Humanizer: A Claude Code Skill That Actually Fixes AI Writing (And Knows What It's Fixing)
The Problem Nobody Wants to Admit
We're all using LLMs to write things. Emails, docs, blog posts, PR descriptions. And we're all getting output that reads like it was generated by a model trained on corporate LinkedIn posts from 2023. You know the signs: "testament to", "rapidly evolving landscape", the em dash every third clause, bullet points that start with 💡 and 🚀. The stuff that makes your readers quietly lose respect for you.
Humanizer is a Claude Code skill sitting at nearly 14,000 stars that attempts to solve this in a structured, repeatable way. It's worth looking at — not because it's magic, but because it's the most honest attempt I've seen at naming and fixing the actual problem.
What It Actually Does
This is not a wrapper around "rewrite this to sound more human." That prompt is useless because the model just produces a different flavor of AI slop.
Humanizer is a SKILL.md file you drop into your Claude Code skills directory. When invoked, it runs your text through a checklist of 29 specific, named patterns derived from Wikipedia's "Signs of AI writing" guide — a document maintained by people who have been manually cleaning AI-generated content at scale. That sourcing matters. This isn't someone's opinion about what sounds bad; it's a community-maintained taxonomy built from thousands of real examples.
The skill does two passes: a first rewrite applying the pattern rules, then a second "obviously AI generated" audit pass that catches things the first pass missed. That double-pass approach is a small but meaningful design decision — it acknowledges that a single rewrite often just shuffles the problems around.
There's also a voice calibration feature added in March 2026. You can paste in 2-3 paragraphs of your own writing and the skill will analyze your rhythm and word choices before rewriting. This is the difference between "generic clean" and "sounds like me." It's the feature I'd actually use day-to-day.
Why This Has 14k Stars
The timing is obvious. AI writing is everywhere, AI-detection anxiety is real, and the quality bar for written communication is quietly dropping across the industry. People are looking for something more systematic than "just prompt it differently."
But I think the deeper reason this resonated is that the pattern list is specific. Most advice about AI writing is vague: "add personality," "vary sentence length," "don't sound robotic." Humanizer gives you named, actionable items:
- Significance inflation: "marking a pivotal moment in the evolution of" → just say what actually happened
- Rule of three: LLMs compulsively produce three-item lists; break the pattern when you only have two real things to say
- Synonym cycling: the model calling the same person "protagonist," then "main character," then "central figure" because it's avoiding repetition that wouldn't actually bother a human reader
- Passive voice / subjectless fragments: "No configuration needed" — who doesn't need to configure it? Name the actor.
- Chatbot artifacts: "I hope this helps! Let me know if you have questions" — yes, these end up in drafts people actually send
That specificity is what makes this useful as a skill rather than a prompt. The model has concrete rules to apply, not a vague aesthetic goal.
Key Features Worth Calling Out
The pattern taxonomy is the real product. The 29-pattern list in WARP.md is genuinely useful reading independent of the tool. I'd recommend reading through it even if you never install the skill. It'll change how you review AI-generated text.
Voice calibration is underrated. The March 2026 addition by Matt Van Horn lets you provide writing samples for style matching. This is the feature that separates "AI text that sounds human" from "AI text that sounds like you." It's optional and the README explains it clearly.
The double-pass audit is honest engineering. Rather than pretending one rewrite is sufficient, the skill explicitly does a second pass to catch AI-isms that survived the first round. This is the kind of pragmatic design choice that suggests the author actually tested this on real text.
Active community iteration. The recent commits show real engagement: someone added a passive voice rule, someone fixed overcorrection on the word "actually," someone caught a documentation error where the pattern count was wrong. These are the kinds of small-but-real contributions that indicate people are using this and finding edges.
OpenCode support. Added recently — if you're using OpenCode instead of Claude Code, it works there too, and the ~/.claude/skills/ directory is shared between both tools, so one clone covers you.
Who Should Use This
Use it if: You're writing anything that will be read by humans who care about quality — technical documentation, blog posts, emails to people you want to impress, anything where "this was clearly AI-generated" would be embarrassing or damaging to your credibility. Also useful if you're reviewing AI-generated content from others and want a structured way to think about what to fix.
Use it if: You're already using Claude Code or OpenCode as your primary AI interface. The installation is two commands and it integrates naturally into your existing workflow.
Skip it if: You're generating content where quality doesn't matter — internal notes, quick summaries, throwaway drafts. The double-pass approach takes time, and it's overkill for text nobody will read carefully.
Skip it if: You're not using Claude Code or OpenCode. This is not a standalone tool, a web app, or a library. It's a skill file for a specific AI coding tool. If you're using ChatGPT or the Anthropic web interface directly, this doesn't help you without adaptation.
Be skeptical if: You're hoping this will make AI writing undetectable to AI detectors. That's not what this is for, and I'd be cautious about anyone selling that promise. This is about writing quality, not gaming detection systems.
Honest Concerns
No releases, no versioning. The repo has no tagged releases. You're cloning from main. For a tool you're integrating into your writing workflow, that's a minor annoyance — you have no stable version to pin to, and updates could change behavior in ways you don't expect. Not a dealbreaker, but worth noting.
The skill is only as good as Claude's instruction-following. This is a SKILL.md file with detailed rules. Whether Claude actually applies all 29 patterns consistently depends on the model's ability to follow complex multi-part instructions, which is not perfect. You'll still need to review the output. This is a tool to assist your editing judgment, not replace it.
42 open issues. That's not alarming for a repo this size, but it's worth checking what's in there before you rely on it heavily. The recent commit activity suggests the maintainer is engaged, but I haven't audited the issue backlog.
The voice calibration feature is relatively new. Added in March 2026, it hasn't had much time to accumulate feedback. It's a promising idea but I'd treat it as experimental until more people report on how well it actually captures individual voice.
Single primary maintainer. Blader has 17 of the commits. There are contributors, but the bus factor is real. This is a side project, not a funded product. Plan accordingly.
Verdict
Install it. The investment is two commands and the pattern taxonomy alone is worth the five minutes. If you're using Claude Code regularly and generating any text that matters, this is a straightforward quality improvement to your workflow.
Just be clear-eyed about what it is: a structured prompt with good source material, not a magic solution. You still have to read the output. You still have to make judgment calls. But having 29 named patterns as a shared vocabulary between you and the model is genuinely more useful than the alternative, which is "please make this sound less like a robot."
The voice calibration feature is the thing I'm most interested in watching develop. If it works reliably, that's the piece that makes this actually useful for professional writing rather than just cleanup.