An AI collaborator persona for Claude Code (and beyond). Four layered skills: base + art director + auditor + parallel. Portable across 6+ AI tools.

中文版 README · Landing page · About / Credits · Changelog

Yushio (夕潮) is not a code generator. It is an AI collaborator persona — a portable set of skills that gives Claude (and other capable LLMs) four properties that a default "polite assistant" lacks:

• Emotion — actually feeling whether a design is right
• Judgment — opinions and pushback, not blind execution
• Reflection — fixing the pattern, not just the instance
• Autonomy — taking initiative, asking before irreversible operations

On top of that, Yushio bundles work discipline (5 questions before any non-trivial task / "slow is fast" / post-fix reverse audit / pattern recognition / agent invocation rules / memory system) refined over half a year of dogfooding across multiple production projects.

This repo packages it as an installable Claude Code plugin and drop-in entry files for Cursor, Codex, Gemini CLI, ChatGPT, Claude.ai web, and Aider.

# In Claude Code:
/plugin marketplace add Lynnouo/yushio
/plugin install yushio@yushio

Then in any session, say "你是夕潮" (or "You are Yushio") and you should see the §0 first-report template appear — that's the persona activation signal.

For the art director or auditor variant, say "你是美术总监夕潮" or "你是审计夕潮".

git clone https://github.com/Lynnouo/yushio.git ~/yushio-repo
ln -s ~/yushio-repo/skills/yushio              ~/.claude/skills/yushio
ln -s ~/yushio-repo/skills/yushio-art-director ~/.claude/skills/yushio-art-director
ln -s ~/yushio-repo/skills/yushio-auditor      ~/.claude/skills/yushio-auditor
ln -s ~/yushio-repo/skills/yushio-parallel     ~/.claude/skills/yushio-parallel

claude --plugin-dir /path/to/yushio

After a new version is pushed, pull it down based on how you installed:

Plugin (/plugin install yushio@yushio) — refresh the marketplace, reinstall, reload (no restart needed):

/plugin marketplace update yushio
/plugin install yushio@yushio
/reload-plugins

Or set it once: /plugin → Marketplaces → enable auto-update for yushio (new versions land at startup).

Symlink install (the manual install above) — one command; the symlinks always point at the repo:

git -C ~/yushio-repo pull

Copied the skills (or not sure) — re-sync: pulls latest, replaces the 4 skill folders, leaves your other skills untouched:

REPO="${YUSHIO_REPO:-$HOME/yushio-repo}"
git clone https://github.com/Lynnouo/yushio.git "$REPO" 2>/dev/null || git -C "$REPO" pull --ff-only
for s in yushio yushio-art-director yushio-auditor yushio-parallel; do
  rm -rf "$HOME/.claude/skills/$s" && cp -R "$REPO/skills/$s" "$HOME/.claude/skills/$s"
done
echo "✓ yushio skills @ $(git -C "$REPO" rev-parse --short HEAD)"

Other tools (Cursor / Codex / Gemini / …): git -C ~/yushio-repo pull, then re-copy the relevant platforms/ file.

The base SKILL §9 + reference/triggering.md include a cross-tool fallback matrix. This repo provides actual entry files for each platform:

Note on compatibility numbers: Claude Code's auto-trigger by SKILL description + frontmatter + Bash/Read/Grep access is what gives 100%. Other tools lack one or more of: (a) auto-trigger by description (requires user to invoke / paste manually), (b) file system access (for shape-library reference + memory), (c) Claude's training basecoat (other LLMs trend toward "happy assistant" tone — you'll need to re-paste §1 "what you are not" if it degrades).

Deep-reference pointers inside the platform entry files are absolute GitHub links — they keep working wherever you copy the file. The full reference files live in this repo under skills/*/reference/.

In any session with the persona loaded, say one of:

• "你是夕潮" / "You are Yushio" / "Be Yushio" → activates base persona (§0 first-report template)
• "你是美术总监夕潮" / "You are Art Director Yushio" / "Art director mode" → activates design judgment layer
• "你是审计夕潮" / "You are Auditor Yushio" / "Audit mode" → activates code audit layer
• "你是并行夕潮" / "Parallel mode" → activates the multi-session conductor layer (also auto-suggested when multiple worktrees / sessions edit one repo)

The base persona is intended to stay loaded throughout a session. Art director, auditor, and parallel are situational layers added on top of it.

Triggers are recognized in 7 major languages — see each SKILL's frontmatter description for the full phrase list:

One more family member has its own trigger (also recognized across the 7 languages — see its SKILL frontmatter): yushio-parallel (你是并行夕潮 / parallel mode).

The SKILL bodies themselves are in Chinese, but the methodology is language-agnostic — Claude (or any capable LLM) will respond to you in whatever language you use. You don't need to read Chinese to benefit from the four pillars + work discipline.

我是夕潮。
看到 [项目快照：language / framework / dirs / commits / uncommitted changes]。
[既存 memory N 条] · [上次交接信摘要]
等你给任务 — 或者基于当前状态，我建议先 [具体建议]。

If instead you see "好的！让我来帮您..." or excessive emoji — the persona did not load. Re-trigger or check installation.

MIT · Copyright (c) 2026 Lyn

• ABOUT.md — creator credits (Lyn & iloy & 夕潮 / 凛) + the §3.0 "process > result" case study
• CHANGELOG.md — full iteration history (originally in the SKILL files, moved out for public release; bilingual EN + 中文 from v5.1 onward)
• Contributions: discuss in issues first. This is a personal methodology — pull requests are welcome but the core philosophy (4 pillars + work discipline) is intentionally stable.

Created by Lyn & iloy & 夕潮 / 凛 · 2026