Claudex Loop
Skill that hardens a plan with two rival models before any code exists. Claude locks intent with you by asking the questions you skipped, then Codex attacks the locked plan round after round until it runs out of objections. Four phases: recon, interrogate, adversarial Codex review, then cross-model build and inspection. Three commands: `/claudex-loop:claudex-loop` for the full loop, `/claudex-loop:codex-review` for a review pass alone, `/claudex-loop:codex-build` to chain into execution. The plugin has been renamed twice: Grill Me Codex, then Crucible, now Claudex Loop. Two settings matter in use: plan review stops after five rounds by default (the `rounds` parameter), and `builder=codex` hands the build to Codex while Claude inspects, which also lets you offload a task onto the other subscription.
Strengths
- The critique comes from a model that did not write the plan, so it has nothing to defend
- The loop runs until objections run dry, not until you get bored
- The locking step forces the questions you skip when you are in a hurry
Limitations
- Assumes two paid subscriptions, otherwise the cost outweighs the service
- On a short task, the adversarial pass costs more time than the plan itself
Best for
- A tech lead framing a migration who wants to know what they missed before committing the team
- Someone working solo with nobody to submit their plan to
- A PO writing a spec who wants it attacked before sending it to the team