Who Watches the Watchmen: Using Codex to Adversarially Review Claude Code's Plans

Chase walks through a Claude Code skill he built called Grill Me Codex. It extends Matt Pocock's Grill Me skill, which he describes as plan mode on steroids: it interrogates you about what you want to build so you and Claude Code end up on the same page before any code gets written. His problem with the original was that the plan still comes from a single model. If you are not technical, you have no way to judge whether the plan Claude produced is actually good.

His fix is to bring in a second model. Grill Me Codex keeps the original planning act and adds a second act where OpenAI Codex reviews the locked plan and debates it with Claude Code over multiple rounds until both models sign off.

The gap in single-model planning

The video opens on the GitHub README for the skill, with the tagline "Two AI models harden your plan before you write a line of code." Chase credits Matt Pocock's original Grill Me skill, and one frame shows Pocock's "Skills For Real Engineers" page, where Pocock positions his agent skills as small, adaptable, and composable, in contrast to heavier approaches like GSD, BMAD, and Spec-Kit.

Chase's argument: skills like Grill Me, GSD, and superpowers align you and Claude Code on intent, but they leave a second problem unsolved. Who watches the Watchmen? You cannot evaluate the plan yourself if you are not an expert, and Claude Code is weak at evaluating its own output. He points out that Anthropic has acknowledged this. Ask Claude Code what it thinks of code it just wrote and it grades itself an A+. The README makes the same point about planning: a model reviewing its own plan is an echo chamber, and a different provider catches what Claude structurally misses.

How the two-act skill works

Act 1 is the original Grill Me flow: Claude Code grills you until the plan is locked. Act 2 is the addition. Claude writes the locked plan to PLAN.md and starts a log at PLAN-REVIEW-LOG.md. Codex then reviews the plan in a read-only sandbox and returns a verdict. Claude revises, the same Codex session resumes so it remembers prior objections and checks whether they were addressed, and the two models go back and forth. The loop is bounded by MAX_ROUNDS (default 5) and terminates when Codex returns "approved." You gate the process twice: once at kickoff and once at final sign-off before any code is written. Codex stays read-only the whole time.

The output is two artifacts: PLAN.md, the clean final plan (the what), and PLAN-REVIEW-LOG.md, the debate record (the why). Installation is a copy into your Claude Code skills directory:

# macOS / Linux
cp -r skills/* ~/.claude/skills/

Chase also notes the design is flexible. Swapping Codex for a local model is straightforward if you want to keep costs down.

Key Takeaways

• Planning skills like Grill Me align you and Claude Code on intent, but a single model still grades its own plan, which is an echo chamber.
• Claude Code rates its own work generously; Anthropic has called out this self-evaluation weakness.
• Grill Me Codex adds a second act where OpenAI Codex adversarially reviews the locked plan in a read-only sandbox and debates Claude Code for up to 5 rounds until it returns "approved."
• The process produces two artifacts: PLAN.md (the final plan) and PLAN-REVIEW-LOG.md (the debate record), with human gates at kickoff and final sign-off.
• The reviewer slot is swappable; you can substitute a local model for Codex to cut costs.

Resources

• Grill Me Codex: Chase's skill, shown as a GitHub repo (grill-me-codex), MIT licensed, built on Matt Pocock's grill-me and grill-with-docs skills
• Matt Pocock's "AI Skills for Real Engineers" skill collection, the source of the original Grill Me skill
• Claude Code and OpenAI Codex, the two models paired in the review loop