grill-me-codex addresses critical gaps in AI-assisted coding by ensuring user-AI plan alignment and validating output quality. It targets developers seeking robust AI-driven planning, offering a novel adversarial review process to prevent echo chambers and enhance plan correctness before code generation.

How It Works

The system employs a two-act adversarial approach. Act 1 involves Claude interrogating the user to lock down a plan. Act 2 then hands this plan to OpenAI Codex, a rival model, which iteratively critiques and tears apart the plan over several rounds. Claude revises based on Codex's feedback until both models approve, ensuring a plan stress-tested by an independent AI perspective. This cross-provider review catches structural flaws Claude might overlook in its own planning.

Quick Start & Requirements

• Install: Copy skill folders into your Claude Code skills directory (~/.claude/skills/ on macOS/Linux, $env:USERPROFILE.claude\skills\ on Windows PowerShell).
• Prerequisites:
  • Codex CLI version 0.130+ (npm install -g @openai/codex@latest).
  • Authenticated Codex account (codex login; ChatGPT account sufficient).
  • Default model configuration in Codex (skills use default, not pinned variants).
• Usage: Invoke via Claude Code commands like /grill-me-codex.
• Links: Matt Pocock's original skills: https://github.com/mattpocock/skills.

Highlighted Details

• Dual-Model Adversarial Review: Leverages OpenAI Codex to critically review Claude's plans, preventing self-confirmation bias.
• Iterative Plan Hardening: The plan undergoes multiple rounds of adversarial critique and revision between Claude and Codex.
• Safety First: Codex operates in read-only sandbox mode; no code generated until user approval.
• Artifact Generation: Produces PLAN.md (final plan) and PLAN-REVIEW-LOG.md (review transcript).

Maintenance & Community

Act 1 skills originate from Matt Pocock's MIT-licensed work. Act 2 and packaging are by Chase AI. Further learning is available via the Chase AI+ Claude Code Masterclass.

Licensing & Compatibility

The core components are licensed under MIT, which is permissive for commercial use and closed-source linking.

Limitations & Caveats

Setup requires specific Codex CLI versions and authentication. The system relies on the user's default Codex model configuration, as specific variants are rejected. The adversarial loop has a default maximum of 5 rounds, terminating on approval or deadlock.