A Claude Code skill that strips the telltale signs of AI-generated writing and brings the writer's own voice forward. 43 named patterns, each with before/after examples, plus a detect-only audit mode.
Most AI text gets caught long before anyone runs a detector. Readers recognize the tells: "testament to", perfect triads, em dashes everywhere, bolded list headers, a "challenges and future prospects" section, an upbeat recap ending. This skill names each tell, shows the fix, and then does the harder half of the job: bringing the writer's own voice forward so the result reads like a person wrote it. It never invents opinions, experiences, or facts to get there, because a fake human voice is as easy to spot as the chatbot one.
It is based on Wikipedia's Signs of AI writing guide, maintained by WikiProject AI Cleanup from observations of thousands of AI-generated texts.
mkdir -p ~/.claude/skills/humanizer
curl -o ~/.claude/skills/humanizer/SKILL.md \
https://raw.githubusercontent.com/WolfurX/humanizer/main/SKILL.md
curl -o ~/.claude/skills/humanizer/LICENSE \
https://raw.githubusercontent.com/WolfurX/humanizer/main/LICENSEOr clone and copy. Claude Code picks it up automatically; it triggers when you ask it to humanize or de-AI a text, when it reviews a draft before publishing, and when it writes user-facing prose of its own.
Edit rewrites the text: patterns out, meaning intact, voice preserved. Guardrails keep the minimum effective edit, so a rough draft with a real voice still sounds like the same person afterward, just without the tells.
Detect audits without rewriting: each pattern found, the offending line quoted, the fix in a few words. It never claims to know whether AI wrote the text. Detectors guess; named patterns are evidence you can check yourself.
Significance inflation, notability name-dropping, superficial -ing analyses, promotional tone, weasel attributions, guesses dressed as facts, formulaic "challenges" sections, faux-insight setups, recap endings, AI vocabulary, filler, hedging, copula avoidance, negative parallelisms, rule-of-three overuse, synonym cycling, false ranges, passive voice and subjectless fragments, em dash overuse, colon reveals, punchy kickers, rhetorical questions, sentence-opener tics, forced metaphors, aphorism formulas, persuasive authority tropes, narrated candor, false concessions, launch-copy introductions, signposting, fragmented headers, reassurance kickers, boldface overuse, inline-header lists, title case, emojis, mismatched quote marks, hyphenation errors, chatbot artifacts, knowledge-cutoff disclaimers, sycophancy, leaked placeholders and tool residue, and the styled callout box that marks a document as machine-made before a single word is read.
Give it a sample of your own writing and it matches your sentence length, word choice, and punctuation habits instead of imposing a house style.
Wikipedia's Signs of AI writing guide, via blader/humanizer (MIT, Siqi Chen), extended with additional patterns; the useful parts of petergyang/no-ai-slop (MIT, Peter Yang) merged in, including the detect mode and the minimum-effective-edit guardrails; and the useful parts of conorbronsdon/avoid-ai-writing (MIT, Conor Bronsdon), including the never-inject guardrails, six patterns, and the whole-piece tests. Pairs with document-design, which guards the layout side of the same problem.
MIT