Summary: I once asked Codex to review a tool written by Claude. The value of the second opinion was another angle that still required verification. Switching models, opening a new window, using neutral wording, or selecting read-only settings cannot independently guarantee that a judgment is independent or correct. This essay retains a four-step method while separating tool output, completed review, and authorization to modify.
I used Claude to write a reusable set of review instructions, then asked Codex to examine that tool. In the original work record, the two rounds reported seven and three issues. Those are the numbers reported at the time. They do not mean that all ten were reverified in this revision, nor do they constitute a current security certification of the complete tool.
The most useful reminder was that a tool exiting normally and leaving a paragraph of text is not enough to show that it completed the review I needed.
Why Bring in a Second Role?
At the time, I used Codex CLI and the selected model to examine code, data, or designs produced by Claude. Codex is a working tool and should not be equated with one permanently fixed model. The original reference to GPT-5.5 belongs only to the account of what I used then.
A different model, role, or context may supply another checking angle, but that does not establish the absence of shared blind spots. The same model placed in a new context may also provide a useful review.
Switching models does not mean that the judgment is now independent and correct.
What I want to preserve is a process in which the reviewer first reads the evidence and reports its own findings, after which the coordinator verifies and handles them. This also belongs within my broader writing about human-AI collaboration: give different roles clear tasks rather than treating another open window as a completed review.
Begin by Defining the Target and Criteria
If I begin by saying, “I have already confirmed that there is no problem,” and then ask another role to confirm it, I may carry my assumption into its work. Neutral phrasing can reduce that kind of leading, but it cannot eliminate bias.
What I prefer to provide is the version to review, the relevant files, expected behavior, known limitations, and the conditions that count as a problem. Necessary context should be explicit; my own conclusion can be held back for later comparison.
A neutral question can reduce leading, but it cannot prove that the review was independent.
Similarly, instructions not to speculate or to report unread scope are working requirements, not evidence that those requirements were followed. That depends on what the reviewer actually read and what support it supplied for its findings.
A Four-Step Method That Preserves Evidence and Disagreement
I organize the method into four steps.
- Define the question. Provide a fixed target, relevant context, and criteria without making my own verdict a premise.
- Preserve the feedback. Retain the original findings, their locations, the reviewed version, and incomplete items so they can be checked later.
- Separate review from modification. This round identifies issues only. Whether and how to change something is decided and recorded by someone with authority to act.
- Return to the source. Check both agreements and disagreements. Two reviewers agreeing is not a reason to skip evidence verification.
Raw feedback, adoption decisions, and modification results need to be preserved separately.
These arrangements make the reasoning process more traceable, but they do not prove statistical independence. Nor do they mean that a reviewer can never help make a correction in another workflow. I separate the roles in this round so that the responsibility and basis for a modification remain clear.
Text Output, Process Completion, and Review Acceptance Are Different
The original tool used a nonblank check to confirm that a final message existed. That can catch one kind of failure: the process exited but left no text. It cannot establish that the text reviewed the correct target, covered the intended scope, or reached a reliable judgment.
The official documentation for non-interactive execution describes codex exec and ways to write the final message to a file. Saving a message can preserve the response, but the existence of that file does not prove that the review was completed.
A nonblank check proves only that some text was left behind.
Acceptance should therefore be separated into several questions: Did the process end normally? Was the complete output retrieved? Was the specified scope actually read? Do the findings have support? Which matters remain unresolved? A result with no findings also needs to be understood together with the scope that was checked.
The old appendix had no default time limit, and a wrapper process exiting would not by itself establish that every child process had stopped. This revision does not present it as a verified fast-failure mechanism or as a guarantee that billing stops at a particular moment.
“Read-Only” Is Not a Total Guarantee About Every Action
I configured the review round to read and report findings only and applied corresponding tool restrictions. But naming target files in a prompt does not automatically create a technical read allowlist. An instruction not to modify files also cannot substitute for actual permission settings.
The official documentation on approvals and security discusses sandboxing and approvals separately. Reading, writing, network access, and the capabilities of external tools need to be confirmed within the actual environment rather than collapsed into a comprehensive guarantee from a single “read-only” label.
Restricting model-generated writes does not mean that every part of the process has no side effects.
A wrapper may write an output file, and a tool may preserve work records. When a model service is used, review material may also be transmitted to that service. Starting the process locally does not mean that the data remains local. Sensitive material should be evaluated before deciding what is appropriate to provide to the service.
Whether Another Review Round Is Worthwhile Depends on the Task and Cost
I no longer describe a second opinion as uniformly inexpensive. The official authentication documentation distinguishes access through a ChatGPT subscription from usage billed through an API key. Availability, allowances, and costs still depend on the account and plan.
The comparison should also include more than the additional model usage. Preparing the target, reading findings, verifying them, making corrections, and checking again all take time. Another checking angle may be worthwhile for important work or issues that are easy to miss. A small wording change may call for a lighter process.
Whether another review round is worthwhile should include the cost of verification and correction.
Available model settings may also change. A default value from the old appendix should not be treated as a fixed answer suitable for every account.
The Appendix Is No Longer a Copy-Ready Template
The original article included a skill and shell wrapper. This revision withdraws the copyable code and retains the method, because understanding an example and establishing that it is suitable to execute are different things.
In addition to the timeout and nonblank-output limitations already described, the old example did not provide complete guarantees for isolating output filenames, detecting and handling output-write failures, or restricting paths. It should not continue to carry promises such as “these are the only writes,” “copy and run,” or a fixed model assignment.
This revision did not execute the old example, retest it against the current CLI, or repackage it as a repaired tool. A similar mechanism should be designed, inspected, and tested separately against current documentation and the requirements of the project.
The existence of a historical example does not mean it can be adopted directly today.
Listen Carefully Without Accepting Everything
When a reviewer has identified one problem accurately, its next finding deserves serious attention. That does not exempt the finding from verification. I may have missed something, and the reviewer may also be mistaken.
Agreed findings need verification, and disputed findings need verification too.
I bring each finding back to the original material and explain which ones I adopt, which I do not, and which lack enough evidence. Any change to the rules should also consider whether the risk they were intended to control still exists. A boundary should not be removed merely because a model appears more confident.
That is the multi-model collaboration method I want to preserve: a second opinion can puncture certainty, but the final decision still belongs to someone with authority to act and must remain connected to reasons that can be checked later.


💬 Comments
Loading...