What is this?
Sloptrim is a local detector for AI-writing patterns: it scores prose files your coding agent saves against documented stylistic tells, using only Python's standard library with no network calls or ML model.
Sloptrim is a local detector for AI-writing patterns: it scores prose files your coding agent saves against documented stylistic tells, using only Python's standard library with no network calls or ML model.Is this a hosted service or something I run locally?It is purely a command-line tool and Claude Code agent plugin: there is no website, no hosted version, and nothing you write ever leaves your machine.
It is purely a command-line tool and Claude Code agent plugin: there is no website, no hosted version, and nothing you write ever leaves your machine.What do I need installed besides my coding agent?You need Node for the hooks and Python 3.9 or newer for the detector itself; nothing else is required.
You need Node for the hooks and Python 3.9 or newer for the detector itself; nothing else is required.Will it catch everything my agent writes?No: it never sees files written by a Bash command since those reach disk without going through Write or Edit, and it only scores the first 262,144 characters of extracted prose per file.
No: it never sees files written by a Bash command since those reach disk without going through Write or Edit, and it only scores the first 262,144 characters of extracted prose per file.How accurate is the AI-writing detection?Across five benchmark arms (GPT-4o, Claude 3.5 Sonnet, o1-pro, paraphrased and humanized variants), Sloptrim scored ROC-AUC between 0.762 and 0.946, but the authors stress AUC is a ranking measure, not a guaranteed yes/no accuracy at the guard threshold.
Across five benchmark arms (GPT-4o, Claude 3.5 Sonnet, o1-pro, paraphrased and humanized variants), Sloptrim scored ROC-AUC between 0.762 and 0.946, but the authors stress AUC is a ranking measure, not a guaranteed yes/no accuracy at the guard threshold.Which file formats can it score besides Markdown?It handles about 20 formats, including .docx, .pptx, .xlsx, OpenDocument files, .epub, .ipynb notebooks and LaTeX, scoring up to the first 262,144 characters of extracted prose.
It handles about 20 formats, including .docx, .pptx, .xlsx, OpenDocument files, .epub, .ipynb notebooks and LaTeX, scoring up to the first 262,144 characters of extracted prose.Can it prove a specific piece of writing was AI-generated?No, the authors explicitly say it cannot prove whether a model wrote something and must not be used as an authorship classifier; a score reflects writing style, never a person.
No, the authors explicitly say it cannot prove whether a model wrote something and must not be used as an authorship classifier; a score reflects writing style, never a person.What's the difference between full and strict mode?/sloptrim full (the default) applies the writing contract plus a guard that nudges above a score of 40, while /sloptrim strict lowers that threshold to 20 and also asks for a character scrub.
/sloptrim full (the default) applies the writing contract plus a guard that nudges above a score of 40, while /sloptrim strict lowers that threshold to 20 and also asks for a character scrub.What are similar projects?yomiyasu, explainer, sepia, linkedin-agent-skill, Tech-Doc-Style-Chinese, planning-with-files, im-not-ai, pm-skills
yomiyasu, explainer, sepia, linkedin-agent-skill, Tech-Doc-Style-Chinese, planning-with-files, im-not-ai, pm-skillsWhich topics is it in?humanizer, claude-code-skills, docx, writing-tools, zero-dependency, linter, claude-code-plugin, ci, testing, coding-agent, markdown, agent-skills
humanizer, claude-code-skills, docx, writing-tools, zero-dependency, linter, claude-code-plugin, ci, testing, coding-agent, markdown, agent-skills