Your AI Architect's Rival Grades Its Homework

What it does

Crucible is a Claude Code skill that treats software planning as an adversarial process. Claude scouts your codebase or research domain, interviews you through a decision map, and locks a PLAN.md. Then OpenAI Codex — running in a read-only sandbox — reviews that plan round after round, hunting for unbuildable subsystems, missing data sources, and design flaws that would corrupt data. Only after Codex exhausts its critiques and you sign off does the optional build phase begin, with the roles flipped: Codex writes the code while Claude inspects the diff like a skeptical contributor.

The interesting bit

The workflow exploits provider rivalry to escape echo chambers. Because Claude and Codex come from different labs, they don’t share the same blind spots; the README notes that in one greenfield CRM run, Codex found one fatal unbuildable architecture, six data-corrupting designs, and seven missing subsystems that Claude’s deeply-researched plan had missed entirely. The product decisions survived untouched — the review only attacked how the plan would break.

Key highlights

• Four-phase workflow: reconnaissance, interrogation, adversarial review, and optional cross-model build
• Codex remains read-only during review; it cannot touch files until the plan is approved
• Every review round produces a PLAN-REVIEW-LOG.md transcript, so the argument is auditable
• Bounded rounds with a hard cap — a flagged deadlock beats a fake “approved”
• Ships as both a Claude Code plugin and bare skills; formerly grill-me-codex

Caveats

• You need both Claude Code and the OpenAI Codex CLI authenticated; it’s a two-vendor workflow by design
• The optional build phase hands Codex full write access, and the README warns that resuming requires bypassing approval flags manually — safety is strictly opt-in
• The published “receipts” come from a single solo-creator CRM project; there’s no broader benchmark or multi-project validation

Verdict

Worth a look if you already live in Claude Code and want a structured, adversarial safety net before expensive implementation mistakes. Skip it if you’re hoping for a one-button, single-model solution — the whole point is the friction between two rivals.