Dual Agent Workflow — Productivity skill for Claude Code

dual-agent-workflow

A reusable **cross-vendor coding-agent policy**: two AI coding tools from two different vendors — **Claude Code** and **Codex** — take turns implementing and reviewing each other's work. Whichever agent writes the code, the *other one* reviews it. Never itself.

That single rule (cross-vendor review) is the point. An AI reviewing its own work is anchored to its own choices and shares its own blind spots; a different vendor is independent by construction and fails differently, so it catches what a self-review never would. Everything else in these templates — the work-type routing, the staged pipeline, the merge gates — exists to serve that rule and keep the process honest.

This repo is a **template**, not a library. You copy one file (`AGENTS.md`) into your own repo, fill in a handful of project-specific placeholders, and both agents pick up the policy automatically — Codex and Claude Code (v2.1.277+) both read `AGENTS.md` natively.

What's in here

File	What it is
AGENTS.template.md	The canonical shared policy. Copy to AGENTS.md in your repo. Both agents read it natively. This is the file you edit.
CLAUDE.template.md	Optional. A thin pointer that imports AGENTS.md (@AGENTS.md). Only needed if your repo already has a CLAUDE.md (Claude Code then reads it instead of AGENTS.md) or someone runs Claude Code older than v2.1.277.
SETUP_PROMPT.md	Hand this to a Claude/Codex session running inside your target repo and it will do the setup for you — inspect the repo, fill placeholders, wire the files, and report back.
GUIDE.md	A from-scratch explainer for people new to the pattern: why it's shaped this way, the roles, the routing table, and two day-in-the-life walkthroughs. Start here if it's your first time.

Quick start

**Option A — let an agent do it.** Open your target repo in Claude Code, attach `AGENTS.template.md`, `CLAUDE.template.md`, and `SETUP_PROMPT.md`, a