What it does
unslop removes common tell-tale habits of AI-written text, such as flattering openers, stock vocabulary, stacked hedges, filler transitions and heavy em-dash use, and keeps the facts and meaning of the original. Code blocks, inline code, URLs, headings, tables, blockquotes and YAML frontmatter are preserved. In deterministic mode they are kept byte-identical; in LLM mode the model is told to preserve them. It runs as a plugin or rules bundle for Claude Code, Cursor, Windsurf, Cline, Gemini CLI and OpenAI Codex, or as a standalone command-line tool installed with pipx. Deterministic mode uses pattern rules with no network calls, while LLM mode outside an assistant needs an Anthropic API key. There are five modes: subtle, balanced, full, voice-match and anti-detector. Voice-match can reuse a saved numeric profile of your own writing. The author describes it as a polish layer rather than a way to defeat AI detectors, and advises turning it off for text that must stay exact, such as legal text, medical warnings or runbooks. The project states that it sends no telemetry.
Capabilities
- Five modes from subtle to full rewrite, including voice-match
- Deterministic pattern mode that works offline with no API key
- Preserves code, URLs, headings, tables and frontmatter
- Sub-skills for commit messages, PR review comments and markdown files
- Optional stripping of reasoning wrappers from agent output
When to use it
- You want AI-drafted cover letters, essays or posts to read in a more natural voice
- You want agent output cleaned up before it lands in docs, commits or review comments
- You need an offline, rule-based pass over markdown files