shirhussain/humanize

humanize-writing is a tool built to make text produced by AI assistants read more like something a person wrote. Instead of only cleaning up text after it is finished, it works while the writing is happening, shaping the output as it is generated so certain telltale patterns never show up in the first place. It can also take a piece of already written text that a person pastes in and rewrite it to remove those same patterns, while keeping quotes, numbers, citations, and proper nouns exactly as they were. The tool ships as an Agent Skill, a plain set of Markdown files following an open format, which means it is not built for just one assistant. It works inside Claude Code, Cursor, GitHub Copilot, Codex, and any other tool that reads a file called SKILL.md. Alongside the skill itself there is a separate linter, a small script written only with Python's built in libraries, that scores a piece of text for how much it resembles typical AI writing, and it can be run in an automated testing pipeline. The project is grounded in outside research, citing Wikipedia's documented signs of AI writing and academic studies that looked at word choice patterns across millions of scientific abstracts. It also notices what kind of writing it is dealing with, adjusting its idea of natural writing depending on whether the text is a casual social media post, a business email, or a formal report, since each of those calls for a different tone. The README documents 36 specific writing patterns it targets, each shown as a short before and after example, covering things like overused stock phrases, excessive em dashes, forced rule of three lists, and chatbot style closing sentences that offer to keep helping. Installing it can be done through a command line tool that detects which AI assistant a person uses and installs the skill for it automatically, by manually cloning the repository into the right folder for a given tool, or as a Claude Code plugin. The project has 4 GitHub stars and is written in Python.