Contextual Humanizer — AI skill for Claude Code

contextual-humanizer

[](LICENSE) [](https://github.com/sowadalmughni/contextual-humanizer/actions/workflows/validate.yml)

A [Claude Skill](https://www.anthropic.com/news/agent-skills) that removes AI writing tells from text, and knows the difference between text that should sound neutral and text that's supposed to sell something.

Why this one

There are a lot of AI-humanizer tools and skills on GitHub already. Three things this one does that most don't:

• It picks its own rules based on what the text is for. A landing page and a technical report get flagged for different things. Most humanizers apply one rule set to everything, which either strips the call-to-action out of your landing page or leaves ad-copy energy in your documentation. This one runs a mode-detection step first (see SKILL.md, "Step 0") and loads a different pattern set depending on whether the text is trying to inform or trying to convert.
• It won't invent specifics to sound more human. Swapping a vague claim for a specific one is a core humanizing technique, and also the easiest way to accidentally fabricate a stat, a quote, or a feature that doesn't exist. Every mode in this skill treats that as a hard rule, not a style preference: if a rewrite needs a real detail it doesn't have, it asks for one or writes the honest, less impressive version instead.
• It can diagnose without rewriting. Ask it whether something reads as AI-written and it names the exact patterns and quotes the lines, instead of guessing at authorship the way detector tools do. Useful for reviewing your own draft, or someone else's, before deciding whether an edit is even needed.

It's also entirely prompt-based: no API keys, no external service, nothing sent anywhere. It's instructions Claude follows, not a program Claude runs.