humanize-writing

A skill that makes AI-assisted text read like a person wrote it, while writing rather than only after the fact, plus a deterministic linter that scores any text for AI-pattern density.

Most tools in this space rewrite finished text. This one works in both directions: it shapes prose as it's generated so the patterns never appear, and it cleans existing text when you paste some in. It knows a cold email shouldn't be humanized like a formal report, it can learn your voice from a few paragraphs of your writing, and it ships with a stdlib-only Python linter you can run in CI.

It is not tied to any one assistant. The skill is plain Markdown in the open Agent Skills format, so it works in Claude Code, Cursor, GitHub Copilot, Codex CLI, Amp, and any other harness that reads a SKILL.md. The linter is a standalone script that needs nothing but Python.

Grounded in primary sources: Wikipedia's Signs of AI writing (WikiProject AI Cleanup), checked against the page as of September 2026, and the corpus studies it cites: Kobak et al. on excess vocabulary across 15M+ PubMed abstracts (Science Advances, 2025), Juzek & Ward on why models "delve" (ACL 2025), Reinhart et al. on grammatical and rhetorical style (PNAS, 2025), and Russell et al. on what expert human detectors notice (ACL 2025).

Install

Any agent, via the skills CLI (installs into Claude Code, Cursor, Copilot, Codex, and others; it detects what you use):

npx skills add shir-danishyar/humanize

Manual, for any harness that reads SKILL.md. Clone it into wherever your tool looks for skills:

git clone https://github.com/shir-danishyar/humanize.git
# e.g. ~/.claude/skills/humanize-writing, ~/.cursor/skills/humanize-writing, ...

Claude Code specifically, as a plugin:

/plugin marketplace add shir-danishyar/humanize
/plugin install humanize-writing@humanize

Whatever the harness, the skill is the same three Markdown files; AGENTS.md at the repo root tells any agent how to load and apply it.

The /humanizer command

Once the plugin is installed in Claude Code, you also get a slash command:

/humanize-writing:humanizer 

It runs the skill in rewrite mode on whatever you give it and returns only the humanized text (or edits the file in place, prose only). Claude Code namespaces plugin commands, hence the humanize-writing: prefix. If you want a bare /humanizer, copy the command file into your personal commands folder; it is self-contained and works without the plugin:

curl -fsSL https://raw.githubusercontent.com/shir-danishyar/humanize/main/commands/humanizer.md \
  -o ~/.claude/commands/humanizer.md

The skill itself stays automatic for audience-facing writing; the command is for when you want to point it at a specific piece of text.

Usage

Generation mode (default). Once installed, the skill applies automatically to audience-facing writing tasks: emails, posts, articles, marketing copy, reports. Ask for the writing; it comes out without the patterns.

Write a launch announcement for our new invoicing feature.