You asked Claude for a blog draft. It came back clean, confident and strangely hollow. Every section closes with a tidy lesson. Three adjectives walk in together. Something “stands as a testament” to something else.
A humanizer skill can fix most of that in one pass. What it can’t do is make a thin article worth reading, and it won’t reliably fool an AI detector either. This guide covers both halves: the skills that work (free on GitHub and paid packs), and the honest limits of what “humanizing” buys you.
Quick answer: The most-used free option is blader/humanizer (about 63,000 GitHub stars as of 1 November 2026, MIT licence), built on Wikipedia’s “Signs of AI writing” guide. Install it with npx skills add blader/humanizer --global --agent claude-code or through /plugin marketplace add blader/humanizer. Pair it with a voice or editing skill, then add real experience yourself. No skill can supply that part.
New to skills? Start with what Claude Code skills are and how to install Claude Code skills.
What does a humanizer skill actually do?
A humanizer skill is a SKILL.md file (plus optional reference files and scripts) that tells Claude how to spot and rewrite the patterns that make machine text read as machine text. It lives in ~/.claude/skills/ (global) or .claude/skills/ (one project), or ships inside a plugin. Claude loads it when your request matches its description, or when you call it by name, such as /humanizer.
The good ones work like a structured copy edit, not a thesaurus. Here is how the leading free skill describes its own process in its README: it marks every tell it finds, strongest first, drafts a rewrite, checks the draft against both the pattern list and the original claims, then writes the final version. It also states that it “does not make things up.” If a sentence needs a name, number or quote that the source lacks, it asks instead of inventing one.
That last rule matters more than any pattern list. A careless humanizer swaps “significant growth” for “growth of 47%” and gives you a fabricated stat that reads more human and is now false.
What a humanizer is not
| It is | It is not |
|---|---|
| A pattern-based edit pass for AI tells | A guarantee against AI detectors |
| A way to cut filler and puffery | A source of facts, stories or expertise |
| Useful on human drafts too | A substitute for knowing the topic |
| One step in an editing workflow | A whole content strategy |
Takeaway: treat a humanizer like a strict line editor with a checklist. It removes noise. It doesn’t add signal.
The common AI tells, according to Wikipedia’s “Signs of AI writing” guide
Most serious humanizer skills trace back to one source: Wikipedia: Signs of AI writing, a field guide written by editors who clean up chatbot text submitted to the encyclopedia. The blader/humanizer README names it directly as “the source for the pattern list.”
The guide groups tells into families. Here are the ones that show up most often in marketing and blog copy.
Language and tone tells
- Puffery and inflated significance. Phrases that call something a “testament” or a “key turning point,” or that describe a “rich” mix of anything. The writing insists something matters without showing why.
- Promotional vocabulary. “Nestled,” “breathtaking,” and the friction-free adjectives you see in ad copy. Fine in an ad, odd in an explainer.
- Superficial -ing tails. A sentence ends with a comma and a present participle (”…, reflecting broader trends”). The clause sounds like analysis and says nothing.
- Vague attribution. “Experts say,” “industry observers note,” with nobody named.
- Negative parallelism. A sentence that denies a small claim to assert a bigger one (“not merely X but Y”).
- Rule of three. Lists of three adjectives or three clauses, used by reflex rather than because there are three things.
Style and formatting tells
- Em dash overuse. Chatbots use them far more often than most human writers.
- Title Case Headings and bolded phrases sprinkled through every paragraph.
- Curly quotes and odd markup pasted from a chat window.
- Tidy section-closing summaries (“In summary…”, “Overall…”) after every heading.
Conversation leftovers
- “I hope this helps!”, “Certainly!”, praise for the question and knowledge-cutoff disclaimers left inside published text.
A caution the guide itself makes
The Wikipedia guide is explicit that these signs are hints, not proof. Humans use em dashes and triplets too. The guide also warns editors against relying on automated AI detectors. That framing should shape how you use any humanizer: the goal is better prose, not a forensic disguise.
| Tell | Before | After |
|---|---|---|
| Puffery | ”This tool represents the future of modern marketing." | "The tool drafts five ad variants from one brief.” |
| -ing tail | ”Sales rose, which shows the strategy worked." | "Sales rose 12% in Q3.” (only if you have the number) |
| Negative parallelism | ”More than software, it is a partner." | "It handles scheduling and invoicing.” |
| Vague attribution | ”Experts agree SEO is changing." | "Google’s 2023 guidance says…” with a link |
Notice the second row. The honest rewrite needs a real number. If you don’t have one, the right fix is to cut the claim, not to invent a figure.
The best free humanizer skills on GitHub
Star counts and last-push dates below come from the GitHub API on 1 November 2026. Stars measure popularity, not quality, so read each SKILL.md before you install it. Every install command was checked against the repo’s own README.
1. blader/humanizer
- Source: github.com/blader/humanizer
- Price: Free, MIT licence
- Stats: about 63,280 stars, last push 28 October 2026
- What it does: Checks for 26 numbered patterns drawn from the Wikipedia guide, ranked by strength. The README says the first five “justify an edit on a single sighting,” while weaker ones (like curly quotes) only count when several tells cluster in one passage. That weighting cuts down on over-editing.
- Install (plugin, Claude Code 2.1.142 or newer):
/plugin marketplace add blader/humanizer /plugin install humanizer@humanizer - Install (Skills CLI, older versions):
Dropnpx skills add blader/humanizer --global --agent claude-code--globalto install it in the current project only. - Also listed on: claudemarketplaces.com and cultofclaude.com.
2. Aboudjem/humanizer-skill
- Source: github.com/Aboudjem/humanizer-skill
- Price: Free, MIT licence
- Stats: about 297 stars, last push 6 September 2026
- What it does: A larger catalogue, 55 patterns per its README, with five voice options, and processing that stays on your machine. The repo includes a pattern reference file with before-and-after pairs and docs for gating commits with a pre-commit check.
- Install:
orclaude plugin marketplace add Aboudjem/10x claude plugin install humanizer@10xnpx skills add Aboudjem/humanizer-skill, or curl the single SKILL.md into.claude/skills/humanizer/as the README shows.
3. hannsxpeter/humanizer
- Source: github.com/hannsxpeter/humanizer (the old
aihxp/humanizerURL redirects here) - Price: Free, MIT licence
- Stats: about 15 stars, last push 8 October 2026
- What it does: A 32-pattern catalogue in six families, each with detect, why and before-after notes. Its README is unusually candid: it lists “statistical token-pattern disruption” as a “side effect only” and says nothing is “tuned for this, verified, or promised.” That’s the right attitude toward detectors.
- Install:
git clone https://github.com/hannsxpeter/humanizer ~/.claude/skills/humanizer
4. content-humanizer in alirezarezvani/claude-skills
- Source: github.com/alirezarezvani/claude-skills, file
marketing-skill/skills/content-humanizer/SKILL.md - Price: Free, MIT licence
- Stats: about 29,110 stars for the whole repo, last push 31 October 2026
- What it does: Per its frontmatter, it “scores AI tells with humanizer_scorer.py and rewrites for rhythm, specifics and personality, then rescores.” The scoring script gives you a before-and-after number, which helps teams agree on what “done” means.
- Install: it ships in the marketing bundle (the README lists 43 marketing skills):
/plugin marketplace add alirezarezvani/claude-skills /plugin install marketing-skills@claude-code-skills
5. hurryingauto3/master-humanizer
- Source: github.com/hurryingauto3/master-humanizer
- Price: Free, MIT licence
- Stats: 1 star, last push 5 August 2026
- What it does: An editorial pass that keeps meaning, citations and code intact. The author states it is not a detector-evasion tool. Low adoption, so treat it as one to read and borrow ideas from rather than a default.
Skills that pair well with a humanizer
A humanizer strips patterns. It doesn’t know your brand, your reader or your offer. These free skills fill that gap.
| Skill | Source | What it adds | Install |
|---|---|---|---|
| copy-editing | coreyhaines31/marketingskills (about 57,800 stars) | Reviews and improves existing marketing copy, or refreshes outdated content | npx skills add coreyhaines31/marketingskills --skill copy-editing |
| copywriting | same repo | Writes conversion copy from a brief | npx skills add coreyhaines31/marketingskills --skill copywriting |
| Marketing plugin (incl. brand-review, draft-content) | anthropics/knowledge-work-plugins (about 34,300 stars, Apache 2.0) | Anthropic’s official plugin: drafts content and enforces brand voice | claude plugin marketplace add anthropics/knowledge-work-plugins then claude plugin install marketing@knowledge-work-plugins |
| Curated lists | ComposioHQ/awesome-claude-skills (about 78,500 stars) | A directory for finding more writing skills | Browse, then install each skill from its own repo |
The marketingskills README also supports /plugin marketplace add coreyhaines31/marketingskills followed by /plugin install marketing-skills if you’d rather take the whole set.
For voice work in depth, see Claude Code skills for brand voice and Claude Code skills for editing.
Paid skill packs: what you get and whether any of them humanize
Here’s the honest finding: none of the paid packs below is a dedicated humanizer. They are broad bundles that include copywriting, content and editing skills. If humanizing is your only need, the free GitHub skills above cover it. Paid packs make sense when you also want a large library of marketing or SEO workflows in one purchase.
Prices were checked on the live pages on 1 November 2026, including the price metadata in the page source.
| Pack | Where | Price | What it covers | Notes |
|---|---|---|---|---|
| Master Marketing Skills Bundle (Inflectual) | Gumroad | $167 one-time | 167 Claude skills across email, SEO, paid ads, strategy, research, copywriting, content and social | Copywriting and content categories are the relevant ones for tone work. The page’s category breakdown doesn’t add up to 167 on its own, so ask the seller for the full list. |
| The Ultimate Claude SEO & AEO Skills Pack (Samuel Uebersax) | Contra | $39 (shown against $79) | SEO and answer-engine audits, planning, writing and optimization | Writing-focused rather than humanizer-focused. The $79 figure is the seller’s comparison price. |
| SuperSkills | superskills-one.vercel.app | $50 one-time | A curated bundle of Claude Code skills across many professional categories | Category counts differ between the vendor’s own pages (15 vs 20). |
None of these figures about skill counts or quality has been independently verified. They come from the sellers’ own pages. Paid packs we found using countdown timers or “only X left” claims were left out.
Free vs paid: the short comparison
| Question | Free GitHub skills | Paid packs |
|---|---|---|
| Dedicated humanizer? | Yes, several | No, humanizing is a side feature at best |
| Can you read the code first? | Yes, every SKILL.md is public | Usually only after purchase |
| Updates | Visible in commit history | Depends on the seller |
| Licence | MIT or Apache 2.0 for those listed | Seller’s terms |
| Cost | $0 | $39 to $167 for the packs listed |
| Best for | Fixing AI tells in drafts | Teams wanting a big ready-made library |
Why “bypass AI detectors” is the wrong goal
Many humanizer tools sell themselves as detector beaters. Building your workflow around that promise is a bad bet, for three reasons backed by research.
Detectors are unreliable in both directions
- OpenAI withdrew its own classifier. In January 2023 OpenAI reported that its AI text classifier correctly flagged 26% of AI-written text and wrongly flagged human text 9% of the time. The company shut it down in July 2023, citing its low accuracy (OpenAI announcement).
- Detectors penalize non-native writers. A 2023 Stanford study by Weixin Liang, James Zou and colleagues, published in Patterns, ran seven GPT detectors on TOEFL essays by non-native English speakers and essays by US eighth graders. The detectors misclassified more than 61% of the TOEFL essays as AI-generated on average, while rarely misjudging the native speakers’ essays (coverage by The Markup). The researchers tied this to low “perplexity,” meaning plainer word choices, which detectors read as machine-like.
- Academic testing found no tool reliable. Weber-Wulff and colleagues tested 14 detection tools in a 2023 study in the International Journal for Educational Integrity and concluded the tools were neither accurate nor reliable, with a bias toward labelling output as human.
Detection is an arms race you don’t need to enter
Researchers including Sadasivan et al. (2023, “Can AI-Generated Text be Reliably Detected?”) showed that simple paraphrasing can drop detector accuracy sharply. Detector vendors then retrain. Any tool promising to beat “every detector” is promising to win a moving race, and that claim is unconfirmed by any independent test we found.
Your reader is the real test
Search engines and customers don’t run your text through GPTZero. They read it. A post that slips past a detector and still says nothing loses the reader anyway. Aim the edit at the person, and detector scores stop mattering much.
Takeaway: use a humanizer to make writing clearer and more specific. If a tool’s main pitch is “undetectable,” treat that as a warning sign.
Google’s stance on AI content
Google doesn’t penalize content for being AI-assisted. It penalizes content that is unhelpful or made at scale to manipulate rankings.
In its February 2023 post Google Search’s guidance about AI-generated content, Google said it rewards “high-quality content, however it is produced,” and that “appropriate use of AI or automation is not against our guidelines.” The same post says using automation “with the primary purpose of manipulating ranking in search results” violates its spam policies.
In March 2024 Google broadened that into its scaled content abuse policy: producing many pages mainly to manipulate rankings is spam whether people, AI or both made them (Google Search spam policies).
What this means in practice:
| Situation | Risk under Google’s stated policy |
|---|---|
| AI draft, edited by a person who adds real knowledge | Low, if the result helps readers |
| AI draft run through a humanizer, nothing else changed | Same quality as before, so the same risk as before |
| Hundreds of humanized pages built from templates to target keywords | High, this is what scaled content abuse targets |
Google’s people-first guidance also asks whether content shows first-hand experience, the first “E” in E-E-A-T. A humanizer can’t add that. More in Claude Code skills for SEO.
How to add real experience (the part no skill can do)
Humanizing does not make thin content good. That sentence is the whole argument of this section.
A humanizer can turn “This strategy can significantly improve results” into “This strategy improves results.” It’s shorter and cleaner, and it’s still empty. What makes it worth reading is a fact only you have: what you tried, what broke, what the numbers were.
A simple workflow
- Draft with Claude, using a writing skill like copywriting or draft-content.
- Add your proof before you humanize. Drop in your own data, screenshots, client quotes (with permission), and named mistakes. Mark gaps as
[NEED: real example]instead of letting Claude fill them. - Run the humanizer. Ask it to flag tells and keep every fact unchanged.
- Run a detector script as a lint, not a judge. For example, a local AI-tell checker that counts em dashes and banned words.
- Read it aloud. Anything you wouldn’t say to a colleague gets cut.
- Check every claim against a source link before publishing.
What “experience” looks like on the page
- A specific before-and-after from your own work, with the date
- The tool or setting you tried that didn’t work, and why
- A screenshot of the real output, not a mock-up
- A named source for every number
- An opinion you’re willing to defend
If you can’t supply any of these, the piece probably isn’t ready, and no skill will fix that. Related reading: Claude Code skills for content writers and Claude Code skills for copywriters.
Common mistakes when humanizing with Claude
- Letting the skill invent specifics. If a rewrite adds a number, name or quote you didn’t supply, delete it. Choose skills that explicitly refuse to invent, as blader/humanizer’s README says it does.
- Running it on everything at full strength. Weak tells such as one em dash or one triplet don’t justify an edit. Over-editing produces a new flat style.
- Chasing a detector score. You’ll start writing for the tool instead of the reader.
- Humanizing before fact-checking. Fix the facts first. Cleaner prose makes false claims more persuasive.
- Installing unread skills. A SKILL.md can include scripts. Read it, especially from low-star repos.
- Stacking three humanizers. Pick one pattern list and one voice skill. Conflicting rules make Claude hedge.
Building your own? See how to build a Claude Code skill. For short formats, the same rules apply to LinkedIn posts and email marketing.
FAQ
What is the best Claude Code skill for humanizing AI writing?
By adoption, blader/humanizer (about 63,000 stars on 1 November 2026). It’s free, MIT licensed and based on Wikipedia’s AI-writing guide. “Best” depends on your needs: Aboudjem/humanizer-skill has a bigger pattern list, and alirezarezvani’s content-humanizer adds a scoring script.
Is there a free Claude humanizer?
Yes. Every dedicated humanizer skill listed here is free and open source.
How do I install the humanizer skill in Claude Code?
Run /plugin marketplace add blader/humanizer then /plugin install humanizer@humanizer (Claude Code 2.1.142 or newer), or npx skills add blader/humanizer --global --agent claude-code.
Can a humanizer make AI text undetectable?
No tool can promise that. Detectors disagree with each other and change over time, and the research cited above shows they are unreliable. Edit for readers instead.
Are AI detectors accurate?
Not reliably. OpenAI withdrew its own classifier after it caught 26% of AI text, and a Stanford study found detectors flagged more than 61% of non-native English essays as AI on average.
Does Google penalize AI-generated content?
Not for being AI-made. Google penalizes unhelpful content and content made at scale to manipulate rankings, however it was produced.
Will humanizing my content help it rank?
Only indirectly. Clearer writing can help readers stay and trust you. Rankings still depend on usefulness, expertise and links.
What are the most common signs of AI writing?
Puffery (calling things a “testament”), promotional words, -ing analysis tails, vague attribution, negative parallelism, rule of three, em dash overuse and chat leftovers like “I hope this helps.”
Is the Wikipedia “Signs of AI writing” page reliable?
It’s a practical guide written by Wikipedia editors, not a scientific study. It says itself that the signs are hints, not proof.
Does Claude write less like AI than ChatGPT?
That’s a common Reddit opinion, but it’s unconfirmed. We found no controlled study comparing them. Both produce recognizable tells without editing.
Can I use a humanizer on my own human writing?
Yes. Human drafts often contain the same filler. The edit works the same way.
Do humanizer skills work in Cursor or Codex too?
Many do. blader/humanizer and Aboudjem/humanizer-skill both document Skills CLI installs for other agents.
Should I pay for a humanizer skill pack?
For humanizing alone, the free options are enough. Paid packs ($39 to $167 among those listed) suit people who want a broad marketing or SEO library.
Is it ethical to humanize AI text?
Editing for clarity is ordinary editing. Using it to pass off AI work where disclosure is required, such as in school or some publications, breaks those rules regardless of the tool.
Can a humanizer add personal experience?
No. It can only rearrange what you give it. Add your own stories, data and opinions yourself.
Where else can I find humanizer skills?
Directories such as claudemarketplaces.com, cultofclaude.com, skillsmp and claudskills list many. Check the underlying GitHub repo before installing.
Sources and methodology
GitHub star counts, licences and last-push dates come from the GitHub API on 1 November 2026. Install commands are copied from each repo’s README. Paid pack prices were checked on the live product pages the same day, including price metadata in the page source.
- Wikipedia: Signs of AI writing
- Google Search Central: Google Search’s guidance about AI-generated content (February 2023)
- Google: Spam policies for Google web search
- OpenAI: New AI classifier for indicating AI-written text (updated July 2023)
- Liang et al., “GPT detectors are biased against non-native English writers,” Patterns, 2023; The Markup coverage
- Weber-Wulff et al., “Testing of detection tools for AI-generated text,” International Journal for Educational Integrity, 2023 (DOI)
- Sadasivan et al., “Can AI-Generated Text be Reliably Detected?”, arXiv, 2023 (arXiv:2303.11156)
This post contains no affiliate links and no hands-on test results. Descriptions of each skill come from its own documentation.
Your next step
Install one free humanizer, run it on your last published post, and look at what it flags. Then do the part it can’t: add one real example, one real number and one opinion to every section. For the wider toolkit, see Claude Code skills for marketers.
