Seyed Ehsan Hadi writes about a local command-line tool and agent plugin that scores prose files for LLM-writing patterns using only the Python standard library, with no network access and no model. It works by hooking into coding agents like Claude Code to evaluate files saved on disk, flagging 71 documented patterns to prompt rewrites.
- Uses stylometry and lexical checks rather than an LLM
- Achieved a ROC-AUC of 0.946 on the GPT-4o benchmark arm
- Supports 20 file formats including .docx, .epub, and .ipynb
- Explicitly warns against using its scores as an authorship classifier
- Uses stylometry and lexical checks rather than an LLM
- Achieved a ROC-AUC of 0.946 on the GPT-4o benchmark arm
- Supports 20 file formats including .docx, .epub, and .ipynb
- Explicitly warns against using its scores as an authorship classifier
