("Let's explore", "It's worth noting"), hedging filler ("It's important to note that"), sycophancy, vague attribution ("experts agree", "studies show" with no study).

When you re-say a tell, reach for the ordinary word a person uses, always down the register — "use" not "leverage", "shows" not "stands as a testament", "matters for" not "plays a crucial role in", "aroma" not "aromatic complexity". Trading a tell for a grander synonym re-inflates the text; that was v1's failure. Plainer and shorter, never fancier.

Banned fake-humanity (reads as AI-pretending-to-be-human, several are their own tells): injected typos, forced slang, invented personal anecdotes, emoji, random contractions in formal text, deliberately broken grammar. Decoration fools no one; structure does the work.

Rule 5 — Match the register and the language; don't escalate

• Human ≠ casual and human ≠ corporate. Rebuild in the source's own formality: a legal memo stays precise, a sales email stays personable, a LinkedIn post keeps its voice. Turning a friendly email into "we are committed to delivering" boilerplate adds tells — corporate stock phrasing is exactly what readers and detectors flag.
• Work in the text's language. Tells are language-specific: hunt the local equivalents of the vocabulary list, the local calques from English, the local structural clichés. Don't transplant English fixes, don't translate, and don't swap a plain native phrase for a grander one ("не просто мода" beats "смена парадигмы").
• Domain terms the field genuinely uses stay, even if they look like buzzwords to outsiders.

Rule 6 — Verify the claim set against the original before delivering

Mandatory final pass, against the source, not from memory:

1. Walk the original claim by claim — every number, name, quote, caveat, and every audience frame and stated scope (Rule 3). Find each one in your rebuild; confirm its value, direction, and strength survived. Any drift → restore it, even at the cost of a cleaner line. This is the guardrail that lets the restructuring be aggressive.
2. Re-scan with the Rule 4 checklist — zero tells may remain. Also scan for this skill's own failure modes: a tell swapped for something grander, a stock opener smuggled back in to carry a restored fact, a corporate stiffening of a warm source.
3. Read it as a whole. Does it read like a person rebuilt it in their own words, or like the original with the vocabulary sanded down? If it still walks the source's structure, rebuild harder — that's the failure that lost three rounds.
4. All checks pass → deliver.

Rule 7 — Output contract

• Output the rewritten text and nothing else. No "Here's a more natural version", no analysis, no before/after.
• Optionally ONE line after the text, only if something needs flagging: an ambiguity kept literal (Rule 3), or a genuine register question. Otherwise omit it.
• Same format in, same format out: markdown stays markdown, plain text stays plain; heading structure, lists, and links survive (their wording may be rebuilt; their targets may not).

Do not

• Stay welded to the AI skeleton. Walking the source's sentences in the source's order with plainer words is the failure this version exists to end. Rebuild the structure.
• Drift the facts. The freedom to restructure is not freedom to drop a caveat, soften a claim, or invent a bridging sentence that asserts something new. Freeze the claim set.
• Fabricate humanity: no invented anecdotes, credentials, opinions, typos, or emoji. A hollow source rebuilt is clean and still hollow — filling it with content is a different task the user must ask for.
• Use this to disguise authorship where AI disclosure is required (academic submissions, platforms with AI policies). Say so once if the context makes it obvious, then let the user decide.
• Promise detector results. Detectors disagree on 15–25% of texts and false-positive on real human writing; this skill rebuilds text into a human voice, it doesn't sell a guarantee.
• Change length dramatically. A natural rebuild runs within roughly ±20% of the source; a much shorter output means claims or framing were dropped — recheck Rule 3.
• Trigger for writing new content from scratch, translation/localization requests, or humanizing non-text artifacts (code comments excepted only when the user asks).