humanizer

A skill that rewrites prose to remove signs of LLM authorship and inject specific, varied, opinionated human voice. Drop-in for any agent harness that loads SKILL.md files (works in ~/.claude/skills/, ~/.codex/skills/, and similar).

Originally adapted from blader/humanizer. This fork tightens the skill around two ideas:

1. Hard gates over soft suggestions. The most common AI tells (em dashes, curly quotes, curly apostrophes inside contractions, bolded inline-list headers, decorative emoji) are zero tolerance and enforced by a literal pre-output scan, not a vibe check. A soft "use em dashes less" suggestion buried halfway down a checklist is exactly why em dashes leak through.
2. One-phase output, two-phase reasoning. The model still drafts, audits its own draft, and revises. But the deliverable is just the final clean text. No "here is a draft, here is what I noticed, here is the final version" three-section response, which gets skipped in agentic flows anyway and clutters output when it doesn't.

Patterns covered are based on Wikipedia: Signs of AI writing, maintained by WikiProject AI Cleanup.

What's in the skill

The skill enumerates 31 named patterns across content, language, style, communication, and filler. A few that the upstream skill underweights or omits:

• Em-dash substitutes. When told "no em dashes," models commonly switch to en dashes, double hyphens, or spaced hyphens that play the same syntactic role. The skill calls these out as also forbidden, with a literal scan step.
• Curly apostrophes inside contractions (don't, it's, we're). The most common quote leak and the hardest to catch by eye.
• Sentence-initial transition pile-up. Opening sentence after sentence with Additionally, / Moreover, / Furthermore, / Notably, / That said,.
• Intra-document register shift. When one paragraph reads like a person and the next reads like a press release, because the rewrite touched some sections and not others.
• "Across" as a soft broadener (across industries, across teams, across the board).
• Tiered vocabulary with diagnostic collocations. Phrase-level patterns like delve into the intricacies of, harness the power of, in today's fast-paced world, through the lens of, the ever-evolving landscape of carry far more signal than single words. Rule #7 splits collocations from words, then tiers each by signal strength so the model knows which are deal-breakers versus which only flag in clusters.

Install

Drop SKILL.md into your agent harness's skills directory. Common locations:

~/.claude/skills/humanizer/SKILL.md
~/.codex/skills/humanizer/SKILL.md

The skill activates when you ask the agent to humanize text, or via /humanizer if your harness exposes skills as slash commands.

Use

Pass text you want humanized. Optionally provide a writing sample for voice matching.

Inline:

Humanize this text. Here's a sample of my writing for voice matching:

[paste sample]

[paste text to humanize]

From a file:

Humanize this text. Use my writing style from ~/Documents/my-writing.md as a reference.

[paste text to humanize]

The skill returns only the final humanized text. If you want the audit (what was AI about it, what changed), ask after the fact.

When to skip the skill

• Code, commits, configs.
• Prose Claude is currently drafting from scratch. The patterns are easier to apply inline as you write than to clean up afterward.

Credit

Adapted from blader/humanizer. Pattern catalog draws from Wikipedia: Signs of AI writing.

License

MIT. See LICENSE.