Humanizer Skill for Claude: What It Does and How to Use It

The humanizer skill is an open-source project on GitHub under the MIT license. Its README says the skill "makes AI-written text sound like a person wrote it, without changing what it says." Anthropic doesn't list a humanizer among Claude's built-in skills, which its help center describes as tools for spreadsheets, Word documents, presentations and PDFs. People also say "Claude humanizer" for saved prompts, chatbots built on top of Claude and dedicated tools such as the Certentis Claude Humanizer, which rewrites a pasted Claude text in a separate step and measures the result.

A skill is a set of instructions that Claude loads when a task calls for it. Anthropic defines skills as "folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks." The core of every skill is a file called SKILL.md: a short header with a name and a description, followed by instructions in plain Markdown. Claude works out which skills are relevant to a request and loads only the instructions it needs, which Anthropic calls progressive disclosure.

Anthropic developed the format and released it as an open standard called Agent Skills. That's why the same SKILL.md also works in other agents, OpenAI's Codex among them.

The humanizer skill looks for 26 writing patterns and rewrites the passages where they appear while keeping the claims. The list is built on Wikipedia's "Signs of AI writing", a field guide that Wikipedia's AI Cleanup project maintains for spotting undisclosed AI-generated content in articles. The skill sorts the patterns into six groups, strongest first:

The instructions in SKILL.md set out four steps. Claude marks every pattern it finds, drafts a rewrite, checks the draft against the patterns and the original claims, and then writes the final version. With pasted text, you get the first draft, a short critique and the final version. Pointed at a file, the skill changes only the prose and leaves code, data and link targets alone. Claude is also told not to invent anything. If a sentence needs a name, number or source that's missing, it should ask you.

The skill's README states the scope plainly: "Humanizer edits for human readers. Getting past AI detectors is not a goal, and detectors still flag most of its output." The Wikipedia page is careful in the same way. It calls its list descriptive, a set of observations rather than rules, and notes that not every text with these signs is AI-generated.

In the Claude web and desktop apps, you upload the skill as a ZIP file. Skills are available on Free, Pro, Max, Team and Enterprise plans and need code execution turned on. These steps follow Anthropic's help article on skills and the skill's README:

1. Turn on code execution. On Free, Pro and Max plans, go to Settings > Capabilities and switch on "Code execution and file creation". On Team and Enterprise plans, the setting sits with an owner, under Organization settings > Plugins & skills in the Policy tab.
2. Download the skill. On the GitHub page of blader/humanizer, choose Code > Download ZIP.
3. Upload it. In Claude, open Customize > Skills, click the "+" button and then "+ Create skill", select "Upload a skill" and choose the ZIP file.
4. Switch it on and use it. The skill appears in your list, where you can toggle it on or off. Claude uses an enabled skill when a request matches its description. If it doesn't, ask directly, for example: "Use the humanizer skill on this text."

GitHub names the folder inside the ZIP humanizer-main, while the skill itself is called humanizer. Anthropic lists a folder name that doesn't match the skill name among the common reasons an upload fails. If that happens, unpack the ZIP, rename the folder to humanizer and zip it again.

In Claude Code, the README installs the skill as a plugin with two commands:

1. /plugin marketplace add blader/humanizer
2. /plugin install humanizer@humanizer

As a plugin, the skill answers to /humanizer:humanizer and needs Claude Code 2.1.142 or newer. On older versions, the README uses the Skills command-line tool instead: npx skills add blader/humanizer --global --agent claude-code. The Claude Code plugin docs describe the same /plugin syntax. Installing from inside a session opens a panel where you review the plugin and choose its scope.

You can also install it by hand. Claude Code reads personal skills from ~/.claude/skills/<skill-name>/SKILL.md, which makes them available in all your projects on that machine. Save the file as ~/.claude/skills/humanizer/SKILL.md and call it with /humanizer. If you already enabled the skill on claude.ai, Claude Code 2.1.273 or newer loads it as well when you sign in with the same account, though not when it runs with an API key. Run /skills to see it listed under claude.ai sync.

Codex and other agents

For Codex, the README's command is npx skills add blader/humanizer --global --agent codex. OpenAI's documentation on skills lists $HOME/.agents/skills as the folder for personal skills and .agents/skills for skills inside a repository. In the Codex CLI or the IDE extension, run /skills or type $ to mention a skill. For other agents, the README uses --agent '*'. Without --global, each of these commands installs the skill only in the current project.

Tell Claude who will read the text, give it a sample of your own writing, name the changes you want, and list what must stay exactly as it is. That follows Anthropic's prompting guide, which recommends clear, explicit instructions and a reason for each one. It calls examples "one of the most reliable ways to steer Claude's output format, tone, and structure." A prompt you can adapt:

Anthropic's guide also advises telling Claude what to do instead of what to avoid. "State each point directly" gives Claude a target, while "don't sound like AI" leaves it guessing. Anthropic suggests one test for any prompt: show it to a colleague with little context on the task. "If they'd be confused, Claude will be too."

Then read the result against your original, sentence by sentence. If you write the same kind of text often, saving these rules as a skill spares you from typing them again.

A skill and a prompt are both instructions. The same model that may have written the draft rewrites it and then checks its own work. Nothing measures the result afterwards, which leaves four limits:

• Your voice can flatten. Without a sample, the skill picks a voice from the kind of text. Its dash rule removes em and en dashes from the final version unless your sample uses them. If you write page or year ranges with an en dash, check them afterwards.
• Facts, citations and meaning go unchecked. The skill tells Claude to keep every supported claim and add nothing, but nothing compares the rewrite with your sources. Check names, numbers, quotations and references yourself.
• Results vary. Two runs on the same text can come out differently. Anthropic's API reference notes that even with temperature set to 0.0, "the results will not be fully deterministic."
• Authorship stays the same. A rewrite doesn't change who or what produced the draft. If your course or employer asks you to disclose AI use, that still applies. See how to cite ChatGPT and AI and academic integrity.

Wikipedia's page makes a related point. It asks editors not to treat the signs as the problems to be fixed, because they can point to deeper, less visible problems in a text. Which patterns matter, and why none of them proves anything alone, is covered in signs of AI writing. How detectors score text is explained in how AI detectors work.

If your worry is that your own writing will be flagged as AI, a rewrite won't settle it. Dated drafts and version history show how you wrote a text. Our guide on what to do if you're falsely accused of using AI explains how to use them.

A dedicated tool helps when you need checks that a chat doesn't run: whether citations survived, whether the language stayed the same, and how the result scores.

The Certentis Claude Humanizer takes 100 to 30,000 characters and up to 3,000 words per run. The result stays in the language of your text, and the rewrite is built to keep citations, links, DOIs and footnote references as they are. Afterwards, the Certentis AI analysis scores your original and the rewrite. You see Human, Mixed and AI probabilities for the result and how the combined Human and Mixed value changed from your original. The page shows the values the analysis actually returned.

You can try it without an account once every 30 days, with up to 300 words.