ClaudeHumanizer

Optimized for Claude Sonnet 4.5 (Recommended) | Also supports: Gemini 2.5 Pro, GPT-5

A professional AI text humanization system using a specialized 11-phase assembly line (with optional phases 6.1, 8.5, and 9.5) to transform AI-generated content into natural, human-like writing while preserving meaning and voice.

Overview

ClaudeHumanizer employs a domain-specialized assembly line where each phase targets one specific aspect of text improvement. This systematic approach eliminates AI detection markers while maintaining quality through:

• Sequential processing - Each phase builds on previous improvements
• Domain isolation - No phase interferes with others' specialized work
• Pattern-based filtering - Intelligent rules for dialogue pauses, light descriptions, finger movements
• Master prohibited list - Prevents reintroduction of AI-associated terms
• Final quality control - Phase 10 catches prohibited words reintroduced by phases 3-9
• Silent operation - Returns improved text without commentary

Quick Start

Important Usage Considerations

Non-Native English Speaker (NNES) Bias Warning AI detectors exhibit documented bias against non-native English speakers. NNES writing often features simpler sentence structures, more limited vocabulary, and reliance on common phrasings—characteristics that detectors may misclassify as AI-generated.

Important: If you are processing text originally written by an NNES writer, be aware that:

• The humanization process may inadvertently simplify vocabulary and sentence structures further
• Some NNES writing patterns are naturally similar to AI patterns (limited lexical diversity, common phrases)
• Over-processing NNES text through all phases might make it appear MORE AI-like to certain detectors
• Consider selectively applying phases rather than the full pipeline for NNES-authored content

Hybrid Text (Human + AI) Guidance AI detectors often fail catastrophically on hybrid texts containing both human and AI-generated content, frequently misclassifying them as either 100% human or 100% AI with no middle ground.

Important: This system assumes 100% AI-generated input. If you have hybrid text:

• Manually identify which sections are human-written vs AI-generated
• Only process the AI-generated sections through the pipeline
• Keep human-written sections completely untouched
• Do NOT run mixed human/AI paragraphs through the system—process them separately
• Consider whether detection is even a concern if substantial human contribution exists

Gemini-Optimized Workflow (NEW)

For users who prefer to use the Google Gemini API, a new, optimized 4-stage workflow is available. This workflow consolidates the 11+ Claude phases into four powerful "mega-prompts" that leverage Gemini's large context window and advanced instruction-following capabilities.

This approach significantly reduces the number of API calls, leading to faster processing and lower costs, while still providing a comprehensive humanization process.

View the Gemini Humanizer Workflow Guide

For full details on the 4-stage approach and instructions on how to use the Gemini prompts and the provided n8n_gemini_workflow.json, please refer to the documentation in the gemini directory.

Model Selection (October 2025)

Choose your LLM based on priorities:

| Model | Best For | Cost | Key Advantage | |-------|----------|------|---------------| | Claude Sonnet 4.5 (Recommended) | Maximum quality | $3-15/1M tokens | "Surgical" edits, natural human tone, best instruction-following | | Gemini 2.5 Pro | Budget/Long texts | $1.25-15/1M tokens | 40% cheaper, 1M context, fastest (372 tok/s) | | GPT-5 | ChatGPT users | Subscription | Literary style, widely available (requires stricter prompting) |

Quick Decision:

• Quality priority? → Claude Sonnet 4.5 (recommended)
• Budget priority or text >100K words? → Gemini 2.5 Pro
• Already have ChatGPT Plus? → GPT-5 (with cautions)

Basic Workflow

1. Download required files: 10 phase prompts + docs/master_prohibited_words.json (+ optional genre-specific lists)
2. Select your model: Claude Sonnet 4.5 (recommended), Gemini 2.5 Pro (budget), or GPT-5
3. Process sequentially: Phase 1 → 2 → 3 → 4 → 5 → 6 → 7 → 8 → (8.5 optional) → 9 → (9.5 optional) → 10
4. Include master list: Required for phases 2 and 10 (contains pattern rules); optional genre-specific lists for Phase 10 when author indicates romance or erotica
5. Use previous output: Each phase processes the result from the previous phase
6. Temperature settings: Use temperature 1.0 for Phase 6 (dialogue), standard temps for others

Execution Methods

Manual Processing

Copy each phase prompt into Claude Sonnet 4.5 with appropriate dependencies:

For phases 2 and 10:
[docs/master_prohibited_words.json content]
[docs/phase_prompt.json]
[input text]

For phases 1, 3, 4, 5, 6, 7, 8, 9:
[docs/phase_prompt.json]
[input text]

Claude Project Setup (RECOMMENDED for Claude users)

Configure custom instructions for automated sequential processing using Claude Sonnet 4.5 (see Usage Guide)

ChatGPT Project Setup (For GPT-5 users)

Configure custom GPT or use Projects feature with GPT-5 model selection (see Usage Guide)

Google AI Studio (For Gemini 2.5 Pro users)

Use Google AI Studio with Gemini 2.5 Pro for automated processing (see Technical Reference)

Automation Integration

Set up n8n, Make.com, or API workflows (see Technical Reference)

Assembly Line Phases

Core Processing Phases

| Phase | File | Domain | Master List | NEW Features | |-------|------|--------|-------------|---------------| | 1 | docs/1_grammar_foundation.json | Grammar errors only | No | - | | 2 | docs/2_ai_word_cleaning.json | AI vocabulary removal | Required | Pattern rules | | 3 | docs/3_overwritten_language_reduction.json | Purple prose + nominalization | No | De-nominalization | | 4 | docs/4_sensory_enhancement.json | Flat passage + extreme specificity | No | Hyper-specific details (v2.3.0) | | 5 | docs/5_subtlety_creation.json | Obvious statements + summaries | No | Summary elimination (v2.4.0) | | 6 | docs/6_dialogue_enhancement.json | Character voice (temp 1.0) | No | - | | 7 | docs/7_weak_language_cleanup.json | Weak language + voice distribution | No | Active/passive monitoring | | 8 | docs/8_strategic_imperfections.json | Rhythm + punctuation inconsistency | No | Enhanced imperfections | | 8.5 | docs/8.5_structural_construction_elimination.json | Syntactic patterns + Rule of Three | No | 31 construction patterns (v1.1.0) | | 9 | docs/9_final_verification.json | AI patterns (N-grams + perplexity) | No | Pattern replacement | | 10 | docs/10_final_ai_word_sweep.json | Word filtering only | Required (+ optional genre lists) | Pure prohibited word removal |

Optional Enhancements

Phase 6.1: docs/6.1_character_dialogue_pass.json - Character-specific dialogue customization for targeted voice refinement (see Customization Guide)

Phase 8.5: docs/8.5_structural_construction_elimination.json - Syntactic pattern elimination (v1.1.0) detecting and restructuring 31 mechanical construction patterns that substitute form for content. Now includes Rule of Three symmetry breaking and insecure paragraph summary elimination (Chaos Method patterns). Recommended for commercial fiction and erotica; optional for literary fiction where patterns may be intentional. Can be used standalone or integrated as standard pipeline phase.

Phase 9.5: docs/9.5_statistical_analysis_hub.json - COMPREHENSIVE STATISTICAL HUB consolidating all quantitative metrics (burstiness, POS distribution, lexical diversity/TTR) into single-pass analysis. Use when AI detection is a concern or text needs statistical optimization. Provides optional detailed metrics report.

Key Features

Architectural Clarity (NEW)

Clear Separation of Concerns:

• Phase 9: QUALITATIVE pattern replacement (N-grams, formulaic phrases, AI patterns)
• Phase 9.5: QUANTITATIVE statistical optimization (burstiness, POS, TTR) - all metrics in one pass
• Phase 10: Pure WORD FILTERING (prohibited words only)

Benefits:

• Single-pass statistical analysis = more efficient
• Coordinated metric optimization = balanced results
• Clear conceptual boundaries = easier to understand and maintain
• Optional statistics phase = skip if text is already optimized

Domain Specialization

• Each phase handles exactly one improvement type
• Clear boundaries prevent interference between phases
• Specialized expertise for consistent results

Pattern-Based Intelligence

• Dialogue Pause Rules - Eliminates "weight of words", "silence stretched", etc.
• Light Description Rules - Replaces "filtering through", "casting shadows" with simple alternatives
• Finger Action Rules - Converts "fingers dancing" to direct action verbs like "typing"
• Pattern matching catches creative variations automatically

NEW: Research-Based Detection Countermeasures

Based on academic AI detector research, ClaudeHumanizer now includes targeted countermeasures for the latest detection methods:

Phase 3 - Nominalization Conversion (v2.4.0)

• Converts AI's abstract nominalized constructions to direct verbal forms
• Example: "The implementation of the solution" → "They implemented the solution"
• Addresses HIGH-priority detection marker explicitly identified in research

Phase 8 - Punctuation Inconsistency Injection (v4.1.0)

• Breaks AI's "machine-like consistency" in punctuation patterns
• Strategic Oxford comma variation, spacing inconsistencies, hyphenation variation
• Addre