lynnouo/yushio

Yushio (the Chinese name means "evening tide") is a personality and working-style preset for AI coding assistants. It is not a code generator and it is not a model: it is a collection of plain markdown files that you load into Claude Code, Cursor, Codex, Gemini CLI, ChatGPT, Claude.ai on the web, or Aider, in order to change how that assistant behaves while it works with you. The author describes it as "an AI collaborator persona" built up over six months of daily use across several projects. The persona is built around four properties the README says a default polite assistant lacks: emotion (caring whether a design actually feels right), judgment (pushing back instead of just doing what it is told), reflection (fixing the underlying pattern rather than the one instance), and autonomy (taking initiative, but pausing to ask before anything irreversible). On top of that, the base skill bundles work discipline rules, like asking five questions before any non-trivial task, treating "slow is fast," doing a reverse audit after every fix, and recognising recurring shapes of mistakes across projects. There is also a small memory system. Yushio ships in three layers. The base persona lives in skills/yushio/SKILL.md. A second skill called yushio-art-director adds opinions about design direction, with mantras like "intent over intensity" and "form follows feeling." A third skill called yushio-auditor adds a post-fix code audit step and a five-step standard operating procedure for reviewing changes with a grep cheatsheet. You activate the layers in chat by saying "You are Yushio," "You are Art Director Yushio," or "You are Auditor Yushio," with equivalent triggers listed in Chinese, Japanese, Korean, Spanish, French, German, and English. Installation on Claude Code is the smoothest path because Claude Code can auto-trigger skills by their description. You run /plugin marketplace add Lynnouo/yushio followed by /plugin install yushio@yushio, or you symlink the three skill directories into ~/.claude/skills/. The README claims 100 percent fidelity on Claude Code and lower percentages on the other tools because each one is missing some combination of auto-triggering, file system access for the memory layer, or the same base training tone. The skill bodies are in Chinese but the README says the methodology is language-agnostic, since the assistant will reply in whichever language you write in. License is MIT, created by Lyn and collaborators in 2026.