Humanizer, Tested: What the Claude Skill Flags and Rewrites

Quick answer: install the skill, then ask in plain language — Please humanize this text: followed by your paste. That is the entire workflow.

npx skills add blader/humanizer --global

/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer

What you actually install

One Markdown file. blader/humanizer’s SKILL.md is 456 lines and about 4,700 words: 35 numbered writing patterns in five groups (content, language and grammar, style, chatbot leftovers, filler and hedging), a section on what not to flag, and a two-pass rewrite procedure. There is no model, no API, and no detector claim anywhere in the repo. The skill is a set of instructions your existing agent follows on your text, which is also why it is not tied to Claude: the repo ships an OpenAI agents config, and the Skills CLI install above works for any agent that reads skills.

The pattern list comes from Wikipedia’s “Signs of AI writing”, the page maintained by WikiProject AI Cleanup, and the repo tracks that page over time. The changelog shows the list growing from 24 patterns in January 2026 to 35 by version 2.10, with entries as narrow as adding figurative uses of “gated” to §7 while keeping real technical uses like CI quality gates.

The test

We asked Claude to write a launch announcement in deliberately maximal AI-marketing style, then ran Humanizer over it (method and date at the bottom). The input, 67 words:

Big news: your notes just got a whole lot smarter. We’re thrilled to announce that PDF search has arrived — meaning every document, every page, and every word is now instantly at your fingertips. No more digging, no more scrolling, no more wondering where that one file went — just type, search, and find. This is more than a feature; it’s a whole new way to work.

The skill reported seven patterns: inflated importance (§1), sales language (§4), “not X but Y” endings (§9), forced groups of three (§10, three separate instances), em dashes (§14), bold text (§15), and a generic positive ending (§25). The rewrite, 45 words:

Your notes just got smarter: PDF search is here. Search now covers the full text of every document, so typing a few words pulls up the file you need. You won’t have to dig through folders trying to remember where that one PDF went.

Both em dashes are gone, all three triads are gone, and the closing sentence was deleted rather than softened, which is what §25 instructs. The factual content (PDF search shipped, it covers full document text) survived intact. That reflects the skill’s hardest rule: a name, number, date, or claim must come from the source, and the rewrite may not invent any.

What it kept is as informative as what it cut. “Your notes just got smarter” survives as the opener. Humanizer removes signs of machine writing; it does not remove marketing intent. If the input argues something, the output still argues it, in fewer and plainer words.

What this test does not show

One run on an easy target: marketing copy written to be caught. The skill’s harder path is voice matching — you paste two or three paragraphs of your own writing and it follows that sample’s rhythm and word choice instead of its default plain style — and we did not measure that here. The false-positive section also matters in real use: it protects deliberate repetition, quoted text, and technical vocabulary, and we did not stress it.

Repo facts as of September 6, 2026: 43,335 stars (988 of them gained that day), 54 commits since the initial one on January 17, 2026, version 2.11.2, MIT license.

Related: AI humanizer or a better model and how to use Humanizer and keep your own voice.

Humanizer runs inside your agent; StashBase isn’t involved. Method: skill file from a fresh clone of blader/humanizer v2.11.2, run through Claude Code non-interactively (claude -p) on September 6, 2026. The sample input was generated by Claude on request to maximize AI-marketing patterns; input and output are quoted verbatim.