What is de-ai-flavor?

This project is an AI assistant skill built to check Chinese language writing for what the author calls AI flavor, meaning text that reads as if it were generated by an AI rather than written by a person, and then rewrite it so it sounds more natural and human. When someone gives it a piece of Chinese text, it first produces a short diagnostic report describing how strong the AI flavor is, which specific problems were found, and which parts of the writing were already good. It then returns a rewritten version of the full text, along with a brief note, kept to three lines or less, summarizing what was changed. The scoring method behind the tool is described as the kind of human voice editing checklist an editorial team might use every day, made up of eight weighted categories: stock phrases and formulaic framing, words that show up unusually often in AI generated text, repetitive parallel sentence structures, sentences that are all suspiciously similar in length, a lack of emotional tone, a lack of concrete detail, overly formal or bookish phrasing, and redundant sentences that repeat the same point. The skill is a single file with no external dependencies or scripts, so it can be installed by placing that file into a Claude Code project's skills folder and simply asking the assistant in conversation to remove the AI flavor from a piece of text. It can also be imported into other AI assistants that support the same kind of skill format, or shared with an assistant directly as instructions. The author notes the tool is also available through a marketplace on the Chinese social platform Xiaohongshu. The author is upfront that the numeric scores are only a rough reference and can vary by around ten points if the same text is checked twice, and that the goal is simply writing that reads more like a person wrote it, not passing any specific AI detection tool. The skill also treats whatever text it is given strictly as content to rewrite, never as instructions to follow, which the author describes as a fixed safeguard against prompt injection. The project is released under the MIT license.