blader's humanizer repository added 407 stars in a single day on GitHub, lifting its star count to 64,559 S¹. The Python package functions as an agent skill designed to remove recognisable machine-writing habits from generated text, and its creators are candid that it will not pass AI-detection tools S¹.
The narrower pitch is what sets this apart. A tool that smooths away AI stylistic habits while conceding that detectors still catch most of its output is staking a smaller claim than the typical "undetectable AI" product. The 16-out-of-16 blind preference test is reported by the maintainers themselves with no outside verification, no documented sample size, and no published methodology S¹. The Wikipedia "Signs of AI writing" page has not been confirmed as carrying official policy status. The open question is whether a 26-item pattern list keeps pace with newer model output, or whether it ages as writing styles evolve.
The project is backed by cfo.ai S¹ and ships as a plugin for seven platforms: Claude Code, Codex, Claude.ai, Claude Desktop, Gemini CLI, GitHub Copilot, and Windsurf S¹. Its most recent commit is dated October 28, 2026, addressing a plugin manifest schema problem P².
Twenty-six patterns drawn from an editor-facing checklist
The maintainers list 26 distinct patterns that Humanizer looks for S¹. The five it treats as the strongest indicators are "Not X but Y" sentence structures, single-sentence endings, aphoristic phrases that seem meaningful but are not, extended introductions before a straightforward point, and rebuttals to objections that no one actually raised S¹. Readers familiar with AI-generated prose will recognise the shapes: the contrast pivot, the tidy concluding line, the pseudo-profound maxim, the slow wind-up, and the phantom counterargument.
The underlying reference is Wikipedia's "Signs of AI writing" page, which the maintainers characterise as a resource that Wikipedia editors consult when spotting machine-generated content S¹. Whether that page holds official policy authority on Wikipedia has not been verified. Humanizer works through the list and rewrites around each pattern, keeping the original meaning intact while altering the surface texture of the prose.
A self-reported preference result with no outside replication
The maintainers state that judges in a blind comparison chose Humanizer's rewritten version over the original AI text in all 16 cases S¹. That result comes from the project itself, is linked to a GitHub issue (#229), and has not been independently reproduced. The number of judges, how they were selected, and what methodology governed the comparison are not documented in the available material.
The more defensible claim is the negative one. The maintainers say plainly that evading AI detectors is not an objective, and that detection tools still flag the majority of text processed by the tool S¹. That position runs counter to what most products in the "humanizer" category advertise. Humanizer is betting on a different outcome: it removes the stylistic habits that make prose feel mechanical, not the statistical signals that detectors rely on.
Plugin installers gain a pre-publication pattern sweep
A content team drafting blog posts inside Claude Code could install Humanizer and run it against drafts before they go live, receiving flagged patterns and suggested rewrites. A technical writer producing documentation through GitHub Copilot could apply the same workflow. The skill runs on any agent that supports the plugin format S¹, which covers most major coding assistants.
The trust question is genuine. Humanizer is open source, so the code is available for inspection, but a star count of 64,559 is not equivalent to a security audit. Anyone adding it to a production agent should review the plugin code first.
The repository's last commit is dated October 28, 2026 P², and the project's GitHub page remains the primary source for all functional claims S¹.
Sources: S1 — blader/humanizer: Agent skill that removes signs of AI-generated writi · P2 — GitHub - blader/humanizer: Agent skill that removes signs of AI-genera · P3 — anthropics/claude-code · P4 — janithcooray/claude-desktop · P5 — GitHub - blader/humanizer: Agent skill that removes signs of AI-genera
Written from 5 sourced items, 4 of them primary.
