Humanizer AI Writing Skill Guide
A practical guide to the Humanizer agent skill, which rewrites AI-sounding text while preserving source claims, code, data, links, and writer intent.
Humanizer AI Writing Skill Guide
Key takeaways#
Humanizer is an agent skill from the blader/humanizer GitHub repository. Its README says it rewrites AI-sounding text so it reads like a person wrote it, without changing what the text says. The skill is distributed as Markdown, which means it can work with agents that support skill-style instructions.
The strongest use case is editing. Humanizer is not a research agent, fact generator, or content spinner. The README is explicit that names, numbers, dates, quotes, citations, and other factual details must come from the source or the writer.
What Humanizer does#
Humanizer targets common signals of AI-generated writing. The repository says it uses 35 patterns from Wikipedia's “Signs of AI writing,” maintained by WikiProject AI Cleanup. It first rewrites the text without treating the original structure as fixed. It then checks the draft against those patterns and against the original claims before rewriting anything that still sounds artificial.
That workflow makes Humanizer useful for documentation, release notes, blog drafts, support replies, and internal memos where the facts are already known. It can also help technical teams avoid the generic rhythm that often appears in LLM output.
How to use it#
Start with a source draft. Ask your agent to use Humanizer on a file or pasted text. The README says the skill shows its work before the final version, including a first rewrite and a short critique of what still sounds artificial. If the input is a file, Humanizer changes prose while leaving code, data, frontmatter, and link targets alone.
For personal writing, provide a writing sample. The README says Humanizer can follow the writer's sample instead of its default style rules. For technical or reference prose, keep the target tone neutral and plain.
Verification checklist#
Before accepting the output, check every name, number, date, quote, and citation against the original source. Confirm that code blocks, YAML, JSON, markdown links, and frontmatter were not changed. Read the final version aloud; the best result should sound direct, not theatrical.
When not to use it#
Do not use Humanizer to hide fabricated claims. It is also a poor fit when the original draft lacks facts, when legal wording must remain exact, or when a strict brand style guide requires sentence-level review.
Source#
Primary source: blader/humanizer on GitHub.