voidful/michelin

Michelin is a skill made for AI coding assistants that tries to make text and visual design written or reviewed by AI feel less like an AI wrote it, in both Traditional Chinese used in Taiwan and English. Rather than just swapping out a few overused words, it works in four steps: first identifying facts and constraints in the text that must not change, then figuring out what specifically feels templated, whether that is structure, tone, visual layout, or wording that does not match how people in Taiwan actually speak, then only fixing the specific issues found, and finally checking the result through a blind side-by-side comparison against real human writing. It can be used to review website copy, product interfaces, documents, presentations, and marketing materials, either checking existing drafts or helping build a sense of a person's or brand's voice and design style from real examples they provide. The project is explicit that outside AI-detection tools are not treated as a reliable judge of success, instead, success means the AI-written version becomes hard to tell apart from human writing in a matched blind comparison, without the quality getting worse just to pass the test. The repository includes a Python-based checker that scans text or web pages for common AI writing patterns and Taiwan-specific language signals, plus a set of scripts for running the anonymous comparison process: preparing test cases, collecting judgments from independent reviewers, and generating a report. It also includes a packaging script that bundles the skill into a distributable file with a checksum. The author notes clear limits: there is no single fixed definition of what sounds human, blind tests can only show something is hard to distinguish under that specific test setup rather than proving anything about its origin, and image or video content still needs human judgment. The project is unrelated to the Michelin Group or Michelin Guide and carries no official rating meaning. It is released under the MIT license.