Claude Humanizer skill review: verdict, safety and alternatives

Humanizer, by blader, rewrites AI-sounding text against 26 patterns from Wikipedia's Signs of AI writing. We rate it Recommended with caveats, 72/100, because three of its own examples change facts.

Get the skill: humanizer repository on GitHub · SKILL.md at the reviewed commit · skills.sh listing

Does Claude have a humanizer skill?

Yes, Claude does have a humanizer skill: blader's Humanizer, with 58,839 GitHub stars on 2026-10-21. It rewrites AI-sounding text against 26 patterns from Wikipedia's Signs of AI writing. We rate it Recommended with caveats, 72/100.

Humanizer is a Claude skill published by blader on GitHub. The documentation for Hermes Agent, NousResearch's agent, credits the author as Siqi Chen (@blader). The skill is one long prompt file, SKILL.md, that the agent loads on request. It teaches the model to spot AI-writing patterns and rewrite around them. The pattern list comes from Wikipedia's Signs of AI writing page, kept by WikiProject AI Cleanup.

Outside tests back the core claim. Judges preferred its rewrites in all 16 blind comparisons (#229). A GPT-5.4 replay on 18 drafts cut a slop score by more than half (#250). Its own examples are the weak spot.

What is a humanizer skill?

A humanizer skill is an Agent Skill that rewrites AI-generated prose so it reads as human writing. It removes tells such as inflated vocabulary and rule-of-three lists. A good one keeps every fact.

An Agent Skill is a folder with a SKILL.md file that an agent loads when a task matches it. Read Agent Skills defined for the full picture. Humanizer groups its 26 patterns into six families at v3.1.0:

• Staging instead of stating: "not X but Y" lines, one-line closers, staged run-ups.
• Rhythm by rule: forced triads, repeated openings, dashes as the universal connector.
• Inflation and borrowed authority: overused AI words, inflated significance, sales language.
• Formatting by rule: decorative bold, decorative headings, curly quotation marks.
• Leftovers from the chat and the draft: chatbot residue, knowledge-limit disclaimers.
• Writing for the wrong reader: re-explaining what the reader already knows.

Is the Claude Humanizer skill worth installing?

Yes, it is worth installing with a fact check. The Claude Humanizer skill scores 72/100, Recommended with caveats, at Moderate confidence. Blind judges preferred its rewrites 16 of 16 times, but three worked examples change facts.

Tell removal is clearly positive, and two outside parties measured it. That alone supports a high-70s score. The worst failure is altered facts in published text. That failure still has a live source in three few-shot examples the maintainer declined to fix. Fact preservation pulls the overall score down to 72.

The page on how verdicts are scored explains the scoring.

Fabrication appears in three worked examples, not in reported 3.x runs, so it lowers fact preservation. Effectiveness stays at 80 because both third-party measurements are small and LLM-judged.

One piece of evidence is missing. Nobody has published a live run on 3.1.0 across several models that counts added or changed facts. Inputs that resemble the §12, §18 and §25 examples matter most for that run.

Strengths and weaknesses in short:

• Measured gains: 16 of 16 blind wins and a slop score cut from 76.3 to 35.4.
• Conservative edits: one comparison counted 14 changed passages and an untouched structure.
• Fast fixes: the Desktop loader and ZIP symlink bugs closed within weeks.
• Fact-changing examples: §12, §18 and §25 add details or change a number.
• Heavy context: about 8,087 tokens load for every call, even one paragraph.
• Silent writes: file mode overwrites the named file without asking.

Humanizer's place among the other writing verdicts is on our list of writing skills for Claude, ranked.

How many tokens does the Humanizer skill use?

The Humanizer skill file measures about 8,087 tokens per load, even for one paragraph. Our estimate for an 800-word blog draft is 9,500 input and 2,600 output tokens. Output runs two to three times the input length.

The default output returns a draft, a list of remaining patterns and a final version. That structure is why output runs two to three times the text length (estimate). Tessl's quality review suggests moving the pattern catalog into a reference file. No release has made that change yet.

Which Claude skill is best for human writing?

For human writing, blader's Humanizer is the best Claude skill we have reviewed, at 72/100. Beautiful Prose scores 62. Width.ai, Hermes Agent and stackademic humanizers have no verdict yet.

The criterion: a skill must remove AI tells while keeping every fact. Pattern counts are as of 2026-10-21.

Four more anti-slop skills have no verdict here yet: avoid-ai-writing, ZeroSlop, agent-style and stop-slop. The Hermes Agent port credits blader under the MIT license. It shares the pattern idea at an older version.

Humanizer cleans a draft after it exists. Drafting is a separate job. The Content Research Writer review covers a co-writer skill for outlines and cited research. That skill scores 58, Conditional. The ranking of all writing skills covers editing skills beyond this group.

Is the Humanizer skill safe to install?

Yes, the Humanizer skill is safe to install. It carries safety grade B: one prompt file plus a local validator, with no network calls or credentials. File mode overwrites the named file without confirmation.

The grade comes from a static read of the files at commit 225a6f3, with these findings:

• One script ships: scripts/validate-package.py checks manifest and version consistency for maintainers. The skill never calls it.
• Network indicators are only github.com and json.schemastore.org URLs inside plugin manifests.
• No environment variables, MCP configs or hooks exist in the package.
• Manifests cover the Claude Code plugin and marketplace, a Cursor plugin and Codex (agents/openai.yaml).
• Input text is treated as material to edit, never as instructions. Release 3.0.0 added that rule after #238.
• File mode writes the final text straight to the named file. No confirmation step exists, and #238 left this unchanged.
• skills.sh audits read Gen Agent Trust Hub Pass, Socket Pass and Snyk Warn.

The file-mode overwrite is the reason for B over A. Commit your file to version control before you point Humanizer at it. Our guide to vetting a skill before install lists the same checks for any skill.

What does the community report about the Humanizer skill?

The community reports better rewrites from the Humanizer skill and no detector wins. A GPT-5.4 replay cut a slop score from 76.3 to 35.4 (#250). Pangram still flags its output.

These are the 20 dated reports we weighed, from GitHub, Hacker News and the open web.

What are the Humanizer skill's known failures?

The Humanizer skill has two known failures at v3.1.0. Its §12, §18 and §25 examples add details or change a number, and models copy examples. It over-corrects text that is already human.

The worst failure for this job is a rewrite that silently invents or alters a fact. The user then publishes that text under their own name. Four open problems remain at v3.1.0:

1. Fact-changing examples. The §12, §18 and §25 before/after pairs add details or change a number. Issue #306 closed as not planned, and PR #233 was declined.
2. Over-correction. Human, terse or journal-style text comes back flattened. The density pre-check request (#93) was declined.
3. Silent overwrite. File mode writes the final text to the named file with no confirmation (#238).
4. Detector flags. Pangram and other checkers still flag the output (#263, #281). The README calls detector evasion a non-goal.

Three older failures are fixed. The Claude Desktop loader bug (#202) closed in v2.10.2. The ZIP symlink rejection (#224) closed in v2.11.1. Shape edits that dropped ranking claims (#212) gained a check in 3.0.0.

What does a fact-changing Humanizer example look like?

A fact-changing Humanizer example turns "over 3,000 square feet" into exactly "3,000 square feet" in §18. Another example in §25 invents a count of vendors. The maintainer declined both fixes.

The §18 pair in SKILL.md v3.1.0 reads as follows. Before:

"Gallery 825 serves as LAAA's exhibition space for contemporary art. The gallery features four separate spaces and boasts over 3,000 square feet."

After:

"Gallery 825 is LAAA's exhibition space for contemporary art. The gallery has four rooms totaling 3,000 square feet."

The §25 after-text adds a detail that the before-text never states:

"Prices are each vendor's published rate. Two vendors publish nothing; call them."

• "Over 3,000" is a floor; "3,000" is an exact figure. The rewrite changes the claim.
• "Two vendors publish nothing" is a new count. A model that copies this pattern invents facts in your text.

The README says "It does not make things up." That line is a publisher claim. The examples above, flagged in #306, show the opposite at v3.1.0.

Does the Humanizer skill work in Codex, Cursor and other agents?

Yes, the Humanizer skill works in Claude Code, Claude Cowork and Claude Desktop, where we verified it. Codex, Cursor, Gemini CLI, GitHub Copilot and Windsurf are publisher claims for other agents. No Codex or Cursor run is documented.

Verified here means public evidence shows the skill loading and working in that agent. Claimed means the publisher lists the agent or ships a manifest for it.

README commit c163acb on 2026-09-28 dropped Cursor and OpenCode from the README agent list. The Cursor manifest remains in the repository. Claude Code is Anthropic's terminal agent. Our list of the best Claude Code skills covers more skills for it. The guide to skills that work in the Claude app covers Desktop and Cowork.

How do you add the Humanizer skill to Claude?

Add the Humanizer skill to Claude Code with /plugin marketplace add blader/humanizer, then /plugin install humanizer@humanizer. In the Claude app, upload the repository ZIP as a skill in Settings. Codex uses one npx skills add command.

The commands match the README as read on 2026-10-21.

How do you add it in Claude Code?

Run the two plugin commands inside Claude Code 2.1.142 or newer:

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

Older Claude Code versions use the Skills CLI instead:

npx skills add blader/humanizer --global --agent claude-code

The plugin answers to /humanizer:humanizer once installed. Our guide to install skills in Claude Code covers plugin and folder installs in more depth.

How do you add it in the Claude app?

1. Open the repository on GitHub and choose Code, then Download ZIP.
2. Open Settings in Claude Desktop or Claude.ai.
3. Upload the ZIP as a skill.

Release v2.11.1 added a humanizer-skill.zip asset after a symlink broke uploads (#224). The steps to upload a skill to the Claude app cover the Settings path in detail.

How do you add it in Codex and other agents?

Codex installs through the Skills CLI with one command:

npx skills add blader/humanizer --global --agent codex

Replace the agent name with '*' to install for every supported agent at once.

How do you humanize content with the Humanizer skill?

Humanize content by typing /humanizer and pasting your draft, or by asking Claude to humanize it. The Humanizer skill returns a draft, remaining patterns and a final version. Diff every number before publishing.

1. Type /humanizer in Claude, or /humanizer:humanizer for the plugin install.
2. Paste the draft, or name the file you want rewritten.
3. Read the list of patterns the skill still finds.
4. Take the final version from the bottom of the reply.
5. Diff every number, name and date against your original.

A plain request works too. The README example prompt is "Please humanize this text" followed by the draft. Established writers can paste a writing sample first. The SKILL.md Voice section then matches the rewrite to that register.

File mode writes straight to disk. Paste text instead when the file has no version control.

Test it yourself with these prompt lines:

• Paste the §18 Gallery 825 before-text and ask for a humanized version.
• Paste three of your own fact-dense drafts, one at a time.
• Diff every number and named detail in each output.
• Count how many facts changed, appeared or disappeared.

Where is the Humanizer skill on GitHub?

The Humanizer skill is on GitHub at blader/humanizer, with SKILL.md at the repository root. We reviewed release v3.1.0, commit 225a6f3. The repository has 58,839 stars and has shipped no release since our review.

• Repository: github.com/blader/humanizer, MIT license.
• SKILL.md: at the repository root, one file of 32,348 bytes (at the reviewed commit).
• Reviewed release: v3.1.0 on 2026-09-28, full commit 225a6f39ac85f76ee48dbad772ea4abe4ed6c9d8.
• Contributors: 22, with six releases in the 90 days before review.
• Changes since review: four README edits on 2026-10-16, with SKILL.md unchanged.
• Copies: we found 2,712 exact copies of the skill.

Directory pages serve older copies. oneaway.io shows a v2.1.1 copy with 24 patterns. Install from the blader repository to get the reviewed version.

Common questions

Will Humanizer help academic writing pass AI detectors?

No. The README calls detectors a non-goal, and Pangram flags its output (#263, #281).

Does the Humanizer skill work in ChatGPT?

Unknown. ChatGPT is not in its agent list, and no public evidence shows it working there.

Is Humanizer an MCP connector?

No. It is a skill, one prompt file the agent loads; see skills versus MCP connectors.

Is there a Humanizer plugin for Claude Code?

Yes. Its own marketplace installs it, and the plugin answers to /humanizer:humanizer.

Why do pattern counts differ between sites?

Versions differ: v3.1.0 has 26 patterns, and the oneaway copy (v2.1.1) shows 24. The Hermes port (v2.5.1) shows 34.

Is there a Chinese version of Humanizer?

Yes. A separate humanizer-zh project exists, and it has no verdict on this site.