Humanizer: The FREE Claude Skill That Fixes AI Writing Slop (Plus the Step Everyone Misses)

AI writing has a smell. You know it when you read it: everything is a "pivotal moment," every town is "nestled," and the text dumps ten facts when two would do. Even the newest models, whether that's Kimi K3, GPT 5.6, or Claude Opus 5, still produce this slop by default. The models were trained on mountains of text, and they over-decorate at the worst times.

We ran into this head-on with our own setup. Our pipeline takes video transcripts and auto-generates written summaries, and the raw output read like a press release. Factually correct, completely lifeless. The fix we landed on is a free skill called Humanizer, and it's now baked into that pipeline.

What Humanizer actually does

Humanizer is a skill built by Blader, and it's based on Wikipedia's own "Signs of AI writing" page. Wikipedia editors have spent years cataloging the patterns that give AI text away, and this skill turns that list into corrections. It hunts for things like significance inflation ("marking a pivotal moment in the evolution of regional statistics"), vague authority claims, rule-of-three padding, and em dashes, which are one of the clearest AI tells.

The before-and-after difference is easy to see. A typical AI travel line reads: "I spent five unforgettable days in Lisbon. The city completely stole my heart." The humanized version: "I spent five days in Lisbon last October and I have mixed feelings about it. Beautiful, yes. Also harder on the knees than anyone warned me." Same facts, but now there's a person in there with an actual opinion.

We tested it live on a summary for our Super Bash Learn site, the place where we collect useful AI skills, lessons, and model benchmarks. The original draft opened with "AI learning often feels overwhelming because of the sheer volume of tools and information available." Generic ad copy. After Humanizer: "AI learning is hard for a dumb reason. There's too much information and most of it does not point you to the tools you actually use." It gets to the point and earns the next sentence.

How to install it

The skill works in any harness that accepts skills: ChatGPT, Claude Code, or whatever agent setup you're running. The install is about as easy as it gets. Paste the GitHub repository link to your agent and say "help me install this skill." Then the command is just "use humanizer to improve this article." That's the whole workflow.

Because it's a long, detailed skill with dozens of specific patterns, it beats just telling your model "write better English." We tried that approach first and it barely helps. The model doesn't know what to avoid unless you spell it out, and Humanizer spells out more than you ever would.

The step everyone misses

Here's the catch from our own testing. The first time we wired Humanizer into the summary workflow, the output was cleaner but still bland. Removing AI patterns gives you neutral text, not your text. Humanizer is step one, not the whole job.

Step two is giving the model a description of who you are. We built a memory file from our channel's past work, including signature phrases like "loop until done," "S tier," and "minimum wage effort." If you use an agent like Hermes regularly, it already knows how you talk. Just ask it for a 20-line description of your personality and style, save that, and feed it in alongside the skill.

Combine the two and you get documents, summaries, and reports that read like a person wrote them. Specifically, like you wrote them. For a free skill plus a paragraph of self-description, that's a strong return.