PlugWiki

Humanizer (blader/humanizer): a Claude skill that edits out AI-sounding writing

blader/humanizer on GitHub is a Claude skill that rewrites AI-sounding text so it reads like a person wrote it, without changing the facts.

Installs
11.5K
on skills.sh
Stars
63.5K
on GitHub
Updated
Oct 2026
last commit

blader/humanizer, Siqi Chen's skill on GitHub, makes Claude find the habits that make text sound machine-written and rewrite them. It checks 26 patterns, such as "not X but Y" contrasts, one-line closers, forced triples, stock AI words and decorative bold. It keeps your facts and adds none.

It works in four steps: mark the tells, draft a rewrite, check the draft against the patterns and the original, then write the final version. The pattern list is built on Wikipedia's "Signs of AI writing" guide, kept by WikiProject AI Cleanup.

Paste text and you get the draft, a short list of what still sounds artificial, and the final rewrite. Name a file and it rewrites only the prose in that file. Give it two or three paragraphs of your own writing and it matches your voice.

It is not an AI-detector bypass. The README says getting past detectors is not a goal and that detectors still flag most of its output. The skill also leaves Claude's chat replies alone. It edits text you hand it. For another take, see stop-slop.

Good for

  • Writers and marketers who draft with Claude and need to edit out the AI tone
  • Developers who want READMEs, docs and PR descriptions to stop sounding generated
  • Editors and reviewers who check text for AI tells before it goes out
  • Anyone who wants a rewrite that matches a sample of their own writing

Install humanizer

Claude Code

  1. 1

    Copy this prompt.

    Install the humanizer skill in Claude Code. Run this command:

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

Then check that it loaded and tell me how to use it.
  2. 2

    Paste it into Claude Code. It shows the change and asks before running it.

Requires Claude Code 2.1.142 or newer.

Use the plugin on Claude Code 2.1.142 or newer. It answers to /humanizer:humanizer. A skill installed with the skills CLI answers to /humanizer.

From the project README. Source.

What's in the skill

When it loads

Rewrite AI-sounding text so it reads like the writer without changing what it says. Use when editing or reviewing prose for AI tells: not-X-but-Y contrasts, one-line closers, staged openers, forced triads, dashes everywhere, inflated claims, sales language, stock AI words, bold labels, or filler. Based on Wikipedia's "Signs of AI writing."

Sections in SKILL.md

  1. Why AI text sounds the way it does
  2. How to work
  3. Voice
  4. What to return
  5. A. Staging instead of stating
  6. 1. Not X but Y
  7. 2. One-line closers and dramatic fragments
  8. 3. Sayings that sound deep
  9. 4. Staged run-up before the point
  10. 5. Arguing with no one
  11. B. Rhythm by rule
  12. 6. Forced triads
  13. 7. Repeated sentence openings
  14. 8. Dashes as the universal connector
  15. 9. Stacked qualifiers
  16. 10. Hyphenated pairs everywhere

Key facts

Author
Siqi Chen (GitHub: blader)
License
MIT
Latest release
v3.1.0, 2026-09-28
First commit
2026-01-17
SKILL.md last changed
2026-09-28 (repo: 2026-10-28)
GitHub stars
63,467 as of 2026-11-02
Installs on skills.sh
11,524 as of 2026-11-01
Based on
Wikipedia: Signs of AI writing
Works in
Claude Code, Codex, Claude.ai and Claude Desktop, plus other agents through the skills CLI
skills.sh audits
Gen Agent Trust Hub pass, Socket pass, Snyk pass

What it includes

  • Five patterns that stage a point instead of making it

    Not X but Y contrasts, one-line closers and dramatic fragments, sayings that sound deep, staged run-ups such as "Let's dive in", and arguing with no one. Patterns 1 to 5 justify an edit on a single sighting.

    Group A: Staging instead of stating

  • Six patterns of sentence rhythm applied by habit

    Forced triads, repeated sentence openings, dashes as the universal connector, stacked qualifiers, hyphenated pairs everywhere, and passive voice with missing subjects. Dashes, qualifiers, hyphens and passive count only when other tells appear nearby.

    Group B: Rhythm by rule

  • Seven patterns that puff up plain facts

    Overused AI words, inflated significance, vague connection words, shallow -ing riders, sales language, borrowed authority such as "experts believe", and avoiding plain is, are and has.

    Group C: Inflation and borrowed authority

  • Three patterns of decoration

    Bold used as decoration and bold-label lists, decorative headings with Title Case, emojis and arrows, and curly quotation marks. Curly quotes count only alongside other tells.

    Group D: Formatting by rule

  • Four patterns left over from the chat window

    Chatbot residue such as "Great question!", knowledge-limit disclaimers and guesses, a heading repeated in the first sentence, and writing about the document instead of its subject.

    Group E: Leftovers from chat and draft

  • One pattern, for replies only

    Re-explaining what the reader already knows. In a reply, the skill leads with the decision.

    Group F: Writing for the wrong reader

  • Mark the tells, draft, check, write the final

    It reads the whole text and marks every pattern, strongest first. It drafts a rewrite that keeps every supported claim and adds no fact, name, number, date, quote or citation. It then reads the draft against the patterns and the original, and writes the final version.

    Try: “Humanize this text: [your text]”

    The four steps

  • What you get back depends on how you call it

    Pasted text returns the draft, a short list of what still sounds artificial, and the final rewrite. A file path writes only the final prose and leaves code, frontmatter and link targets alone. Embedded use, such as a commit message, returns only the final text.

    Try: “Humanize the prose in docs/launch-post.md”

    Output modes

  • Your writing sample overrides the patterns

    Paste two or three paragraphs of your own writing and the rewrite follows your rhythm, word choice and punctuation, including dashes if you use them. Without a sample, personal writing keeps its opinions and technical text stays neutral.

    Voice matching

  • Quotes, titles and names stay as written

    It does not edit quotations, titles or names. Text from before November 30, 2022 is not treated as AI-written. It relies on several tells together, not one.

    When it leaves text alone

From the official docs

What people say

People like the cleaner tone and the fact preservation, and many note that it will not fool AI detectors.

What people like

  • Users say it catches a lot of AI tone when added to a review step.
  • A user replayed it on 18 drafts: important details kept in 17 and length cut about 7%.
  • Its rewrite beat the original on quality 16 times out of 16 in the author's blind test.
  • Recommended when you want the rewrite to match a sample of your own writing.

Common complaints

  • Detectors such as Pangram still flag most of its output.
  • It does not change how Claude talks in chat, which surprises some users.
  • Some readers say they can still spot the writing after a humanizer pass.
  • Skeptics say a model told to drop its habits drifts back to them.

In their words

GitHub issueAug 2026
“It is not currently possible to defeat Pangram through pure LLM generation through any skill or prompt (I have tried!)”
bladerGitHub issueApr 2026
“I think though, maybe you all are missing the point, the output of this *reads* less like AI output to a human audience.”
greg-randallGitHub issueJul 2026
“The skill does NOT change how the AI talks to _you_, it changes how the text it prepares for you is formatted and written.”
davidsneighbourHacker NewsJul 2026
“I use the humanize skill to clean up AI written work before handing it over to colleagues.”
3uler
3 more quotes
Hacker NewsAug 2026
“It’s not perfect but definitely catches lots of slop.”
moltarHacker NewsSep 2026
“I find LLM writing unbearable (even with humanizer skills and all the other bullshit, I can smell it from a mile away).”
bdangubicBlogSep 2026
“Pick Humanizer when you want the rewrite to match a sample of your writing, and add `deslop` for code.”
Flavio Copes

Examples

  • Rewrite pasted text with the slash command (README).

    /humanizer [paste your text here]

    Source

  • Ask in plain language (README).

    Please humanize this text: [your text]

    Source

  • Rewrite the prose in a file and leave code and frontmatter alone (README).

    Humanize the prose in docs/launch-post.md

    Source

  • Match your own voice with a writing sample (README).

    /humanizer Here's a sample of my writing for voice matching: [paste 2-3 paragraphs of your own writing] Now humanize this text: [paste AI text to humanize]

    Source

  • Codex default prompt from the repo's agents/openai.yaml.

    Use $humanizer to rewrite this text in my voice without changing its facts.

    Source

  • Clean a pull request description before you post it. The skill's embedded mode returns only the final text.

    Write the PR description for this branch, then run it through /humanizer.

Comparison

SkillAuthorCleans upHow it worksInstall
humanizerThis pageSiqi Chen (blader)Writing26 patterns from Wikipedia's guide; drafts, self-checks, can match your voice/plugin install humanizer@humanizer
stop-slopHardik PandyaWritingPhrase and structure lists; scores text 1-10 on 5 dimensionsnpx skills add https://github.com/hardikpandya/stop-slop --skill stop-slop
unslopLauren Tan (pstack)Writing28 numbered rules; runs when you type /unslopnpx skills add https://github.com/cursor/plugins --skill unslop
deslopCursor (cursor-team-kit)CodeRemoves habit-driven code from the branch diff against mainnpx skills add https://github.com/cursor/plugins --skill deslop

Releases

active Six releases since July 2026 and a commit on 2026-10-28.

  • v2.0.0 2026-01-17: Complete rewrite based on the Wikipedia guide, committed the same day as the first commit.
  • v2.9.0 and v2.9.1 2026-07-22: Two releases on the same day.
  • v2.11.0 2026-08-17: Release. v2.11.1 followed on 2026-08-18.
  • v3.0.0 2026-09-06: Consolidated the pattern list, which had grown to 35, into 25.
  • v3.1.0 2026-09-28: Current release with 26 patterns. The last commit, on 2026-10-28, dropped the $schema key from the Claude plugin manifest.

Questions

How do I install the humanizer skill in Claude Code?

Use the plugin route in the install box: add the blader/humanizer marketplace, then install the plugin. It needs Claude Code 2.1.142 or newer, and the plugin answers to /humanizer:humanizer. On an older version, use the skills CLI command listed second. A skill installed that way answers to /humanizer.

Does humanizer work in Claude.ai, Claude Desktop, Codex or ChatGPT?

Claude.ai and Claude Desktop: yes, download the repo as a ZIP and upload it as a skill in Settings. Codex: yes, with the skills CLI command in the install box. ChatGPT is not covered by the README. The skills CLI command for every agent also names Gemini CLI, GitHub Copilot and Windsurf.

Does humanizer get past AI detectors like Pangram, GPTZero or Turnitin?

Not reliably, and that is not its goal. The README says detectors still flag most of its output, and the author wrote that no skill or prompt can currently beat Pangram. One user reported a low Turnitin score, while others report Pangram catching every rewrite. Use it to improve how text reads, not to hide where it came from.

What patterns does humanizer remove?

It checks 26 patterns in six groups. A is staging instead of stating, B is rhythm by rule, C is inflation and borrowed authority, D is formatting by rule, E is chat and draft leftovers, and F is re-explaining what the reader already knows. The tool list above names every group. Counts changed over time, so pages that quote 24 or 29 patterns are out of date.

Where does the pattern list come from?

From Wikipedia's "Signs of AI writing" guide, maintained by WikiProject AI Cleanup, plus reviews of AI text on Wikipedia and elsewhere. That page is advice for editors, not policy. It warns against treating the signs as the only problems to fix, because that can make detection harder.

Humanizer vs stop-slop: which should I use?

Use humanizer when you want a rewrite pass that keeps your facts and can match a sample of your writing. Use stop-slop when you want a shorter rule set you can edit into your own version. The comparison table above also covers unslop and deslop, which cleans up code instead of prose.

Why does Claude still use em dashes in chat after I installed it?

The skill rewrites text you hand it, such as a draft or a file. It does not change how Claude talks to you in chat. Run /humanizer on the text you want cleaned.

Is there a Chinese version of humanizer?

Yes, community-made. Humanizer-zh by op7418 is a Chinese adaptation based on humanizer v3.0.0, and it installs as /humanizer-zh. jiji262/humanizer-chinese covers Simplified Chinese, and nagameTW/humanizer-zh-tw covers Traditional Chinese. The core repo keeps language variants out.

Are there humanizer skills for Finnish, Korean or French?

Finnish: Hakku/finnish-humanizer has 27 patterns for AI-generated Finnish. French: ferr079/humanizer-fr is a fork of humanizer v3.1.0 with 29 patterns adapted to French. Korean: several unrelated projects share the name, including dotoricode/korean-humanizer and a skill inside NomaDamas/k-skill, so check the repo before you install.

Is the humanizer skill safe to install?

Yes, the risk is low. It is one file of instructions. It runs no code and makes no network calls. skills.sh lists passing results from Gen Agent Trust Hub, Socket and Snyk. One thing to know: in file mode it writes the rewrite into the file you name, so commit or back up first.

Information

updated Nov 2, 2026

Repository
blader/humanizer

blader/humanizerskills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 23
  1. humanizer README · github.com
  2. humanizer SKILL.md · github.com
  3. Wikipedia: Signs of AI writing · en.wikipedia.org
  4. Blind study: the rewrite pass wins 16/16 on quality, and does not move AI detection · github.com
  5. humanizer releases · github.com
  6. humanizer on skills.sh · skills.sh
  7. Commits to SKILL.md · github.com
  8. Turnitin report · github.com
  9. Pangram report · github.com
  10. humanizer CHANGELOG · github.com
  11. stop-slop · github.com
  12. still talks with emdashes even after installation · github.com
  13. Humanizer-zh · github.com
  14. Language variants · github.com
  15. humanizer-chinese · github.com
  16. finnish-humanizer · github.com
  17. humanizer-fr · github.com
  18. korean-humanizer · github.com
  19. korean-humanizer on skills.sh · skills.sh
  20. validate-package.py · github.com
  21. Choosing anti-slop skills for writing and code · flaviocopes.com
  22. humanizer replay on 18 drafts · github.com
  23. agents/openai.yaml · github.com

Updated Nov 2, 2026.