Tutorial: Claude Code Adversarial Plan Review with Codex CLI

Adversarial Plan Review with /grill-me-codex and OpenAI Codex CLI

Every planning skill for Claude Code runs into the same ceiling: you’re asking one model to both write a plan and grade it. This tutorial shows you how to extend Matt Pocock’s /grill-me with an adversarial, multi-round review powered by Codex CLI — so two independent AI tools sign off before you write a line of implementation code. You’ll end the process with a hardened plan.md that has survived iterative critique, not just the first draft that felt right.

1. Grab the grill-me-codex and grill-with-docs-codex skill files from the pinned comment under the original YouTube video. These extend Pocock’s existing skills without replacing them — both sets coexist in your skills directory.

2. Drop both files into your Claude Code skills directory alongside Pocock’s originals. The grill-me-codex skill is structured in two acts: Act 1 runs the same deep planning questions as the base skill; Act 2 hands the resulting plan to Codex for adversarial review.

3. Write a prompt describing the feature you want to build — include constraints, relevant existing infrastructure, and acceptance criteria — then invoke /grill-me-codex. The skill reads your codebase before generating questions tailored to your context.

4. Work through the planning questions Claude Code surfaces. The demo runs ten questions; each presents two or three labeled options alongside a recommendation with explicit rationale. You can accept a recommendation outright by typing do your recommendation, or specify a different option — the skill handles both paths cleanly.

5. Once all questions are resolved, Claude Code writes plan.md — the single source of truth for the build — and initializes plan-review-log.md, which will record every exchange between Claude and Codex across all review rounds.

6. Codex CLI launches headlessly with a shared session ID so it retains memory across rounds. It reads plan.md and writes Round 1 findings to plan-review-log.md. In the demo, it surfaces 11 issues — security holes, correctness gaps, and unbounded inputs. Claude Code acts as arbiter: it absorbs valid findings, rejects weak ones, and updates plan.md before returning control to Codex.

Warning: this step may differ from current official documentation — see the verified version below.

7. Codex reviews the updated plan.md in Round 2, returning four findings — including false fixes from Round 1 that Claude Code claimed to address but never wired correctly. The plan is updated again.

8. The loop continues until Codex issues an APPROVED verdict or five rounds complete. In the demo, Round 3 produces approval with three low-severity non-blockers that don’t halt implementation.

9. Before writing a single line of implementation code, read through the final plan.md and the open items at the bottom of plan-review-log.md. The log captures every issue raised and every fix applied across all rounds — it doubles as a debugging reference if something breaks later.

10. To swap Codex for a local or cheaper model, open the skill file and update the model invocation line. The iterative loop logic is model-agnostic.

How does this compare to the official docs?

The skill stitches together Claude Code’s planning layer and Codex CLI’s review capabilities in a way neither tool documents as a first-party integration — which makes it worth examining what each tool’s official documentation actually says about headless invocation, session continuity, and multi-round review before you build a production workflow around this pattern.

Here’s What the Official Docs Show

Act 1 gives you a solid working model for the /grill-me-codex workflow — this act layers in the verified prerequisites and version specifics that keep it reproducible across installs. Two official sources cover the available ground: Claude Code’s product pages and the openai/codex GitHub README.

Step 1 — Grab the skill files from the pinned comment

No official documentation was found for this step — proceed using the video’s approach and verify independently.

Step 2 — Drop both files into your Claude Code skills directory

The video’s approach here matches the current docs exactly. One prerequisite the video skips: Claude Code requires an authenticated Anthropic account before any skill can be installed or invoked. It’s also worth flagging before you build a multi-round loop — the Free tier may hit context limits mid-review. As of June 2026, Pro runs $17–$20/month and Max starts at $100/month; extended agentic sessions lean toward the latter.

Step 3 — Write your feature prompt and invoke /grill-me-codex

No official documentation was found for this step — proceed using the video’s approach and verify independently.

Step 4 — Work through the planning questions Claude Code surfaces

No official documentation was found for this step — proceed using the video’s approach and verify independently.

Step 5 — plan.md is written; plan-review-log.md is initialized

No official documentation was found for this step — proceed using the video’s approach and verify independently.

Step 6 — Codex CLI launches headlessly; Round 1 findings written to the log

The video’s approach here matches the current docs exactly. The openai/codex README explicitly states Codex CLI “runs locally on your computer,” and the repository structure confirms headless terminal invocation is the intended use pattern. One reproducibility gap: the tutorial does not specify a tested version. The current release is v0.137.0 — pin to this when following Steps 6–8.

Step 7 — Round 2 review surfaces false fixes from Round 1

No official documentation was found for this step — proceed using the video’s approach and verify independently.

The openai/codex repo does include an AGENTS.md at root level — worth consulting for how Codex CLI interprets task instructions and structures its outputs across rounds.

Step 8 — Loop runs to APPROVED or five-round cap

No official documentation was found for this step — proceed using the video’s approach and verify independently.

Step 9 — Read plan.md and open items in plan-review-log.md before implementation

No official documentation was found for this step — proceed using the video’s approach and verify independently.

Step 10 — Swap the model by editing the skill file’s invocation line

No official documentation was found for this step — proceed using the video’s approach and verify independently.

The Codex CLI README does confirm model selection is a real, configurable parameter — the terminal UI screenshot shows gpt-5.2-codex medium as the active model. The specific skill-file editing method the video demonstrates is a custom integration layer; no official source documents it directly.

Useful Links

1. Sign in – Claude — Claude Code product entry point covering authentication requirements and current subscription pricing across Free, Pro, and Max tiers.
2. GitHub – openai/codex: Lightweight coding agent that runs in your terminal — Official openai/codex repository with README, installation guide, AGENTS.md, and release history for Codex CLI v0.137.0.