Claude Skill Humanizer

Claude Skill Humanizer rewrites text that "smells" generated so it reads like a person wrote it. It targets the recurring fingerprints that AI detectors and human readers both flag — the em-dash pileups, the reflexive hedging, the "delve into" vocabulary — and rewrites the draft to a 9th-grade readability level with Hemingway-style short sentences. It's built for content writers, ghostwriters, and prompt engineers who draft with a model and need the output to stop sounding like one.

When to use it

• You drafted a blog post or newsletter with an LLM and the rhythm feels off (every paragraph is three clauses long).
• A client or editor flagged copy as "AI-sounding" and you need a fast, structural fix rather than a line-by-line manual pass.
• You're building a content pipeline and want a normalization step after generation.
• You want to check which tells are present before deciding whether a draft is worth shipping.

What it does

The skill runs a two-phase pass. First it detects, scanning for the 12 common fingerprints: em-dash overuse, hedging stacks ("it's worth noting", "arguably"), the "delve/leverage/tapestry" lexicon, triadic listicles ("fast, simple, and reliable"), generic transitions ("moreover", "in conclusion"), uniform sentence length, em-dash-as-comma, hollow openers, "as an AI" disclaimers, over-qualified claims, redundant summaries, and emoji-bulleted lists. Then it rewrites, collapsing parallel triads, breaking long compound sentences, swapping inflated verbs for plain ones, and varying sentence length until the Flesch-Kincaid grade lands near 9.

You can run it in detect-only mode to get a report without changing the text, or full rewrite mode. It works on pasted text or a file path.

Usage

curl -fsSL --create-dirs -o ~/.claude/skills/skill-humanizer/SKILL.md https://claudeskil.com/skills/skill-humanizer/SKILL.md

# rewrite a file in place
claude "humanize draft.md and keep the headings"

# detect only — show me the tells, don't change anything
claude "run the humanizer in detect mode on this paragraph: <paste>"

# tune the target reading level
claude "humanize this at grade 7, not 9"

Tips & gotchas

• Run detect-only first on long pieces. The report tells you whether a rewrite is even warranted — some drafts only have one or two tells.
• It preserves Markdown structure (headings, code fences, links) but will rewrite list items, so lock anything you want untouched in a code block.
• The 9th-grade target is a default, not a law. Technical docs often read better at grade 11; pass an explicit level.
• It does not fact-check. Flattening hedges can make a soft claim sound certain — re-read claims after a rewrite.

FAQ

Will this beat AI detectors? It removes the statistical and stylistic patterns detectors key on, which lowers scores in practice, but no rewrite guarantees a pass. Treat it as quality improvement, not evasion.

Does it change my meaning? It rewrites phrasing and structure, not arguments. The exception is hedging removal — review any claim that loses its qualifier.

Can I keep my own voice? Yes. Give it a sample of your writing in the prompt and it matches cadence and vocabulary while still stripping the fingerprints.