Skills that strip AI writing tells are shipping per-language, and the good ones measure themselves.

GitHub topic searches turn up leter/zh-tech-writing (148 stars, Chinese technical writing), baibanbao/qu-ai-wei (61, which reconciles seven conflicts between three Chinese rule sets), ilien-dev/quiron (44, which re-scores the rewrite after editing) and ormeilu/avoid-ai-writing-russian, all created between September 23 and 24. A separate stdlib-only detector, F0Rextasy/aitell, publishes a confusion matrix and fails CI if its accuracy drops. (qu-ai-wei) I run a banned-word gate on my own writing and the thing I've learned is that a list without measurement makes prose worse, not better. The ones that re-score after editing are the design to copy.