How Humanizer Uses False-Positive Guards to Protect Quoted Text, Proper Names, and Meta-Discussions

Humanizer employs three specific false-positive guards—quoted text detection, proper name recognition, and self-referential pattern exclusion—to prevent accidental rewriting of content that merely contains stylistic patterns rather than actively using them.

The blader/humanizer repository implements a sophisticated pattern-matching engine designed to identify and remove "tells" in prose, but it risks damaging legitimate content without robust safeguards. Understanding these false-positive guards is critical for users who want to ensure the tool preserves quotations, brand names, and analytical discussions about language patterns.

The Three Core False-Positive Guards in Humanizer

Humanizer's defense system relies on context-aware detection to distinguish between active usage of a stylistic pattern and passive mentions within protected contexts.

Quoted Text Protection (Version 2.8.1)

According to the source documentation in README.md lines 184-185, version 2.8.1 introduced "a safeguard for quoted text" that prevents the engine from modifying content wrapped in quotation marks. This protection covers both straight quotes ("…") and curly quotes ("…"), ensuring that cited speech, book titles, and direct quotations remain intact regardless of whether they contain patterns like not-X-but-Y.

Proper Name Recognition

The tool recognizes when a pattern-like phrase functions as part of an identifier rather than a stylistic construction. In SKILL.md lines 62-64, the "When not to act" section explicitly mandates leaving watched phrases alone when they appear inside "a proper name." For example, a product officially named Not-X-But-Y will not be rewritten to Y instead of X because the engine identifies the capitalized, hyphenated structure as a proper noun.

Self-Referential Pattern Exclusions

Humanizer detects meta-discussions where the text analyzes or explains a pattern rather than employing it. The same SKILL.md section requires preserving passages that "discusses the phrase rather than uses it." When encountering sentences like "The not-X-but-Y pattern is the strongest tell," the engine recognizes the explanatory context and bypasses transformation.

How the Guards Work: Weak-Alone Logic

These safeguards operate as weak-alone protections, meaning the pattern is only ignored when the surrounding context explicitly matches one of the three criteria above. This design prevents over-aggressive editing that could alter meaning, break proper nouns, or erase useful meta-information while maintaining the tool's effectiveness on actual stylistic tells.

The logic resides primarily in SKILL.md, which serves as the core skill definition, while agents/openai.yaml points to this configuration to enforce the behavior across agent implementations.

Practical Code Examples

The following examples demonstrate how Humanizer applies these false-positive guards in practice:

Quoted Text Protection:

> Original input
The phrase **"not X but Y"** appears in the title of the paper.

> Humanizer output (unchanged)
The phrase **"not X but Y"** appears in the title of the paper.

Proper Name Protection:

> Original input
Our product is called *Not-X-But-Y* and it solves the problem.

> Humanizer output (unchanged)
Our product is called *Not-X-But-Y* and it solves the problem.

Meta-Discussion Protection:

> Original input
The **not-X-but-Y** pattern is the strongest tell in Humanizer.

> Humanizer output (unchanged)
The **not-X-but-Y** pattern is the strongest tell in Humanizer.

Summary

• Quoted text safeguard: Added in v2.8.1, protects content inside straight or curly quotation marks from transformation.
• Proper name guard: Identifies pattern-like phrases that function as brand names, product titles, or identifiers and excludes them from editing.
• Self-referential exclusion: Prevents rewriting when text discusses Humanizer's patterns analytically rather than using them stylistically.
• Weak-alone implementation: Guards only activate when specific contextual triggers are detected, ensuring surgical precision.
• Source location: Defined in SKILL.md (lines 62-64) and documented in README.md (lines 184-185).

Frequently Asked Questions

How does Humanizer detect quoted text versus regular text containing quotes?

Humanizer parses the context surrounding quotation marks to determine if content is encapsulated as a citation or title. According to README.md, the v2.8.1 safeguard specifically targets phrases wrapped in "…" or "…" characters, treating the enclosed content as protected regardless of internal pattern matches.

Will Humanizer rewrite product names that contain stylistic patterns like "Not Only X But Also Y"?

No. The proper name guard in SKILL.md explicitly preserves watched phrases when they appear within proper names or titles. If the capitalization, hyphenation, or context indicates the phrase serves as an identifier (e.g., Not-X-But-Y), Humanizer leaves the text untouched to prevent damaging brand identities or official nomenclature.

Why does Humanizer ignore discussions about its own patterns?

The "When not to act" section in SKILL.md requires the tool to distinguish between using a pattern and mentioning it. When text analyzes, explains, or references a pattern meta-linguistically (e.g., "The not-X-but-Y construction reveals bias"), the engine classifies this as explanatory rather than stylistic content, preserving the original phrasing to maintain analytical clarity.

Where are the false-positive guard rules defined in the codebase?

The primary definitions reside in SKILL.md at lines 62-64, which establishes the "When not to act" protocol. High-level documentation appears in README.md at lines 184-185, noting the quoted-text addition in version 2.8.1. The agents/openai.yaml file references these configurations to ensure consistent application across the OpenAI agent implementation.

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