A skill for Claude Code and Codex · Method and business
cross-review is the skill that sends the opposite model to review whatever was just built: Codex reviews Claude Code and Claude Code reviews Codex, with a closed verdict, evidence run by the reviewer itself, and no more than two rounds. The reviewer only reads and tests, it never edits anything, and approval requires at least one test case the author never listed.
Solo skill: unlocks right away, no guarantee after download. Purchase terms.
Or get all 25 skills plus 9 bonuses in the Supreme Pack for R$ 497.
The problem
A review done by the same model that wrote the code inherits the exact same blind spot, and a review with no rules turns into an endless back and forth or an approval out of sheer fatigue. Three symptoms, and what changes with the skill.
| Without the skill | With cross-review |
|---|---|
| You ask Claude to review what it just implemented, it reads its own summary, agrees with itself, approves it, and the authorization bug ships to production. | The reviewer is always the model opposite the author, runs in read-only mode, and judges what actually runs, not the implementer's account of it. |
| The review approves by running only the tests the author listed, so it repeats the exact same blind spot as whoever wrote the code. | Approval requires the command and output attached, plus at least one case outside the author's list. Blocking requires reproduction, with input, command and observed output. |
| A review with no end: every round raises a new issue, the fix never closes, and nobody knows if the code is actually ready to ship. | Three closed verdicts, and no more than two rounds. The second round only checks whether the original blockers got fixed, with no new fronts opened. |
What you get out of it
A verdict you can actually use as a gate, not just a loose opinion.
approved, approved with caveats, or blocked, no middle ground
every blocker comes with the input, command and output that reproduce the defect
what needs to pass, what exists to prove the bug, and what's just exploration
rounds, verdicts and scope boundaries logged so you can resume or audit later
How it works
The diff, the files and the context of what changed, with zero secrets in the material that goes to the other model.
If Claude Code wrote it, Codex reviews it, and the reverse holds too. The reviewer runs in read-only mode and can't edit or delegate.
The reviewer runs the contracted tests plus its own, then decides between approved, approved with caveats, or blocked.
If it's blocked, the author fixes it and round 2 validates only the original blockers and whatever the fix might have broken.
Verdict and state saved. A timeout, a failed tool, or a test that couldn't be run never turns into an approval.
Before you buy
Each review draws on the subscription quota of whichever model is reviewing.
Its limits
The reviewer inspects, tests and decides, and the fix stays with whoever implemented it. That's a hard rule: a reviewer that edits stops being independent.
A good review depends on a testable target. When the environment blocks a test from running, the verdict states that limitation instead of quietly approving anyway.
To orchestrate the whole project, with classified risk, executors and gates, enterprise-agentic-workflow, also in the catalog, calls cross-review on every delivery.
What you receive
The price
Just cross-review
R$97one time payment
Bought it solo and now want it all? Message me on WhatsApp and I will deduct what you already paid.
Supreme Pack
R$ 1.665R$497or 10x on a card
Buying all 25 solo adds up to R$ 1.665. The pack pays for itself by the ninth skill. See everything the Supreme Pack includes
05 · Method and business
Before you ask
Whoever didn't write the code reviews it. If Claude Code implemented it, Codex gets the diff and context in read-only mode, runs the tests, looks for a case the author didn't cover, and returns a closed verdict. The reverse works the same way.
You need both installed and accessible, because the review always runs on the model opposite the one that wrote the code. That's exactly what removes the blind spot of a model reviewing itself.
No more than two in normal use. The first is complete, and the second only checks whether the blockers got fixed. In lite mode, there's just a single round.
After touching login, credentials, payments, authorization, concurrency, a migration or a destructive operation, when the same fix has failed twice, or any time you want the other model's opinion before shipping.
Yes, the license covers your work for third parties. What's off limits is reselling or redistributing the skill.
One-time payment, 12 months of updates. No refunds after download; the Supreme Pack comes with a 7-day guarantee.
