For a long stretch, every build I worked on started in Claude. I’d talk through the approach there, the order of the work, the parts most likely to break, and only then move to Codex to actually write it. That split held up because newer models had gotten much better at the planning part, but the thinking and the building still lived in different tabs. Then Codex added Plan mode, the thinking moved into the project itself, and the Claude step fell away.
I’ve paid for Claude, Gemini, and ChatGPT for months — this is the one I recommend
Three subscriptions later, I keep opening the same one.
Claude did the thinking, and Codex did the typing
My builds always started in one tab and finished in another
The division of labor felt natural. I’d open Claude Code, which had become my default for working through a build, and reason out what to do first, which files to touch, and where the whole thing might fall apart. Once that was settled, I took the plan to Codex, and it turned it into working code.
It worked because the two tools were good at different jobs, and I trusted the thinking to happen before a single line got written. For a while, that was enough.
The catch took me a while to admit. I was describing my project to Claude from memory, including the folder layout, the naming, and which files actually mattered. Then I’d paste the result into a tool that couldn’t see any of it, and Codex would work out half that context again on its own.
So my setup in the other tab was doing less than I thought. Every handoff lost a little, and I was the one carrying it across.
Plan mode reads the project before it touches a thing
Codex didn’t get a bolted-on planner. It just learned to read before it writes. You switch it on from the plus button in the composer, where Plan mode sits right under Files and folders, and from then on, Codex gathers context and proposes a plan instead of jumping straight to edits.
It stays read-only until you sign off, and you control how far it can reach with the approval setting next to it. “Ask for approval” keeps it checking in before it edits files or touches the internet, “Approve for me” only stops for anything risky, and “Full access” takes the guardrails off.
To see the loop end to end, I’ll start with something small, such as:
Add input validation to the signup form.
It reads the relevant files, lays out a short plan, and waits. I approve, and it builds. Nothing changes on disk while I’m still deciding, which is the whole point.
- OS
-
Windows, macOS, Linux
- Developer
-
OpenAI
- Price model
-
Free, paid plans available
A vague prompt gets questions, not a wrong guess
I tell it to challenge me, and it actually does
The part that sold me is what happens when I don’t have the full picture. If a request is fuzzy, Codex questions it before committing to anything. When I know the direction but not the details, I’ll say so directly:
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 interviews me. It asked whether the cache should live in memory or survive a restart, how long entries should last, which calls were worth caching at all, and how I wanted stale data cleared out. Those are the exact questions I used to answer over in my Claude planning chat.
Now they happen where the code lives, so the gaps I’d usually leave unsaid get caught before they turn into wrong code. Handing the same brief to different models can land you in very different places, and a lot of that gap comes down to what each one bothers to ask first.
More reasoning turns a real task into a careful plan
Plan mode also lets me set how hard Codex should think before it answers. The reasoning control sits next to the model picker, with Low, Medium, High, and Extra High to choose from on GPT-5.5. A quick edit gets Low. Anything with real moving parts gets Extra High, where it spends noticeably longer working on the problem before it proposes a thing.
I leaned on that for a heavier job.
Migrate the auth flow from session cookies to JWTs, and plan it before changing anything.
On Extra High, the plan came back grounded in my actual files, naming the middleware it would rewrite, the routes that needed new checks, and the tests it expected to update. It read like something I’d have written myself, only faster.
I’ve watched Codex map out a real build this way before writing a line, and the plan is usually close enough that approving it is the easy part. Then it runs.
I use ChatGPT, Claude, and Gemini every day, but never for the same job
After months of trying to cut it down to one, the three-way split is still standing.
Where I want to take plan mode next
Saved plan templates and longer runs are the next test
What I’m curious about now is making the plans reusable. For the work I repeat, a saved plan template would mean Codex starts from a known shape instead of building one from scratch each time. From there, I want to push Plan mode into longer runs, the kind that span several steps and a chunk of an afternoon, and wire the approved plans into automations so the routine parts kick off on their own.
