humanizer

A Korean-language writing analyzer that detects the characteristic patterns of AI-generated Korean text and rewrites it to read as natural human writing. It is scoped strictly to Korean, leaving English translations and other languages aside, and it grounds its detection in empirical linguistics, chiefly the KatFishNet paper (ArXiv 2503.00032v4), citing measured accuracies such as 94.88% AUC for comma patterns, 82.99% AUC for part-of-speech diversity, and 79.51% AUC for spacing. The goal is to make text sound as though a native Korean speaker wrote it while preserving the original meaning and intent, which makes it a fit for cleaning up ChatGPT, Claude, or Gemini output in Korean.

Detection spans 40 distinct patterns organized into six prioritized categories: punctuation (led by comma overuse), spacing rigidity, part-of-speech diversity, vocabulary (AI buzzwords, unnecessary Sino-Korean, English calques, plural and pronoun overuse), structural monotony, and translation-ese. Each pattern carries metadata for validation level, a severity of S1 critical, S2 strong, or S3 weak, and the version it was added. Findings are reported severity-first with concrete instances and the scientific or empirical rationale, followed by a natural rewrite and an optional summary of key changes.

Several guards protect meaning. A mandatory six-point self-check verifies that numbers, dates, and proper nouns, causal direction, negation, direct quotations, core claims, and formality level all survive the rewrite. An over-rewriting guard estimates a change rate and escalates: under 30% passes, 30 to 50% adds a caution line, and above 50% halts and asks the user before producing a more conservative rewrite. Version 1.6.0 adds a single-letter Naturalness Grade from A to D that summarizes how natural the rewritten text reads. The skill is permitted to use Read, Write, Edit, Grep, Glob, and AskUserQuestion tools.