cc-codex-dev

CC orchestrate, Codex implement, CC verify — A Claude Code skill for cross-model AI collaborative development.

What is this?

A Claude Code slash command (/codex-dev) that delegates coding tasks to OpenAI Codex while Claude Code handles scoping, review, and verification.

The core idea: Split work by context needs, not by role. Cross-model review catches blindspots that same-model review misses.

Design Principles

This skill was designed based on guidance from Anthropic and Karpathy:

1. Split by context, not role — Agents don't need "personas". They need the right files in their context window. (Anthropic: How we built our multi-agent research system)

2. Start simple, evolve with data — Single Codex call is the default. No parallel execution in V1. Add complexity only when real usage data proves it's needed. (Anthropic: Building Effective Agents)

3. Verification > Orchestration — Hard gates (build/test) run before LLM review. Deterministic checks are more reliable than LLM judgment. (Karpathy: March of Nines — a 10-step workflow at 90% per step = 35% overall reliability)

4. Fewer steps, harder checks — Every handoff compounds failure. Keep the chain short: scope → delegate → verify → report.

5. Abort on scope creep — If Codex needs files outside the allowed list, new dependencies, or schema changes, it stops immediately.

Workflow

/codex-dev "implement feature X"

Step 1: CC scopes (reads code, identifies files, defines contract)
Step 2: CC delegates to Codex (with task + files + constraints + stop conditions)
Step 3: Codex implements
Step 4: CC verifies (build/test first, then LLM review of diff)
Step 5: Report (PASS/FAIL + next steps)

Installation

Option A: Copy to your commands directory

mkdir -p ~/.claude/commands
cp codex-dev/SKILL.md ~/.claude/commands/codex-dev.md

Option B: Clone and symlink

git clone https://github.com/Lntanohuang/cc-codex-dev.git ~/.claude/skills/cc-codex-dev
ln -s ~/.claude/skills/cc-codex-dev/codex-dev/SKILL.md ~/.claude/commands/codex-dev.md

Prerequisites

• Claude Code installed
• Codex CLI installed and authenticated (codex --version)

Usage

In Claude Code, type:

/codex-dev "add a statistics page showing student accuracy and weak points"

Or just say: "let codex write this" — CC will auto-trigger the skill.

When to use

• Well-scoped implementation tasks with clear file boundaries
• Tasks that benefit from a second model's perspective
• When you want CC to focus on architecture decisions while Codex handles implementation

When NOT to use

• Trivial changes (rename, typo) — just do it directly
• Vague/unscoped tasks — scope first, delegate after
• Pure exploration — use CC's built-in Explore subagent

References

• Anthropic — Building Effective AI Agents (PDF)
• Anthropic — How we built our multi-agent research system
• Anthropic — Claude Code Subagent Docs
• Karpathy — March of Nines / Agentic Engineering

License

MIT