Skip to content

Humanizer

What it is

Humanizer is a core community skill for Claude Code, OpenCode, and agentic platforms supporting the MCP 3.0 Task Protocol. It is focused on removing robotic "AI-isms" and making AI-generated output feel more natural and human-readable through pattern-based refinement and voice calibration.

What problem it solves

It addresses the common issue of AI output sounding overly mechanical, generic, or obviously templated. In the era of Claude 5.1 and Gemma 3, it remains the de facto standard for "de-botting" agentic outputs before they are shared with humans or published as final artifacts.

Where it fits in the stack

Development & Ops / Output Refinement. It sits at the Interaction Layer, transforming the "raw" reasoning of an LLM into human-centric communication. It is frequently used in FastMCP 3.0 pipelines to ensure tool-generated reports maintain a natural tone.

Typical use cases

  • Copy Polish: Improving generated marketing copy or blog posts before publication.
  • Agentic Communication: Post-processing drafts from coding or writing agents to ensure they match a team's voice.
  • Voice Matching: Calibrating an AI to write exactly like a specific human user based on writing samples.
  • MCP Tool Output Refinement: Smoothing out raw data outputs from MCP servers into conversational summaries.

Strengths

  • Native Claude Code Integration: Works as a simple slash command (/humanizer) inside the terminal.
  • Pattern-Based Auditing: Derived from the Wikipedia Signs of AI writing guide.
  • Voice Calibration: Supports "soul injection" by matching sentence rhythm and rhythm from user-provided samples.
  • Privacy-First: Operates entirely locally within the agent's context; no external API calls for the humanization logic itself.
  • MCP 3.0 Compatible: Can be orchestrated as a sub-skill within larger agentic workflows.

Limitations

  • Subjectivity: "Human-like" writing is inherently subjective and varies by domain.
  • Model Dependent: Success depends on the underlying model's ability to handle the "soul injection" step (Claude 3.5 Sonnet and newer, or Gemma 3 27B+ are recommended).

When to use it

  • When generated output needs a natural, non-robotic tone for human consumption.
  • When you want to maintain a consistent personal or brand voice across agentic outputs.
  • As a final step in an automated content generation pipeline.

When not to use it

  • In technical documentation where standard "AI-like" clarity (bolded headers, bulleted lists) is actually preferred.
  • When factual accuracy is the primary bottleneck; humanization does not fix underlying hallucinations.

Getting started

Installation

Humanizer is distributed as a Markdown-based skill (SKILL.md). Install it by cloning into your agent's skills directory:

Claude Code 

mkdir -p ~/.claude/skills
git clone https://github.com/blader/humanizer.git ~/.claude/skills/humanizer

OpenCode 

mkdir -p ~/.config/opencode/skills
git clone https://github.com/blader/humanizer.git ~/.config/opencode/skills/humanizer

CLI examples

The skill is invoked using the /humanizer slash command within supported agentic terminals.

Basic Humanization

/humanizer "This model marks a pivotal moment in the evolution of AI."

Voice Calibration

Provide a sample of your own writing for the agent to match: 

/humanizer --calibrate
# [Follow prompt to paste 2-3 paragraphs of your own writing]

MCP-based Refinement

/humanizer --mcp-source logs.txt

API examples

Programmatic Invocation (Markdown Skill pattern)

While Humanizer is a text-based skill, it can be invoked programmatically by injecting the skill definition into a system prompt.

import os

# Conceptual example of injecting Humanizer skill into a system prompt
skill_path = os.path.expanduser("~/.claude/skills/humanizer/SKILL.md")
with open(skill_path, "r") as f:
    humanizer_skill = f.read()

system_prompt = f"You are a helpful assistant. Use the following skill when requested:\n{humanizer_skill}"
# Then call the model (e.g., google("gemma3-27b-it")) with "/humanizer <text>" in the user prompt

Sources / References

Contribution Metadata

  • Last reviewed: 2026-07-21
  • Confidence: high