How to Use the Humanizer Claude Skill and Keep Your Own Voice

Humanizer is a skill that edits the AI sound out of a draft. At first that seems backwards.

A model wrote the flat sentences. Now a model is supposed to make them sound like you?

How would it know what you sound like? And what keeps it from changing what you said?

Both questions are fair. It doesn’t know your voice until you show it some of your writing. And it can change your meaning, so you compare the edit with the original.

Here is the short version. Install Humanizer, paste your draft into your agent, and ask for an edit. To keep your own voice, add two or three paragraphs you wrote earlier. When the edit comes back, read it against the original. Pay most attention to the facts, and to anything you weren’t sure about.

How do you install Humanizer and run it?

With Claude Code, run this command in your terminal first.

npx skills add blader/humanizer --global

You need Node.js and npm. npm’s guide covers installing them.

--global installs the skill for every project. Leave it off to install only in the current one.

Reload Claude Code’s skills after the install. Then go back to the chat and enter:

Use Humanizer to edit the blog post below.
Keep my opinions, my facts and the things I'm unsure about. Cut filler and repetition.
Don't invent experiences or add conclusions for me.

[paste your draft]

In Claude Code you can also type /humanizer and paste your text. The project README lists the other ways to install it.

/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer

What does Humanizer change?

Humanizer (blader/humanizer on GitHub) is an open-source writing skill. It is a Markdown file of editing instructions.

The agent you already use reads the file and follows it. So you need an agent that supports skills.

For a first try, pick a short passage where you know exactly what you meant. Then you can tell whether the edit changed it.

The example below is a work summary about organizing research notes. Every example in this article was constructed to explain the editing method. None is a real work record or actual Humanizer output.

Last Saturday I put my notes from 18 interviews into one folder. This reorganization greatly improved the efficiency of our data management and provides strong support for future research. There is one line I want to quote, though, and I can’t confirm who said it, because I didn’t write down names at the time.

(1) It cuts judgments the draft can’t support.

The passage says how many notes were organized, and what went wrong with the quote.

It gives no evidence that anything became more efficient. And “strong support” never says what was supported.

A first pass can delete those two judgments.

Last Saturday I put my notes from 18 interviews into one folder. There is one line I want to quote, but I can’t confirm who said it, because I didn’t write down names at the time.

(2) It keeps the problem you haven’t solved.

The open problem is the unknown speaker. Putting the notes in one folder doesn’t bring back names that were never written down.

The edit keeps that problem visible.

(3) It checks its own draft.

Humanizer’s editing procedure has the agent mark the problems before it rewrites. Then the agent checks the draft for leftover filler and changed facts, and writes the final version.

When you paste text, it returns the first draft and a short check along with the final version. If you only want the final version, add: Return only the final text.

(4) It can’t add details you never gave it.

The edit above already states the problem with the notes. What happened next? The original never said, and Humanizer has no way to know.

Its rules tell the agent to ask for a missing detail, or to write a simpler sentence. If you want the follow-up in the piece, you have to supply it.

Suppose the author then provides this note. It can go after the edited passage.

I went back to the recordings and added speaker names to the notes from 6 interviews. I’m leaving the other 12 until I need them.

The note says that 6 sets of notes now have names. It doesn’t say whether the speaker of that one quote was identified.

So the edit can’t report the attribution problem as solved.

Leaving the other 12 alone is the author’s decision. It should stay in.

How do you keep your own voice?

Sometimes the filler is gone and the post still doesn’t sound like you. Then give Humanizer some of your own writing.

(1) Pick two or three paragraphs you’re happy with.

Try to pick pieces written for the same kind of reader. The tone of a personal essay may not suit a technical tutorial.

(2) Say what you want kept.

It might be the occasional joke at your own expense. Or the sentences where you admit that something you tried didn’t work.

That tells the agent what it shouldn’t polish away.

(3) Watch for tidy sentences that mean something else.

Back to the work summary. Suppose the author also wrote this.

I meant to finish all 18 in one go, and I only got through 6. The other 12 can sit for now. I don’t want to listen to my own interview recordings again for a while.

Polished like this, the passage is tidier and means something else.

Speaker information has been added for 6 interviews, and the remaining 12 will be completed in phases as the research requires.

The original holds a plan that fell short. It also holds an author who doesn’t feel like listening to more recordings.

“Completed in phases” covers both. It also gives the author a plan they never made.

For a personal blog, it’s enough to tighten the sentences.

I meant to finish all 18 and got through 6. The other 12 can wait. I don’t want to listen to my own interview recordings again for a while.

This version keeps the reluctance. It doesn’t add a joke to seem friendly.

When you give Humanizer a sample, you can tell it directly to keep this kind of writing.

Use Humanizer to edit this blog post.

Voice reference:
[paste 2–3 paragraphs of your own writing]

Use the reference only for tone and sentence rhythm.
Take facts from this draft and the source material.
Keep my hesitation, my jokes at my own expense and my personal judgments.
Don't invent experiences, add opinions or tell jokes for me.

Source material:
[paste the relevant facts; leave out if there are none]

Draft to edit:
[paste your draft]

Under Humanizer’s voice rules, your sample takes priority over its general editing rules. That includes your punctuation habits.

(4) A sample doesn’t always make a visible difference.

Bing Wu’s earlier Humanizer vs. Stop Slop comparison added a writing sample to a short English draft. The result didn’t meaningfully change.

That was one draft and one sample. Try it on your own writing.

What should you check after the edit?

Put the original and the edit side by side. A new version that reads smoothly makes changes in meaning easy to miss.

For the example above, the check looks like this.

(1) Watch the small qualifying words.

Words like “some,” “may” and “in this test” can look like padding. Deleting them can turn a single observation into a general claim.

(2) Reject additions the original can’t back up.

Suppose an edit says the reorganization “saved time finding notes.” The original gives no basis for that. Say so.

Delete "saved time finding notes." I didn't measure any time.
Keep the problem that I can't confirm who said the quote, and keep the line about not wanting to listen to more recordings.
Don't add follow-up plans for me. Keep the other edits.

(3) Read the piece yourself.

Fix the sentences that sound awkward. If a sentence already sounded like you before the edit, it’s fine to change it back.

Can Humanizer edit a Markdown file directly?

Yes. If the article is already saved as Markdown, give Humanizer the file path.

Keep a copy first, or keep the original in version control. Then you can compare afterwards.

Use Humanizer to edit the prose in drafts/research-notes.md.
Keep the limits on each claim and my first-person judgments.

Humanizer’s file mode tells the agent to write the final text back to the file. It leaves code, commands, paths, YAML metadata and link targets unchanged.

If your tutorial has runnable examples, check those parts before you publish anyway.

If you keep your past writing in StashBase, you can search it for a passage to use as a voice reference. You can also keep the source material next to the draft you’re editing.

For a one-off edit, pasting a few paragraphs is enough.

Related: Humanizer and similar Claude skills covers Stop Slop and No AI Slop, and How to Write with AI Without Losing Your Voice covers the writing method around the edit.

Disclosure: I’m one of StashBase’s developers. Humanizer guides the edit inside your agent; StashBase can be used to organize the material. This article is based on the project README and version 3.0.0 of the skill, checked on September 19, 2026. The interview-notes examples are editorial illustrations. No separate comparison test was run, and no AI-detection scores were evaluated.