# The Humanizer Workflow: A 4-Step Process for Naturalizing AI Text

> Transform AI text with the Humanizer workflow. Learn the 4 steps to eliminate robotic phrasing and achieve natural, human-sounding prose.

- Repository: [Siqi Chen/humanizer](https://github.com/blader/humanizer)
- Tags: how-to-guide
- Published: 2026-09-09

---

**The Humanizer workflow consists of four distinct steps—marking AI "tells," drafting a rewrite, checking for remaining robotic phrasing, and writing a final polished version—to transform AI-generated text into natural, human-sounding prose.**

The `blader/humanizer` repository provides a structured methodology for converting AI-generated output into authentic, human-like writing. Defined primarily in the repository's [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) file, the Humanizer workflow implements a systematic four-stage process that preserves factual accuracy while eliminating robotic linguistic patterns. This guide examines each step in detail and demonstrates how to execute them using the Claude skill configuration.

## Step 1: Mark the Tells

The workflow begins with a comprehensive scan of the input text to identify and flag every AI-generated pattern, or "tell." In [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md), this step requires scanning the entire input once and flagging patterns starting with the strongest indicators first.

This detection phase targets both sentence-level clues and larger-scale structural issues. Writers must identify repeated closers, staged introductions, and other formulaic constructions typical of AI output. The marking process creates a roadmap for the subsequent rewrite by cataloging every element that contributes to the artificial tone.

## Step 2: Draft the Rewrite

Once tells are marked, the workflow moves to drafting. This stage produces a first rewrite that preserves every supported claim from the original text while restructuring the prose for human flow.

According to the source code analysis, you may trim dull sections, merge or split paragraphs, and rearrange structure as needed. However, the workflow imposes strict constraints: you must **never add new facts, names, dates, quotes, or citations** that are not present in the source material. If a missing detail is required for clarity, the workflow dictates that you either ask for the information or compose a simpler sentence that does not require the missing data.

## Step 3: Check the Draft

The third stage involves quality assurance through oral reading and pattern verification. Practitioners read the draft aloud to detect any remaining AI-sounding phrasing that survived the initial rewrite.

This verification step confirms that no factual content was added or lost during editing, unless a specific pattern explicitly required deletion. The workflow specifically highlights five common surviving tells to watch for during this review:

- **Not-X-but-Y** contrasts
- **One-line closers**
- **Dash** constructions
- **Triad** structures (rule of three)
- **Bold labels**

Identifying these persistent patterns ensures the final output achieves the desired natural tone.

## Step 4: Write the Final Version

The final step delivers a polished version where each point is stated naturally without "patch-by-patch" fixes. This stage focuses on holistic refinement rather than discrete corrections.

Writers adjust sentence length for variety and rhythm. If any sentence remains awkward after individual editing, the workflow requires rewriting the surrounding paragraph to ensure smoother flow. This comprehensive approach prevents the choppy, stitched-together quality that often characterizes partially edited AI text.

## Implementation and Usage Examples

The Humanizer workflow is accessible through the Claude skill defined in [`.claude-plugin/plugin.json`](https://github.com/blader/humanizer/blob/main/.claude-plugin/plugin.json). The following examples from [`README.md`](https://github.com/blader/humanizer/blob/main/README.md) demonstrate practical invocation methods.

### Direct Command Invocation

Invoke the workflow using the slash command:

```text
/humanizer

[paste your AI‑generated text here]

```

Or use natural language prompting:

```text
Please humanize this text: [your text]

```

### File Mode Processing

Process entire documents by specifying file paths:

```text
Humanize the prose in docs/launch‑post.md

```

### Voice Matching

Match a specific writing style by providing a sample:

```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 text to humanize]

```

## Repository Structure and Key Files

The Humanizer workflow is implemented across three primary files in the repository:

- **[`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md)** – Contains the complete four-step workflow definition, full prompt specifications, and all pattern definitions for identifying AI tells.
- **[`README.md`](https://github.com/blader/humanizer/blob/main/README.md)** – Provides installation instructions and concrete usage examples that illustrate each workflow stage.
- **[`.claude-plugin/plugin.json`](https://github.com/blader/humanizer/blob/main/.claude-plugin/plugin.json)** – Links the skill to Claude agents, exposing the workflow through the `/humanizer` command interface.

## Summary

- **Mark the tells** by scanning input and flagging AI patterns from strongest to weakest indicators.
- **Draft the rewrite** while preserving all factual claims and avoiding the introduction of new data.
- **Check the draft** by reading aloud and eliminating the five common surviving tells: not-X-but-Y contrasts, one-line closers, dashes, triads, and bold labels.
- **Write the final version** with natural flow, varied sentence lengths, and paragraph-level smoothing rather than patchwork fixes.

## Frequently Asked Questions

### What are the four steps in the Humanizer workflow?

The four steps are: (1) Mark the tells by scanning and flagging AI patterns, (2) Draft the rewrite while preserving facts and restructuring prose, (3) Check the draft by reading aloud and removing remaining robotic phrasing, and (4) Write the final version with natural flow and adjusted sentence variety.

### Where is the Humanizer workflow defined?

The complete workflow definition resides in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) at the root of the `blader/humanizer` repository. This file contains the full prompt instructions, the four-step methodology, and detailed pattern definitions for identifying AI-generated text characteristics.

### Can Humanizer add new facts during the rewrite?

No. The workflow explicitly prohibits adding new facts, names, dates, quotes, or citations that are not present in the source material. If a detail is missing but required for clarity, the instructions specify either requesting the information from the user or writing a simpler sentence that does not depend on the missing data.

### How does voice matching work in Humanizer?

Voice matching allows the workflow to adapt to a specific writing style by analyzing a sample of the user's own writing. You provide 2–3 paragraphs of your previous work before submitting the AI text to humanize, allowing the system to mirror your personal tone, rhythm, and stylistic preferences in the output.