Claude and Codex: Two AIs, One Workflow
Here's something most people don't think about: why use one AI when you can make two of them check each other's work?
I've been running Claude Code as my primary agent — it builds, it plans, it architects. But I recently added OpenAI's Codex CLI as a second opinion layer that sits right inside the same workflow. Claude does the work. Codex reviews it. Two different models, two different perspectives, and I'm the one who decides what ships.
Why Not Just Use One?
Every model has blind spots. Claude is incredible at generating code and creative work, but it can be too agreeable sometimes. It'll build exactly what you asked for without pushing back on whether you should be building it that way in the first place.
Codex doesn't care about your feelings. When I tell it to review something, it comes back with things like "this claim is unsourced," or "you're blurring a native feature with an ecosystem convention." It's the hard ass in the back of the room that keeps the work honest.
How It Actually Works
The setup is surprisingly simple. I installed three skills into Claude Code's global config:
/codex — the core skill. It lets Claude invoke Codex CLI
directly using codex exec. The cool part is that Codex adapts its role based
on what I'm doing. If I'm writing code, it becomes a strict code reviewer. If I'm making
a Remotion video, it becomes a creative director critiquing pacing and transitions. If
I'm writing a blog post, it becomes an editor.
/codex-review — adversarial code review from five perspectives:
security, correctness, compliance, performance, and maintainability. It also runs automated
checks like TypeScript compilation, ESLint, and secret scanning.
/codex-plan-review — when Claude makes a plan, I can ask Codex
to tear it apart before I approve it. It finds blind spots, missing steps, and assumptions
I didn't catch.
The best part is that Claude decides when to use Codex on its own. For complex plans, it consults Codex behind the scenes and incorporates the feedback before I even see the plan. For simple stuff, it skips Codex entirely. No wasted tokens, no micromanaging — Claude just knows it has a teammate available when the work gets hard.
A Real Example
Yesterday I wrote a blog post about building Claude's persistent brain. Before publishing, I said "get Codex's take on this." It came back with ten bullet points of criticism — flagged missing prerequisites in my setup prompt, called out unsourced claims, pointed out that I was conflating native Claude features with third-party ecosystem conventions, and even suggested the giant prompt block should be moved to a collapsible section.
Some of the feedback I agreed with. Some of it was Codex being overly strict for a casual blog. But the point is that I got a genuine second perspective from a completely different model, without leaving my terminal.
The Bigger Picture
We're heading toward a world where the best workflows won't be one AI doing everything. They'll be multiple models playing different roles — one builds, one reviews, one fact-checks — with a human making the final call. It's not that different from how good teams already work. You don't ship code without a review. You don't publish without an editor. Why would your AI workflow be any different?
The dual-model setup costs almost nothing extra on the Claude side since Codex only runs when you invoke it, and its output gets summarized before entering Claude's context. On the Codex side, the usage is generous enough that you can lean into quality without worrying about limits.
If you're already paying for both subscriptions, you might as well make them talk to each other.
The Setup Prompt
Make sure you have Codex CLI installed first (npm install -g @openai/codex) and
authenticated. Then open Claude Code at your home directory and paste this:
"I want you to set up Codex CLI as your always-available collaborator inside Claude Code. Codex is your teammate — use it freely whenever you need a second perspective. Here's exactly what to do: 1. INSTALL THREE CODEX SKILLS into ~/.claude/skills/: a) Core Codex CLI access: git clone --depth 1 https://github.com/skills-directory/skill-codex.git /tmp/skill-codex-temp cp -r /tmp/skill-codex-temp/plugins/skill-codex/skills/codex ~/.claude/skills/codex rm -rf /tmp/skill-codex-temp b) Plan review skill: git clone --depth 1 https://github.com/cathrynlavery/codex-skill.git /tmp/codex-hook-temp mkdir -p ~/.claude/skills/codex-plan-review cp /tmp/codex-hook-temp/skills/codex/SKILL.md ~/.claude/skills/codex-plan-review/SKILL.md rm -rf /tmp/codex-hook-temp c) Adversarial code review: mkdir -p ~/.claude/skills/codex-review curl -sL https://raw.githubusercontent.com/pauhu/claude-codex-review/main/skills/codex-review/SKILL.md -o ~/.claude/skills/codex-review/SKILL.md 2. ADD THIS SECTION to your ~/.claude/CLAUDE.md (create the file if it doesn't exist): ## Codex (Your Collaborator) - Codex CLI (codex exec) is always available. Use it freely whenever you need a second perspective. - You decide when to call Codex — don't wait for the user to ask. Use your judgment: - Complex plans → get Codex's review before showing the user. Incorporate feedback, present the refined plan. - Simple/trivial plans → skip Codex, just show the plan directly. - Code reviews, architecture decisions, unfamiliar APIs → consult Codex. - Creative work, content, research → Codex can help as editor, fact-checker, or critic. - Codex adapts role to context: code → reviewer, video → creative director, content → editor, research → fact-checker, design → UX critic. - Always use the latest model: run codex exec -m gpt-5.4 (current latest). - Tell Codex to be concise. Summarize its output for the user — don't dump raw responses. - Codex usage is unconstrained. Only Claude token efficiency matters. 3. RUN A SECURITY AUDIT on all three skill files to make sure nothing is malicious. 4. TEST the setup by running: codex exec -m gpt-5.4 --sandbox read-only 'Respond with one sentence confirming you are Codex CLI and the connection works.' After setup, Claude should freely consult Codex whenever it needs help — reviewing complex plans before showing them, getting second opinions on architecture, fact-checking research, and critiquing creative work. No need to ask Claude to use Codex — it should use its own judgment."
That's it. Three skills, a few lines in your config, and you've got two AIs keeping each other honest.