The blader/humanizer repository trended on GitHub in week 45 of 2026 and shows about 66.2k stars on its GitHub page as of November 8, 2026. It is not a model checkpoint. It is an agent skill that rewrites prose using Wikipedia’s Signs of AI writing — the same pattern list WikiProject AI Cleanup maintains for encyclopedia editors.
If you ship docs, changelogs, or blog drafts through Claude Code or Codex, Humanizer is a concrete answer to a question explainx.ai keeps hearing: how do I keep agent output from sounding like default ChatGPT cadence without lying to readers?
TL;DR
| Question | Answer |
|---|---|
| License | MIT (README) |
| Primary source for patterns | Wikipedia: Signs of AI writing |
| Pattern count | 26 in README tables |
| Claude Code trigger | /humanizer:humanizer (plugin path in README) |
| Codex install | npx skills add blader/humanizer --global --agent codex |
| Detector evasion? | Explicitly not a goal (README) |
| Blind test claim | README cites 16/16 judge preference in issue #229 — community report, not a peer-reviewed benchmark |
Why GitHub is amplifying it now
Trending skills usually ride a felt pain. explainx.ai’s Kevin Boone piece on trusting AI web content documents how readers downgrade pages that read like uniform model voice. Humanizer attacks style tells, not factual hallucination — complementary to harness false-alarm research on audits.
The repo README says it is sponsored by cfo.ai. Sponsorship does not change the skill mechanics: pattern list, rewrite pass, critique pass, final pass — all described in open SKILL.md.
Compared with benjitaylor/skills Stageshow or claude-animated-shorts, Humanizer is text-only and editorial. It fits next to jakubkrehel skills for UI in a growing skills marketplace pattern explainx.ai tracks in what are agent skills.
The before/after the README proves
The README’s launch example is a shared drafts feature announcement. Before text stacks Not X but Y, emoji rockets, “Let that sink in.”, and inflated significance. After text states what shipped, what it cost in files, and pricing in plain declarative sentences.
That example is marketing prose, but the same patterns show up in internal RFCs, support macros, and agent-generated PR descriptions. Humanizer is aimed at those defaults.
The five patterns it edits on sight
Rounded page with a pen and green spark, matching prose cleanup in agent workflows
The README ranks patterns by strength. The top five:
- Not X but Y — “It is not just a feature, it is a shift.”
- One-line closers — dramatic fragments after every section.
- Sayings that sound deep — faux philosophy instead of claims.
- Staged run-ups — “Here is the thing” before the fact.
- Arguing with nobody — straw objections nobody raised.
Patterns marked weak alone need clusters before a rewrite — the README warns careful human writers may use any one deliberately.
Full pattern map (high level)
The README groups all 26 patterns into six families:
| Family | Examples |
|---|---|
| A. Staging instead of stating | Patterns 1–5 above |
| B. Rhythm by rule | Forced triads, repeated openings, dash spam |
| C. Inflation and borrowed authority | “Delve,” “testament,” vague associations, fake expert consensus |
| D. Formatting by rule | Bold decoration, emoji headings |
| E. Chat and draft leftovers | “Great question!” wrappers, knowledge-limit disclaimers |
| F. Wrong reader | Re-explaining what the audience already knows |
Each row in the README includes before and after micro-examples — useful when you tune your own style guides for agents.
How the skill runs (workflow)
From SKILL.md and the README:
- Mark tells in the incoming text, strongest first.
- Draft a rewrite without treating the original outline as sacred.
- Critique the draft for remaining tells and factual drift.
- Ship final prose; for files, change prose only and leave code, data, frontmatter, and link targets alone.
When you paste text interactively, the skill shows first draft, critique, final — transparency that helps humans learn patterns, not only outsource editing.
Voice matching: paste 2–3 paragraphs of your own writing and Humanizer will mimic rhythm and punctuation quirks, including dashes if your sample uses them.
Installation paths (copy from README)
Claude Code (plugin):
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer
Requires Claude Code 2.1.142+ per README; older clients fall back to Skills CLI.
Codex and multi-agent:
npx skills add blader/humanizer --global --agent codex
Every agent the CLI knows:
npx skills add blader/humanizer --global --agent '*'
Claude.ai / Desktop: download ZIP and upload as a skill in settings — README path for non-CLI users.
For install hygiene, mirror checks from ECC affaan-m harness skills: read SKILL.md, verify licenses, and know what network calls your agent may still make outside the skill.
What developers are saying on Hacker News
An older HN thread on Humanizer (item 47109489) is mostly normative, not benchmark-heavy. Commenter davydm called it “another step in the war” they would not champion — a values objection to polishing AI prose. wps lamented that automated polish drains meaning from writing. Those reactions are not hands-on throughput numbers; they warn that humanizer culture can collide with authenticity goals. Use the skill when clarity is the goal, not when voice is the product.
Limits you should plan for
The README is explicit:
- Detectors may still flag output — do not use it for academic integrity evasion.
- Facts must not be invented — it should ask when a rewrite needs missing data.
- Technical reference stays neutral; personal essays keep quirks.
Humanizer does not replace citation checking-style discipline or legal review for regulated copy. Keep a human editor on the hook for numbers that appear only once in a draft — pattern stripping does not catch invented statistics.
Pattern deep dive — inflation and borrowed authority
Patterns 12–18 in the README are where enterprise blogs fail reviews. The README calls out overused tokens such as “delve,” “testament,” and “landscape,” plus inflated significance lines like “marking a pivotal moment.” Enterprise risk teams increasingly grep for those words after compliance incidents.
Pattern 17 — borrowed authority is especially toxic in B2B copy: “Experts believe” without naming an expert, or a logo wall of outlets with no linked study. Humanizer’s fix is blunt: name a real source or delete the claim. That aligns with explainx.ai’s guidance in how to read model cards — citations must point somewhere inspectable.
Pattern 18 swaps “serves as / features / boasts” for plain “is / has.” It sounds trivial; it removes an entire class of press-release passive voice agents love.
Pattern deep dive — formatting and chat leftovers
Patterns 19–25 catch bold decoration, emoji headings, and chatbot residue such as “Great question!” wrappers. Those lines often survive multi-agent pipelines when one model drafts and another formats.
Pattern 25 — writing about the document shows up in agent PRs: “This function was added to replace…” Humanizer pushes the model to describe behavior, not diff meta.
Using Humanizer inside a daily agent loop
A practical loop for doc teams:
- Draft with your coding agent under a normal system prompt.
- Run
/humanizeron paragraphs that face customers or regulators — not on every code comment. - Diff the prose-only changes; reject any fact edits you did not authorize.
- Run your existing style linter and legal review — Humanizer is not compliance.
If you orchestrate with loop engineering patterns, add Humanizer as an optional pass after factual QA, not before technical QA. Order matters: humanizing hallucinations just produces polite hallucinations.
Comparison — what Humanizer does not replace
| Tool or habit | Humanizer overlap |
|---|---|
| Manual copy edit | Overlaps — Humanizer automates pattern removal |
| AI verification tax workflows | Complements — still need fact checks |
| Plagiarism or rights scanners | No overlap |
| Claude Security scans | No overlap — security not voice |
Custom SKILL.md voice guides | Can combine — feed your samples for voice match |
Wikipedia maintenance context
Pen tracing a wavy line from stacked letters, echoing Wikipedia’s Signs of AI writing editorial guide
WikiProject AI Cleanup does not maintain Humanizer, but both lean on the same Signs of AI writing page. Editors remove slop from articles; Humanizer removes slop from drafts before you paste into CMS tools.
That shared origin is a feature: when your comms team asks “why does this read like ChatGPT,” you can point to a public pattern list instead of taste arguments. It also means when Wikipedia updates the guide, skill authors may need to merge upstream pattern changes — watch the repo CHANGELOG.
Sponsorship and supply chain
The README discloses cfo.ai sponsorship. Sponsorship does not change the MIT license or the Wikipedia attribution. Still follow normal supply-chain hygiene from ECC skills install guidance: read SKILL.md, check plugin marketplace signatures, and pin versions in team docs.
Ethics for teams
If you edit customer-facing text:
- Disclose AI assistance where your policy requires it — rewriting does not remove machine origin.
- Keep human review on numbers, security claims, and comparisons.
- Do not use Humanizer to launder plagiarized or scraped content — pattern stripping is not rights clearance.
Related reading
- What are agent skills — complete guide
- ECC affaan-m — skills install risks
- Kevin Boone on AI web content and trust
- benjitaylor Stageshow UI skill
- Claude animated shorts Remotion skill
- Andy Hall on audit agents and false alarms
- Loop engineering with coding agents
Star counts and install commands match the blader/humanizer README as of November 8, 2026.