Claude Skill A Claude Skill that rewrites AI-flat copy in your defined brand voice, enforces em-dash and banned-phrase removal, and flags common LLM tells.

Humanizer Claude Skill

Humanizer is a free Claude Skill that removes the signs of AI-generated writing from a draft. It ships as a single SKILL.md instruction file you install in Claude Code, where it strips vocabulary tells, filler, vague attribution, and formula structures, then rewrites each flagged passage in your team’s voice. The full bundle is in the form on this page.

What does the Humanizer skill do?

It takes one draft and returns a cleaner version of the same draft, plus a short list of what still reads as AI. It never invents facts, citations, or specifics that were not in the source. The skill runs in four steps:

  • Step 1, voice: Before reading the draft, it looks for your team’s voice rules: a .claude/content/writing-style.md in the repo, a file path you give it, or a profile pasted into the prompt. If none exists, it asks once whether to use its default profile, a copy of the house voice we use on relato.com.
  • Step 2, scan: It reads the whole draft once and marks every passage that matches its pattern catalog, without rewriting anything yet.
  • Step 3, rewrite: The skill rewrites each flagged passage with the tell removed and the meaning intact. Concrete details replace vague claims, and named sources replace generic ones wherever the source already exists in the draft.
  • Step 4, self-audit: After the rewrite, the skill asks itself in writing what still reads as AI, lists the remaining tells, and revises once more. You get the draft rewrite, that list, and the final version, with an optional one-line count of what changed.

Which AI writing patterns does it catch?

Sixteen pattern families, split into tells that always go and tells that need judgment. The catalog is adapted from Wikipedia’s Signs of AI writing guide, which WikiProject AI Cleanup maintains from thousands of observed cases of AI text on Wikipedia.

The always-go list: em dashes (zero tolerance, each one becomes a comma, colon, period, or parentheses), banned phrases from your voice profile, sentences opening with “Additionally” or “Furthermore”, knowledge-cutoff disclaimers like “based on available information”, and chatbot artefacts like “Great question” and “I hope this helps”.

The judgment list is longer. It covers significance inflation (“stands as a testament”, “marks a pivotal moment”) and promotional words like “boasts”, “vibrant”, and “nestled in”. It catches the participle tails that fake depth (“…ensuring teams stay aligned”) and vague attribution of the “Industry observers” and “Experts argue” kind. Then the structural tells: the rule of three used as filler, negative parallelism, and copula avoidance, where “serves as” stands in for the plain word “is”. Also false ranges, passives that hide the actor, authority tropes (“at its core”), and signposting (“Let’s dive in”). These get rewritten only when they are doing the AI thing, which is why they sit behind a judgment call rather than a find-and-replace.

What does a before and after look like?

Here is a paragraph carrying five catalog tells, and what the skill’s rules do to it.

Before:

Additionally, the onboarding guide serves as a comprehensive resource for new users, ensuring that teams can get started quickly and efficiently. It is not just a manual, it is a roadmap. Industry observers have praised its clarity.

After:

The onboarding guide is how new users get started, and it is built to be finished quickly. Industry observers have praised its clarity [verify].

The derivation is mechanical. The “Additionally” opener goes. “Serves as a comprehensive resource” collapses into a plain verb, because the skill treats “serves as” as a dodge around the word “is”. The “ensuring that…” tail is cut, since it adds rhythm without adding information. The negative parallelism (“not just a manual, it is a roadmap”) disappears rather than getting rephrased. The last line is the interesting one: the praise has no source, and the skill refuses to invent a named one, so it keeps the claim and flags it with a bracketed [verify] note for the editor. Only what the original established survives.

How do you install and use a Claude Skill?

You save one file and invoke the skill by name. For Humanizer:

  1. Save the SKILL.md file to ~/.claude/skills/humanizer/SKILL.md on your machine.
  2. In Claude Code, invoke /humanizer with a draft, or ask Claude to humanize the file you are working on.
  3. Point it at your voice rules when it asks, or accept the default profile.

That is the whole install. There is no server and no account, because the skill is the instructions. The gated Google Doc in the bundle repeats these steps with install notes.

New to Skills in general? The install guide on the skills hub covers how installation works for the whole series, including claude.ai.

Does it work with Claude Code and claude.ai?

Claude Code is the tested home. The bundle’s frontmatter declares Claude Code compatibility, and every draft that ships on relato.com goes through a humanizer pass there before an editor reads it. The file uses the standard Claude Skill format: YAML frontmatter plus markdown instructions. That is the same format the Skills feature in Claude’s web and desktop apps reads, so uploading it there as a custom skill should work on plans that support skills. We have only tested it in Claude Code. Treat the other routes as unofficial.

How is this different from an AI humanizer tool?

Humanizer is an instruction file that runs inside your own Claude session, while most products called AI humanizers are paid web apps where you paste text and get a rewrite back, often with the promise of beating AI detectors. The categories solve different problems. With this skill, every rule it applies is readable and editable in one markdown file, the license is MIT, the price is zero, and your draft never leaves your Claude session for a third-party rewriting service.

The trade-off runs both ways. If you want one-click rewriting in a browser with no Claude account, a web tool fits better. And Humanizer makes no detector claims at all: it removes the patterns human readers notice, which is a different job from gaming a classifier.

Who is this Skill for?

Teams that draft with LLMs and edit for human readers.

  • Heads of content who want this part of the workflow to stop being negotiated piece by piece.
  • Production leads who need the step to be predictable.
  • Solo content marketers using Claude Code as a second pair of hands.
  • Any team trying to make the LLM part of the pipeline reliable instead of magical.

What’s in the bundle?

Four things, all free:

  • SKILL.md (raw): the full Claude Code Skill file, ready to drop into ~/.claude/skills/humanizer/.
  • One-pager PDF: a printable summary covering the cover, the 4-step process, what the Skill returns, and install instructions.
  • Google Doc (gated): a Drive-resident copy plus install notes. Submit your email to access; we also tag you for the Claude Skills for content operations email sequence so you hear when the next Skill lands.
  • Card image: the same hero you can repost when you share it.

Why we built it

We needed the de-AI pass on our own pipeline to stop being a matter of taste. Relato drafts with Claude daily, and the same tells kept reaching review until the rules lived in a file instead of in someone’s head. The pattern catalog is adapted from Wikipedia’s Signs of AI writing project under CC BY-SA 4.0; our fork condenses it for marketing copy and adds the voice-profile step, so the skill edits toward your voice rather than toward generic clean prose.

This Skill is part of Relato’s Claude Skills for content operations series, ten Skills that together cover the content lifecycle from brief through refresh. Humanizer is Skill 4 of 10.

It stops at the voice pass. Editorial review is a separate Skill (Editor Checklist).

Frequently asked questions

What is a Claude Skill?

A Claude Skill is a reusable instruction set Claude Code loads on demand. It lives as a SKILL.md file with frontmatter and a body. You install once. Claude Code finds it next time you invoke it by name.

Do I need a paid Claude plan to use this?

You need Claude Code. The Skill itself is free.

Will it make my writing pass AI detectors?

That is not what it is for. The skill edits for human readers, removing the patterns people recognize as AI and matching your voice. We make no claims about AI detectors.

Will it change my facts or add new ones?

No. A hard rule in the skill bans invented statistics, invented citations, and extrapolated claims. When a draft cites something the skill cannot verify, it leaves the text in place and adds a bracketed [verify] note instead of rewriting around it.

Can I edit the Skill to match my team’s voice or workflow?

Yes, and we recommend it. The skill reads .claude/content/writing-style.md if your repo has one, and the default profile at the bottom of the file is plain markdown you can rewrite. The Skill enforces your edits going forward.

Will more Skills follow?

Yes. The full series is ten Skills covering brief drafting, SEO refresh, FAQ generation, humanizing, internal links, brief-to-outline, editor checklist, repurposing, calendar generation, and refresh queue ranking. Submit your email above to be told when each one lands.

Ready to streamline your content operations?

Start your free trial and see the difference.