blader/humanizer

Humanizer is an agent skill for people who draft text with a language model and want the result to read like a person wrote it, not like a model. It rewrites prose in place—removing stock phrasing, inflated significance and formatting habits—while keeping names, numbers, quotes and citations tied to the source. It is not a tool for evading AI detectors; the author says detectors still flag most of its output.

What it does

The pattern list comes from Wikipedia's Signs of AI writing. The strongest five are "not X but Y" constructions, one-line closers like "Let that sink in", sayings that sound deep, staged run-ups ("Here's the thing"), and arguing with an objection nobody raised. Humanizer checks 26 patterns in total, but marks several as weak on their own—dashes, qualifiers, hyphenated pairs, passive voice, curly quotes only justify an edit when several such tells share a passage. This matters because the tool does not treat every stylistic quirk as proof of machine authorship.

How to use it

The skill answers to /humanizer after installation. You can paste text into the chat, ask in plain language, or give it a file path. The most distinctive workflow is voice matching: paste two or three paragraphs of your own writing first, and the rewrite follows that sample's rhythm, word choice and punctuation, including deliberate quirks.

When you paste text, it shows the rewrite, a short critique of anything still artificial, then the final version—so you can see what changed. When you point it at a file, the README says it edits only prose and leaves code, data, frontmatter and link targets untouched. The author states that factual details must come from the source or the writer; if a sentence needs a missing detail, it asks rather than inventing one.

Requirements and platforms

It is a skill, so it needs a compatible agent. The README lists Claude Code (plugin, requiring version 2.1.142 or newer; older versions use the Skills CLI), Codex via npx skills add, Claude.ai and Claude Desktop via a ZIP upload, and the Skills CLI for agents such as Gemini CLI, GitHub Copilot and Windsurf. For agents the CLI does not know, SKILL.md can be copied into the skill folder. The repository is MIT-licensed.

Limits

The blind test cited is a self-reported case: in one review, judges preferred the rewrite over the original 16 times out of 16 (blind-test issue). That is one evaluation, not a general accuracy or style guarantee. The stated goal is readability for human readers, not detector evasion, and the author explicitly says detectors flag most output. Whether a rewrite preserves meaning remains something the user should check, especially for sourced facts.

The project is popular by stars, but a HN commenter noted uncertainty about how well it works (HN comment); stars indicate attention, not measured output quality.