How Humanizer Makes AI Text Sound More Human: Inside the 25-Pattern Engine

Humanizer transforms AI-generated text into natural human prose by scanning for 25 specific "tells" in LLM output and applying handcrafted rewrite rules that remove generic statistical patterns while strictly preserving factual accuracy.

The blader/humanizer repository is a Markdown-based skill that rewrites robotic AI content into authentic human writing. According to the source code in SKILL.md, it operates through a deterministic pipeline that identifies common LLM artifacts—such as forced triads, passive constructions, and over-used "deep-sounding" phrases—and replaces them with direct, claim-driven sentences without hallucinating new information.

The Pattern Engine: Identifying AI Artifacts

At the core of Humanizer's ability to make AI text sound more human is the Pattern Engine, defined entirely within SKILL.md. This component scans input text for 25 distinct patterns that indicate AI authorship, ranking each occurrence by strength and frequency as documented in the repository's README.md.

These patterns target statistical artifacts that large language models favor, including:

• The "Not X but Y" construction
• Forced triads (groupings of three items)
• Over-used "deep-sounding" vocabulary
• Passive voice and filler qualifiers
• Unnecessary emphasis (bold, headings, emojis)

When the engine detects these tells, it marks them for rewriting—strongest patterns first—ensuring the most obvious AI signals are prioritized for transformation.

The Four-Stage Rewrite Pipeline

After pattern detection, Humanizer processes text through a multi-stage pipeline defined in the repository's documentation.

Draft Generation Without Structural Lock

The Draft Generator creates a first rewrite that does not treat the original sentence structure as fixed. Located in the logic defined by SKILL.md, this stage applies specific rewrite rules to each marked tell, converting generic AI phrasing into concise alternatives while maintaining the original factual claims.

Fact Guard: Preventing Hallucination

The Fact Guard component enforces a strict "no-invented-facts" policy documented in README.md. It guarantees that no new names, dates, numbers, or quotes are fabricated during rewriting. If the system encounters a required detail that is missing from the source text, Humanizer prompts the user for clarification rather than hallucinating content—a critical distinction from standard paraphrasing tools.

The Critique Pass

Following the initial draft, Humanizer executes a Critique Pass that scans the rewritten text for any residual AI-style phrasing. As described in the repository's usage documentation, this step reviews the first draft for remaining generic patterns before producing the final version.

Optional Voice Matching

When provided with a writing sample, Humanizer activates Voice Matching capabilities. According to README.md, the system mirrors the sample's rhythm, word choice, punctuation patterns, and "deliberate quirks"—such as specific dash usage—to align the output with the user's personal writing style.

Installation and Usage

Humanizer installs as a global or local skill via npm and operates through simple text commands. The complete logic resides in SKILL.md, with validation helpers available in scripts/validate-package.py for CI checks.

Install the skill once:

npx skills add blader/humanizer --global

Call the skill directly in any chat-compatible agent:

/humanizer

[Paste AI-generated text here]

To humanize an entire file while preserving code blocks and frontmatter:

Humanize the prose in docs/launch-post.md

For voice matching, provide a sample before the target text:

/humanizer
Here's a sample of my writing for voice matching:
[Paste 2-3 paragraphs of your own writing]

Now humanize this text:
[Paste AI-generated text]

Summary

• Pattern Detection: Humanizer uses 25 handcrafted rules in SKILL.md to identify statistical AI artifacts like forced triads and generic phrasing.
• Structure-Agnostic Rewriting: The Draft Generator rewrites content without fixing the original sentence structure, prioritizing natural flow over rigid paraphrasing.
• Fact Preservation: The Fact Guard prevents hallucination by requiring all names, dates, and quotes to originate from the source material.
• Quality Assurance: A Critique Pass reviews drafts for residual AI patterns before finalizing output.
• Voice Adaptation: Optional voice matching analyzes user-provided samples to replicate personal rhythm and punctuation quirks.

Frequently Asked Questions

How does Humanizer prevent inventing fake facts when rewriting text?

Humanizer implements a Fact Guard system that mandates all names, numbers, dates, and quotes must originate from the source text. If the rewrite requires information not present in the input, the tool asks the user for clarification rather than fabricating details, as specified in the README.md documentation.

What are the 25 patterns that Humanizer detects in AI text?

The 25 patterns include specific linguistic tells such as "Not X but Y" constructions, forced triads, over-used AI vocabulary, passive voice, and unnecessary formatting emphasis. These patterns are prioritized by strength and frequency in the SKILL.md file, targeting the statistically likely phrasing that LLMs tend to generate.

Can Humanizer match my personal writing style?

Yes, through the Voice Matching feature. When you provide a 2-3 paragraph sample of your own writing, Humanizer analyzes and replicates your rhythm, word choice, punctuation habits, and deliberate quirks—such as specific dash usage—to ensure the output aligns with your authentic voice rather than a generic human average.

Is Humanizer a code library or a standalone application?

Humanizer is neither a traditional library nor a binary application. It is a Markdown-based skill contained primarily in SKILL.md that works with any AI agent supporting Markdown skills (such as Claude or OpenAI-compatible agents). It requires no code execution and installs via npx skills add blader/humanizer.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}