Ryugi 流儀

ryoo-ghee · /ˈrjuːɡi/ · Japanese for “a school's own way of doing things”

Artifacts for coding agents.
Your agent draws its options instead of describing them. You pick, comment and pass.

Ask an agent for a settings page and you get three paragraphs describing three layouts. Ask for a design and you get a wall of Markdown you can't point at.

Ryugi gives the agent pages to put that work on, and gives you a way to answer on them:

• Artboard — every option drawn side by side; pick one.
• Document — a design doc with the agent's questions inline; comment on any sentence.
• Whiteboard — boxes and arrows you both draw on.
• HTML — one page with real sliders to tune a component or a motion.
• Verify — the built screen next to what you approved, and a report of what the agent proved by running it; pass or reject.

Your picks, comments and verdicts go straight back to the agent session that made the page. Every artifact lives in your project, opens from one local server, and becomes a link when you want to share it.

npx ryugi-artifacts install

It asks one question and installs Ryugi into Claude Code:

• Artifacts only — send Show me three directions for a settings page and an artboard opens.
• The full workflow — artifacts plus skills that brainstorm, spec, build under TDD and verify. Send Let's make a react todo list and the agent opens an artboard before it writes code.

Harnesses · CLI · Full workflow · Ryugi Cloud

Use it when there is more than one way it could look.

The agent draws every option side by side and marks the one it recommends. Pick, comment on any pixel, and Submit sends your answer back.

Use it when the decision is structure, architecture or tradeoffs.

The agent asks its questions inside the doc and links the boards it drew. Comment on a sentence or edit in place; your edits and the agent's merge live.

Use it when the parts and arrows are still the question.

You both draw on it. Select what you drew, ask about it, and the answer lands next to your sketch.

Use it when one thing needs tuning: a component, a type scale, a motion.

A one-page bench whose sliders keep their values when the agent rewrites the page.

Use it when the work is built and you decide whether it matches.

The built page sits next to the design you approved. Pass it, or reject it and say what is off; it comes back fixed.

What a screen can't show goes in a verify report: one line per claim the agent checked by running something — a test run, an API response, a migration count, which build is deployed — with the raw output one click away, and a "Not verified" list of what it couldn't check and why. You read it top to bottom and pass, reject or waive; screens still get judged on the artboard, and the report links to them.

Use it when you'd rather point at the thing than describe it in the terminal.

Pin a comment on a board, a sentence or a sketch, and it reaches the agent session that opened the page. In Claude Code it arrives on its own, with no message from you; the agent applies it, replies under your comment and updates the page while you watch. Questions, picks and verdicts travel the same way.

Use it when you want last week's mockup back, or a link to share.

npx ryugi-artifacts

Local first, no subscription, no agent session needed. One server covers every project on the machine, with ⌘K to jump and links that survive a restart. publish shares one to Ryugi Cloud.

For a ryugi command in your terminal, install the CLI globally:

npm install -g ryugi-artifacts
ryugi
ryugi /path/to/project

ryugi-artifacts remains an alias. Installing the plugin alone does not add ryugi to your PATH; the agent uses the plugin's launcher. The global CLI is optional and updates separately with ryugi update. See the CLI guide for commands and options.

Artifacts work on their own. If you also want the agent to build from what you approved, pick the full workflow at install (or run npx ryugi-artifacts install --mode full later). Its skills make the agent reach for the artifacts at the right moment, and they fire on their own:

flowchart LR
  I[Idea] --> B[Brainstorm] --> A[Artifacts] --> S[Spec] --> T[Build · TDD] --> V[Verify] --> L[Land]

After you sign off, the agent slices the spec into tasks, proves each one red-green, keeps a ledger of judgment calls across compaction, gets one fresh review of the branch, and reports how to land it. Debugging and worktrees are covered too. How it works →

Have Node.js and npm available in the agent's environment. The plugin calls its own launcher, which downloads ryugi-artifacts on first use and caches each release under ~/.cache/ryugi/artifacts/ by default. It checks for updates at most once a day and can reuse a compatible cached release offline. The first run needs network access and a writable cache.

Ryugi replaces Claude's hosted Artifacts with the Artboard, documents, whiteboards and HTML benches. Turn the hosted ones off by adding this key to the top level of ~/.claude/settings.json — keep your other settings:

{
  "disableArtifact": true
}

That is the only flag Ryugi needs; leave the rest of your Claude Code settings alone.

Fork of Superpowers v7.0.0 by Jesse Vincent and Prime Radiant (UPSTREAM.md). Issues and PRs welcome; AGENTS.md has the rules.

Skills, harness integrations and docs: Ryugi License 1.0 — free to use, modify and redistribute, with two exceptions. The Restricted Parties in Schedule A (Tencent, Alibaba, ByteDance, Baidu, Huawei, Xiaomi, Meituan, JD, PDD, NetEase, Kuaishou, Didi, Ant Group, iFlytek, SenseTime, and products such as WorkBuddy, Qwen, Qoder and Trae) receive no rights at all, and nobody may scrape the skills into a marketplace. This is source-available, not an open source license; material derived from Superpowers and contributions made before 2026-11-04 stay MIT (Exhibit B).

The artifacts application, published to npm as ryugi-artifacts, is proprietary (npm license: UNLICENSED) and not part of this repository. Releases up to 1.51.2 were published under AGPL-3.0-only.