What are the advantages of the Codex's Plan Mode?

That division was fine because newer models were much better at planning, but thinking and building were still in different tabs. Then the Codex added Plan Mode, the thinking was moved into the project itself, and the step of using Claude was eliminated.

Claude handled the conceptualization, Codex handled the creative work.

The division of labor happened quite naturally. Everyone would open Claude Code, the application that had become the default for executing a construction project, and reason about what to do first, which files to edit, and where the entire project might encounter problems. Once everything was in order, they transferred the plan to the Codex, and it was transformed into working code.

This approach works because these two tools are good at doing different things, and people are in the habit of thinking before writing any code. For a while, that was enough.

It took a while for people to realize the problem. They described their project to Claude from memory, including the folder layout, naming conventions, and which files were actually important. Then, they pasted the results into a tool that couldn't see any of that information, and the Codex would automatically reprocess half of that context.

Therefore, the settings in the other tab worked less efficiently than they thought. Each transfer took a little bit of time, and they were the ones who suffered the consequences.

Plan Mode reads the project before making any changes.

The Codex doesn't come with a built-in planning function. It only learns to read before writing. You enable this mode from the plus button in the editor, where Plan Mode is located just below Files and Folders, and from there, the Codex gathers context and suggests a plan instead of jumping straight to editing.

It will be read-only until you confirm, and you control its permissions using the approval settings next to it. "Ask for approval" will require prior approval before editing files or accessing the internet, "Approve for me" will only stop anything that poses a risk, and "Full access" will remove all restrictions.

To see the entire loop, let's start with a small example, such as:

Add input validation to the signup form.

It reads the relevant files, generates a brief plan, and waits. Once you approve it, it begins building. Nothing changes on the drive while you decide, which is the purpose of this.

A vague prompt will lead to more questions, not a wrong guess.

What leads many people to use the Codex's Plan Mode is what happens when they don't have the whole picture. If a requirement is unclear, the Codex will question it before making any commitments. When you know the direction but not the details, be upfront about it:

I want to add caching, but I haven't worked out the details. Challenge my assumptions and ask whatever you need before you plan it.

Then it will interview you. It asks whether the cache should be stored in memory or persist after a reboot, how long entries should remain in the cache, which calls are worth caching, and how you want to clear old data. These are exactly the same questions you answered in your planning conversation with Claude.

Now, these issues occur right where the code is executed, so vulnerabilities you might otherwise overlook are discovered before they become flawed code. Giving the same summary to different models can lead to very different results, and much of that difference depends on whether each model bothers to ask certain questions beforehand.

More careful reasoning transforms a practical task into a meticulous plan.

Plan Mode also allows you to set the level of inference the Codex should perform before making a decision. The inference control is located next to the model selector, with Low, Medium, High, and Extra High levels to choose from on GPT-5.5. A quick edit would be selected at Low. Anything with actual moving components would be selected at Extra High, where it spends significantly more time processing the problem before making a suggestion.

Use that as leverage for a more challenging task.

Migrate the auth flow from session cookies to JWTs, and plan it before changing anything.

At the Extra High level, the plan is based on actual files, specifying the middleware it will rewrite, the new routes it needs to check, and the tests it anticipates updating. It's like writing it yourself, only faster.

The Codex also maps out a real build in this way before writing the code, and the plan is usually accurate enough that approval is very easy.

How should Plan Mode be further developed?

What everyone is curious about now is how to make plans reusable. For repetitive tasks, a saved plan template would mean the Codex starts from a known shape instead of building from scratch each time. From there, users can push Plan Mode to longer runs, extending the number of steps and connecting approved plans to automation processes so that frequently occurring components automatically start.