Humanizer Skill for Claude: Make AI Text Sound Human
Install the humanizer skill in Claude Code with two commands, then pick a DeepSeek Harness (dsh) humanizer plugin for English, Chinese or Russian text.
Last updated: 2026-11-15
Quick answer
A "humanizer skill" is an agent skill — a folder holding a SKILL.md — that rewrites text so it stops reading like a default model answer: no "It's not just X, it's Y", no "Let that sink in" closers, no bold used as decoration. The one most of this search lands on is blader/humanizer (MIT), built on Wikipedia's Signs of AI writing: it installs in Claude Code with two slash commands and works through 26 patterns, strongest first.
Two caveats are worth taking from the project's own README before you install anything. It edits "for human readers", and — in the README's words — "does not try to get past AI detectors, and detectors still flag most of its output". Its headline quality claim, a 16-out-of-16 blind preference in its own issue thread, is the project's own test, not an independent benchmark.
Everything on this page was checked against the repositories and their READMEs on 2026-11-15; commands are copied verbatim.
What the skill actually changes
Humanizer names 26 "tells" and sorts them from strongest to weakest. It edits any of the first five the first time it sees one:
| # | Pattern | What it looks like |
|---|---|---|
| 1 | Not X but Y | "It's not just a feature, it's a shift." |
| 2 | One-line closers | "Let that sink in." after every section |
| 3 | Sayings that sound deep | "At its core, what really matters is..." |
| 4 | Staged run-up | "Here's the thing." / "Honestly?" |
| 5 | Arguing with no one | "I'm not saying X, but..." |
The rest fall into five more groups: rhythm by rule (forced triads, repeated sentence openings, dashes as the universal connector), inflation and borrowed authority ("marking a pivotal moment", "experts believe"), formatting by rule (bold as decoration, decorative headings, curly quotes), leftovers from the chat ("Great question!", "I hope this helps!"), and writing for the wrong reader.
The workflow is the part people miss: it marks every tell it finds, drafts a rewrite, critiques the draft for anything still artificial, then checks the result against the original claims. Pointed at a file, it changes only the prose and leaves code, data, frontmatter and link targets alone.
Install it in Claude Code
From inside Claude Code, add the marketplace and install the plugin:
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer
The README states the plugin needs Claude Code 2.1.142 or newer. On older builds it points at the Skills CLI instead:
npx skills add blader/humanizer --global --agent claude-code
Then call it with /humanizer and paste the text — or just ask in plain language ("humanize this text: ..."). To match your own voice, include 2–3 paragraphs of your writing with the request; the README says it then follows the sample's rhythm, word choice and punctuation.
The same skill in other agents
| Agent | Command (README verbatim) |
|---|---|
| Codex | npx skills add blader/humanizer --global --agent codex |
| pi | pi install git:github.com/blader/humanizer |
| Every agent the Skills CLI knows | npx skills add blader/humanizer --global --agent '*' |
| Claude.ai / Claude Desktop | Download the repo ZIP (Code → Download ZIP) and upload it as a skill in Settings |
Drop --global to install it for one project only. For an agent the Skills CLI does not recognise, the README's fallback is to copy SKILL.md into that agent's skill folder.
Humanizer options, compared
Only repositories verified through the GitHub API on 2026-11-15 are listed. "Language" is the language each project's rules and README target, not a claim that it refuses other languages.
| Project | Language | License | Install (README verbatim) | Link |
|---|---|---|---|---|
| blader/humanizer | English (rules quote English patterns) | MIT | /plugin marketplace add blader/humanizer then /plugin install humanizer@humanizer | GitHub |
| Vladimir-Human/humanizer-ru | Russian only (SKILL.md: "Только для русского языка") | MIT | npx skills add Vladimir-Human/humanizer-ru | GitHub · directory page |
| lynote-ai/dsh-humanizer | English + Chinese rule catalogue | BSD-3-Clause | dsh plugin add github:lynote-ai/dsh-humanizer | GitHub |
| DEEP-IOS/dsh-humanizer | Chinese (long-form fiction and articles) | MIT | dsh plugin --profile web add dsh-humanizer@next | GitHub |
| hellogit2021/avoid-ai-writing-cn | Chinese | MIT | dsh plugin --profile web add github:hellogit2021/avoid-ai-writing-cn | GitHub |
| whh110112/human-writing-skills | Multilingual editing, long-form | MIT | dsh plugin --profile web add https://github.com/whh110112/human-writing-skills/releases/download/dsh-v0.3.1/dsh-advanced-human-writing.tgz | GitHub |
What separates them in practice:
blader/humanizeris the packaged Claude Code experience: a/humanizerslash command, a plugin manifest, a committed Cursor rule. Best starting point if you work in Claude Code.Vladimir-Human/humanizer-ruis the only Russian specialist here. Beyond the rewrite rules it shipsscripts/check_facts.py(re-compares numbers, dates and names between original and edit) andscripts/lint_style.py(scores surviving tells), and its README says the scripts are part of the workflow, not optional.lynote-ai/dsh-humanizermakes no model calls at all: 8 deterministic tools that emit scores, rule lists and rewrite briefs, and the agent does the rewriting. It also learns a per-author "voice profile" (1–100 similarity score) stored in~/.dsh/voice-profiles.json.DEEP-IOS/dsh-humanizertargets Chinese long-form work (novels, articles) and adds a SQLite-backed memory workbench so progress, character relations and foreshadowing survive compaction. Its README is explicit that the alpha line has no blind-evaluation proof yet.hellogit2021/avoid-ai-writing-cnis the lightest Chinese option: three spoken triggers — "去掉AI味" rewrites, "只标记" audits without editing, "改文件<路径>" edits in place — plus "写的不错" which appends the AI-isms it just saw to a personal word list under%APPDATA%\avoid-ai-writing-cn\user-patterns.md.whh110112/human-writing-skillsis a writing skill family rather than a pure de-AI pass: multilingual editing, chunked long-form collaboration and continuity review, shipped both as skills and as a dsh plugin tarball.
Using them with DeepSeek Harness
A dsh skill is the same object as a Claude Code skill — a folder with a SKILL.md. Project skills live at .dsh/skills/<name>/SKILL.md inside your workspace, and dsh re-injects the catalogue on the next prompt, so no restart is needed; the install-skills walkthrough shows the seven-step version with screenshots, and skill installation, hands-on covers the global-vs-project choice.
Three of the table entries are dsh plugins, which land in a profile and need a restart:
dsh plugin --profile web add github:lynote-ai/dsh-humanizer
dsh plugin --profile web add dsh-humanizer@next
dsh plugin --profile web add github:hellogit2021/avoid-ai-writing-cn
dsh plugin add forwards to pnpm inside ~/.dsh/profiles/<profile>/, which is why the source spec has to be something pnpm understands — see Installing plugins you can trust for the full command shape and a vetting checklist before your first third-party install.
For anything Claude Code already has, the cheaper route is a copy: drop the skill folder into .dsh/skills/ and the agent picks it up on the next message.
When a humanizer is the wrong tool
- Beating a detector. None of these projects claim it. The blader README says detectors still flag most of its output; use the skill to make text readable, and do not present rewritten text as unassisted human writing where that matters.
- Preserving facts automatically. Only
humanizer-ruships a fact-comparison script. Elsewhere, the read-back of numbers, dates and names is on you — check the diff. - Non-English text with an English rule catalogue. Chinese and Russian need their own rule sets; the projects above that target those languages say so explicitly.
- Policy-sensitive writing. Coursework, submissions and anything with a disclosure requirement have their own rules; a rewrite does not change what you are allowed to submit.
FAQ
What is the humanizer skill for Claude Code?
An agent skill that rewrites text so it stops sounding machine-written. The reference implementation is blader/humanizer (MIT): one SKILL.md plus a Claude Code plugin manifest, built on Wikipedia's "Signs of AI writing" and covering 26 patterns. You invoke it with /humanizer and paste the text.
Does a humanizer skill get past AI detectors?
No. The blader/humanizer README says outright that it "does not try to get past AI detectors, and detectors still flag most of its output". The skill edits for human readers — natural rhythm, specific facts, no filler — which is a different goal from beating a classifier.
How do I install it?
In Claude Code: /plugin marketplace add blader/humanizer, then /plugin install humanizer@humanizer. On older Claude Code builds the README points at npx skills add blader/humanizer --global --agent claude-code. Codex uses npx skills add blader/humanizer --global --agent codex, and --agent '*' installs for every agent the Skills CLI supports.
Is there a humanizer for Chinese or Russian text?
Yes. For DeepSeek Harness there is DEEP-IOS/dsh-humanizer and hellogit2021/avoid-ai-writing-cn (both Chinese, MIT), plus lynote-ai/dsh-humanizer (English + Chinese rules, BSD-3-Clause). For Claude Code and other skill-aware agents, Vladimir-Human/humanizer-ru is a Russian-only skill (MIT) with its own fact-checking and style-lint scripts.
Can I use these in DeepSeek Harness instead of Claude Code?
Yes. A skill is a folder with a SKILL.md, and dsh reads project skills from .dsh/skills/<name>/SKILL.md — so a plain SKILL.md copied from a Claude Code skill works. Several of them also ship as dsh plugins, installed with dsh plugin --profile web add <source>.
Sources
Every claim above comes from these pages, read on 2026-11-15:
- github.com/blader/humanizer — README: Wikipedia-derived pattern list, the 26 patterns, install commands per agent, the Claude Code 2.1.142 requirement, the "does not try to get past AI detectors" statement, the 16/16 blind-test claim (issue #229), MIT license. GitHub API snapshot taken 2026-11-15: MIT, Python, created 2026-01-18, pushed 2026-11-10, about 69k stars (a 2026-11-15 snapshot, not a fixed number).
- github.com/Vladimir-Human/humanizer-ru — README and
SKILL.md:npx skills addinstall, Claude Code marketplace pair, Gemini CLI extension, manual copy paths,check_facts.py/lint_style.py, Russian-only scope. - github.com/lynote-ai/dsh-humanizer — README:
dsh plugin add github:...install line, 8 deterministic tools, no model calls,~/.dsh/voice-profiles.json, BSD-3-Clause. - github.com/DEEP-IOS/dsh-humanizer — README: npm
latest/nextdist-tags,dsh plugin --profile web add dsh-humanizer@next, tool list, SQLite memory workbench, the alpha-line honesty note, MIT. - github.com/hellogit2021/avoid-ai-writing-cn — README: install line, the four spoken triggers, the personal word-list path, peer range
>=0.1.0-rc.6 <0.3.0, MIT. - github.com/whh110112/human-writing-skills — README: release-tarball install command, skills-plus-plugin layout, multilingual editing and continuity review, MIT.
The third-party plugins and skills above are community projects, not DeepSeek or Anthropic products, and none of them is affiliated with this site. Install commands come from each README at the date shown; re-check the README before you run one, because the source spec (branch, tag, path) changes with releases.
Related reading: the Impeccable skill, install skills the right way, the best DeepSeek Harness skills to install, skill installation, hands-on, Jane Street agent skills.
