Stop pasting text into humanizer

Follow this StackScout guide in my project: "Stop pasting text into humanizer"
https://stackscout.app/guides/humanizer-in-your-pipeline

Read it, then set up what it describes. When you're done, give me a one-line summary.

Almost everyone uses blader/humanizer the same way: paste a paragraph, ask for a human version, copy the result back. That works, and it is the weakest of the three modes the skill actually has. Here is the setup that changes what your agent produces by default.

What the skill is

One markdown file. 33 numbered patterns, each with a before and an after, taken from Wikipedia's "Signs of AI writing" page. MIT licensed, 32,720 stars on 2 August 2026.

Install it from the author, so updates and security stay with them:

npx skills add blader/humanizer --global

The three modes, and the one that matters

The skill names its own invocation modes. Straight from SKILL.md:

• Pasted text (default). You paste, it answers with a draft, an audit and a final version.
• File mode. You point at a file. It rewrites the file in place and reports a summary instead of pasting the whole thing back.
• Embedded mode. "Another task or agent is using this skill as one step of a larger job." No draft, no audit, no ceremony. Just the finished prose.

Embedded is the one worth wiring up, because it moves the skill from something you remember to use to something that runs whether you remember or not.

Put this in your CLAUDE.md or AGENTS.md:

## Writing rules
Any prose you produce for a human to read — commit messages, PR descriptions,
docs, release notes, release posts — goes through the `humanizer` skill in
embedded mode before you show it to me. Output the final text only.

That single paragraph is the whole integration. Your commit messages stop announcing their own significance, your PR descriptions stop being three bullet points with bolded headers.

Give it your voice, or it will give you its own

The second thing nobody does: hand it a sample of your own writing before it rewrites anything. The skill treats that sample as outranking its own rules:

"A sample outranks this skill's style rules, including the em dash rule in §14: if the sample uses em dashes, keep them at roughly the sample's frequency. Matching the author beats scrubbing the tell."

Without a sample you get a human voice. With one you get yours. Keep three or four paragraphs you actually wrote in a file and point at it:

Read voice-sample.md first, match its habits, then humanize CHANGELOG.md.

The honest limit, and what to do about it

The repo contains SKILL.md, README.md, AGENTS.md, an OpenAI manifest and scripts/validate-package.py. That script checks the packaging, not your text. So the skill is a prompt with no verifier: the same model that wrote the rewrite decides whether the rewrite is clean.

If you want a number instead of a promise, count the tells yourself. This repo ships a small counter that reads the skill's own pattern lists at runtime and reports hits:

node scripts/ai-tells.mjs draft.md --skill ~/.claude/skills/humanizer/SKILL.md

draft.md · 162 words
26 tells · 160.5 per 1000 words

  6  §7. Overused "AI Vocabulary" Words
       crucial, enhance, fostering, landscape, testament, vibrant
  4  §3. Superficial Analyses with -ing Endings
  3  §18. Emojis

It only counts what is countable: watch lists, em and en dashes, curly quotes, emojis, bold-header lists, title-case headings. Judgement calls like rule of three or hedging stay out, because counting opinions is how you get a number nobody believes. A hit is a suspicion, not a verdict — the skill says it plainly: look for clusters, not single tells.

Where it belongs in a stack

Last. commit-pr-writer writes the message, social-content writes the post, and humanizer is the pass everything goes through before another person reads it. It edits prose, so it never competes with the skills that decide what the prose says.

One caution: over-editing kills real writing. The skill knows this and carries a "what NOT to flag" list — perfect grammar, formal vocabulary, one em dash, a single short sentence. If your text comes back flatter than it went in, you passed it prose that was already human, and the fix is to not run it.