中/EN
Frontline Hotspot
Frontline Hotspot

Humanizer: 33 Rules to Wash the AI Tone Out of Your Articles

Based on Wikipedia's AI Cleanup project, 33 AI-writing signature checks-one-line install, two-pass auto-rewrite to strip the machine feel. With limits analysis and hands-on path.

Published July 22, 20264 min read
<!-- humanizer-ai-text-deai-tool | hotspot | Humanizer 33 Rules -->

The WeChat article you wrote gets "This is AI-written, right?" as the first comment.

It's not that you write badly-it's that LLM output is too "signature.""In today's rapidly evolving digital era," "not only... but also...," "let's explore together." These patterns have become AI-text fingerprints; readers spot them at a glance.

Humanizer, freshly open-sourced on GitHub, is built to solve exactly this. Live for less than a week, Star count broke a thousand, comments uniformly "finally someone did this."

What It Is

Humanizer is essentially a structured prompt instruction (SKILL.md), designed to run as an Agent Skill inside AI coding tools that support the skill protocol-Claude Code, Cursor, etc. One-line install:

bash
npx skills add blader/humanizer

Once installed, toss the text you want rewritten to it, and it auto-runs two passes: the first rewrites by the rules, the second does an "AI-tone audit" to clean any machine-feel expressions that slipped through.

No API key, no deployment-just a pure Markdown instruction file, MIT license, take it and use it.

Core: 33 AI-writing Signature Checks

Humanizer's rule library comes from Wikipedia's WikiProject AI Cleanup-a group of editors who, while cleaning AI-generated junk entries, summed up 33 "AI signature" patterns. Four categories:

Content level: meaning inflation ("marks a key moment in the evolution of..."), name-dropping, vague attribution ("experts believe"), the universal challenge formula ("despite challenges... continues to thrive").

Language level: AI high-frequency words (additionally, testament, landscape, showcasing), copula avoidance (won't say is/has, prefers serves as/features/boasts), three-part parallelism, synonym rotation, fake scope ("from the Big Bang to dark matter").

Style level: dash overuse, bold overload, inline title lists, Title Case headings, emoji decoration, hyphenated word pairs (cross-functional, data-driven).

Communication level: chatbot tails ("hope this helps!"), sycophantic tone ("great question! you're so right!"), filler phrases ("In order to" -> "To"), over-hedging ("could potentially possibly" -> "may").

Each rule comes with a Before/After comparison. For example:

Before: "I recently spent five unforgettable days in Lisbon, and let me tell you - this city completely stole my heart."

After: "I spent five days in Lisbon last October and still have mixed feelings about it. Beautiful, yes. Also harder on the knees than anyone warned me."

The rewrite isn't "more concise"-it's more like a real person talking, with hesitation, concrete details, imperfect sentence rhythm.

An Underrated Feature: Voice Calibration

Most people install and use it directly, but Humanizer has a hidden trick: feed it 2-3 paragraphs of your own writing, and it analyzes your sentence rhythm, word preferences, and expression habits, then rewrites in your style instead of outputting a "generic clean version."

This matters a lot for creators-you don't need to write every piece from scratch yourself. AI drafts, Humanizer rewrites in your voice, readers can't tell it's machine-written.

Limits and Caveats

A few honest words:

  1. It's not an anti-detection tool. It won't pass detectors like GPTZero, because those look at token probability distributions, not surface patterns. Its goal is "reads unlike AI," not "fools the algorithm."

  2. Limited Chinese support. The 33 rules are mostly summarized from English corpora; Chinese AI text has its own signatures ("it's worth mentioning," "all in all," "in the context of..."), and you need to add rules yourself.

  3. Can't replace fact-checking. It only handles style, not content correctness. Fake data AI made up is still fake data after a style pass.

How to Use It

Simplest path:

bash
# 1. Install (needs Node.js)
npx skills add blader/humanizer

# 2. Call in Claude Code
/humanizer
[paste your article]

If you don't use Claude Code, just copy the SKILL.md from the repo into any tool's skill directory that supports skill instructions. Cursor, Windsurf, OpenCode all work.

For batch content production (e.g., a workflow generates drafts then publishes), you can extract the 33 rules into part of a system prompt as post-processing after the writing node.

One-line Summary

The problem with AI writing was never "can't produce," it's "looks AI the moment it's produced." Humanizer turns "de-AI-ifying" from mysticism into an executable 33-item checklist, open-source and free. Whether you use the tool or not, the 33 rules alone are worth saving as a writing self-check list.

Related