# Humanizer Four-Step Rewrite Process: Converting AI Text to Natural Prose > Discover Humanizer's four-step rewrite process to convert AI text into natural prose. Learn how to transform robotic writing into human-like content effectively. - Repository: [Siqi Chen/humanizer](https://github.com/blader/humanizer) - Tags: how-to-guide - Published: 2026-09-05 --- **Humanizer transforms AI-generated content into human-like prose by scanning for 35 distinct AI-writing patterns, drafting and reviewing the text aloud, validating factual integrity through two critical questions, and applying final polish including dash-rule corrections.** The `blader/humanizer` repository implements a rigorous **four-step rewrite process** that strips robotic phrasing from machine-generated text while preserving every factual claim. Defined in the repository’s [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) file, this workflow systematically eliminates AI artifacts—from inflated importance statements to dash misuse—across all input modes including pasted text, file mode, and embedded mode. ## Step 1: Identify and Tag AI Patterns The process begins with a forensic scan of the source material. According to [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 42-44, Humanizer reads the input and marks each instance of the **35 AI-writing patterns** it tracks, including: - **Inflated importance** – exaggerated significance claims typical of LLM outputs - **Overused AI words** – terms that signal synthetic authorship - **Dash misuse** – incorrect em-dash or en-dash application Each detected pattern is tagged for subsequent correction, creating a diagnostic map of the text’s robotic elements before any rewriting occurs. ## Step 2: Generate Draft and Auditory Review Once patterns are flagged, Humanizer produces a first-pass rewrite. As specified in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 44-45, the system then performs an internal auditory review, reading the draft aloud to evaluate: - **Rhythm and flow** – sentence cadence and structural variety - **Verb choice** – active versus passive constructions - **Formality calibration** – tone matching the intended audience This review simulates human editing practices, ensuring the draft moves beyond mere grammatical correctness to achieve natural readability before proceeding to validation. ## Step 3: Ask Two Critical Questions Before finalizing, the workflow enforces a strict quality gate defined in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 45-48. The system interrogates the draft with two mandatory questions: 1. **"What still sounds AI-generated?"** – Any remaining patterns are flagged for additional editing cycles until the prose passes human-voice detection. 2. **"Did the rewrite add or remove any fact, name, number, date, quote, citation, ranking, or other claim?"** – This ensures **factual fidelity**. Unsupported additions or lost claims are treated as errors; the draft must preserve all original factual content exactly, changing only stylistic presentation. ## Step 4: Write the Final Version The concluding phase refines the approved draft into final prose. Per [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 48-49, this step addresses any remaining awkward sentences by rewriting the surrounding paragraph rather than isolated words. Crucially, the system applies the **dash rule** (documented in section 14 of the skill definition), correcting punctuation misuse identified in step 1. The output contains no invented facts and maintains the writer’s original voice, having passed through the complete four-step workflow. ## How to Invoke the Four-Step Process Humanizer exposes this workflow through a simple CLI interface. The process executes identically whether processing clipboard content or entire files. ### Pasted Text Mode ```bash # Direct skill invocation for clipboard content /humanizer [Paste AI-generated paragraph here] ``` The engine executes the four steps sequentially: highlighting detected patterns, returning a draft rewrite, running the two critical questions automatically, and outputting the final humanized version. ### File Mode ```bash # Rewriting an entire document Humanize the prose in docs/launch-post.md ``` When operating on files, Humanizer reads [`docs/launch-post.md`](https://github.com/blader/humanizer/blob/main/docs/launch-post.md), runs the four-step process on each paragraph, and writes only the revised prose back to the file. Upon completion, it prints a summary of changes made, with [`scripts/validate-package.py`](https://github.com/blader/humanizer/blob/main/scripts/validate-package.py) ensuring the skill metadata remains synchronized with the workflow definitions. ## Summary - **Pattern Detection** – [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 42-44 define the initial scan for 35 AI-writing artifacts including dash misuse and inflated language. - **Auditory Review** – Drafts are evaluated for rhythm, verb strength, and formality before validation. - **Factual Integrity** – Two mandatory questions prevent hallucinations and ensure no claims are added or removed during rewriting. - **Final Polish** – The dash rule (section 14) and contextual paragraph rewriting produce publication-ready human prose. - **Universal Application** – The workflow functions identically across pasted text, file mode, and embedded implementations. ## Frequently Asked Questions ### What are the 35 AI-writing patterns Humanizer detects? The patterns include structural markers like **inflated importance statements**, **overused AI terminology**, **incorrect dash usage**, and other syntactic tells typical of large language model outputs. The complete catalog lives in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md), where each pattern is defined with specific detection criteria to guide the rewriting process. ### How does Humanizer ensure factual accuracy during rewriting? During step 3, the system explicitly asks whether any "fact, name, number, date, quote, citation, ranking, or other claim" was altered. According to the source code in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 45-48, any unsupported addition or lost claim is treated as an error, forcing a rewrite until factual parity with the source is guaranteed. ### Can I process entire documents or only pasted text? Humanizer supports **three input modes**: direct pasted text via the `/humanizer` command, **file mode** for rewriting complete documents like [`docs/launch-post.md`](https://github.com/blader/humanizer/blob/main/docs/launch-post.md), and embedded mode for API integrations. The four-step process executes identically regardless of input method. ### What is the dash rule applied in step 4? Documented in section 14 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md), the dash rule corrects em-dash and en-dash misuse identified during the initial pattern scan. Rather than fixing punctuation in isolation, the system rewrites the surrounding paragraph to ensure the corrected dashes flow naturally within the sentence rhythm, preserving the human voice established in earlier steps.