About this skill

Problem

A document can look well-structured and professionally worded while still being AI-generated or based on an AI first draft. Judging by tone alone is unreliable, especially when the text contains industry terms, precise parameters, changelogs, or numbered sections, since those can come from prompts or model training data. This skill turns AI-trace review into a rule-based scan with scores, paragraph hits, and quoted evidence.

How It Works

It reads .docx, .pdf, .txt, .md, or pasted text, splits the content by paragraphs, and checks 32 writing patterns, including AI high-frequency vocabulary, rule of three, self-disclosure statements, and low syntactic fingerprint variation. Each dimension is scored from 0 to 10, then weighted into an overall 0-100% AI likelihood, with the report written to outputs/ai-audit-...md.

Limits

The score is a review signal, not a final verdict. Standardized contracts, manuals, or policy documents may naturally contain templated phrasing. Domain detail and precise numbers cannot independently prove human authorship. Pay special attention when a text has rich case details but weak causal logic, or when multi-round human edits leave a single underlying syntactic fingerprint.