What is unslop?

Unslop is a skill for Claude that tries to make AI generated English text read more like something a person actually wrote. Most tools that claim to do this just swap out a handful of overused words for other words, but this project argues that approach barely works, because word lists go stale fast and the words were never the real giveaway in the first place. The author points to a 2026 study from the University of Maryland and Google DeepMind that compared over sixty thousand pieces of writing from humans and five AI models. Even after researchers cleaned up the obvious wordy cliches in AI text, a detector looking only at the underlying structure, not the vocabulary, still caught it almost as often as before. The real tells are things like AI text always spelling out the point of each paragraph, following the same claim then support then conclusion pattern every time, and wrapping every ending up too neatly. To address that, the skill works on three levels at once: cleaning up formatting quirks like unusual dashes and repetitive bold text, replacing overused phrases and rhetorical patterns with plainer ones, and finally reshaping the deeper structure and confidence level of the writing so it stops sounding uniformly polished. It also lets a user teach it their own personal writing style by feeding it samples of their past writing, so future edits sound like that specific person rather than a generic idea of good writing. The author is candid that an early version was still spotted as AI generated when they posted it publicly, and used that feedback to build a newer version that catches more subtle patterns like tidy one-line closing statements. It installs as a skill file inside Claude Desktop or Claude Code, and is triggered with phrases like asking to humanize or rewrite a piece of text. A companion project adapts the same approach for Russian language text.