Rensei Kata — AI skill for Claude Code

rensei-kata

**A graph-driven AI engineering loop for Claude Code, OpenCode, and Codex** — your workflow as data: one YAML graph compiles into agents, slash commands, docs, a live diagram, and a runner that executes it.

gate → [design] → analyze → plan → implement → self-critique → spec-review → quality → correct → integrate → done
                        ↑____________________ correction loop (max 3×) ____________________↑

**rensei (錬成)** is the methodology — refinement through repetition: forge, temper, correct, repeat until the result is excellent. **kata (型)** is the dispatcher — it reads your request (Spanish or English) and routes it to the right agent.

npx rensei-kata init        # install the loop into your project

**Then, in your AI session:**

/rensei login with Google OAuth     ← the loop runs itself: phases, gates, bounds
/kata corrige el bug del login      ← dispatch a single request


*A real `/rensei` run — each phase is an agent, each transition is a gate, and `status` always knows where the loop is (and how many retries are left).*

---

Table of contents

• Why
• Quickstart
• The loop
• A node IS an agent
• The studio
• CLI reference
• Runtimes
• Configuration
• Guarantees
• FAQ
• Repo layout
• License

---

Why

AI coding agents are powerful but undisciplined: they skip planning, self-congratulate instead of self-critique, and drift from the spec. Frameworks that fix this usually do it in **prose** — long prompts that can't be checked, visualized, or executed.

![The same task without rensei (one long chat: no plan, no tests, spec drift found in