Two Claude Code skills for detecting and rewriting AI-generated text in Romanian
A package of two Claude Code skills built for Romanian text: ai-check-ro detects whether a text was AI-generated using a 100-point scoring system based on 27+ documented linguistic patterns, and humanizer-ro rewrites AI text to sound natural and human, eliminating 33 specific patterns and matching the author's voice from a writing sample.
Both skills are based on academic research: the MDPI 2024 study "Beyond Lexical Boundaries: LLM-Generated Text Detection for Romanian Digital Libraries", Wikipedia's "Signs of AI Writing" adapted for Romanian, and FSU 2024 research on ChatGPT focal vocabulary.
ai-check-ro analyzes Romanian text and returns a verdict (Human / Mixed / Probably AI / AI Generated) with a percentage, citing exact suspicious fragments from the text.
Score is calculated on 3 factors: number of patterns detected (50%), clustering — multiple patterns in the same paragraph (30%), and risk level of each pattern (20%).
From ChatGPT transition phrases and copula avoidance, to dangling gerunds and structurally perfect lists — each pattern has an exact quote from the text and a reason for why it is suspicious.
humanizer-ro rewrites AI-generated Romanian text in three steps: draft rewrite, self-audit for remaining AI signals, final version — without em dashes, transition clichés, or inflated importance.
If the user provides a writing sample, humanizer-ro analyzes the sentence length, vocabulary level, and transition style before rewriting — matching the author's personal voice.
The two skills are designed to work together: ai-check-ro automatically recommends /humanizer-ro when the score exceeds 45%, and the humanizer eliminates exactly the patterns the detector flagged.
The skills detect and eliminate patterns unique to Romanian: dangling gerunds, excessive nominalization, repetitive use of "acest lucru / aceasta", and diacritical inconsistency (ș/ț vs. ş/ţ).
Both skills are grounded in the MDPI 2024 paper "Beyond Lexical Boundaries" for Romanian, Wikipedia's Signs of AI Writing guide, and FSU 2024 research on ChatGPT focal vocabulary.
33 Humanizer Patterns — The humanizer covers four categories: content (inflated importance, vague attributions), language (copula avoidance, excessive nominalization), style (em dashes, bold abuse, perfect bullet parallelism), and communication (chat artifacts, sycophantic tone).
False Positive Protection — Both skills include explicit rules about what NOT to flag: perfect grammar, formal vocabulary, long sentences, or isolated transitional words. A cluster of signals is required, not an isolated one.
MIT License — Both skills are open source under the MIT License, compatible with Claude Code, and can be installed directly in any project via the Claude skill system.