Three habits worth carrying into every session
You can drive Codex now. Here are three working habits that make any terminal AI tool, Codex or Claude, go better. None of them are features to memorize. They're ways of working.
Ask it to explain before it changes anything
On a real task, before you let the tool start editing, you can say: "Don't change anything yet. First tell me your plan." You read the plan, and if it's heading the wrong way, you correct it before a single file moves. Catching a wrong direction at the plan stage costs nothing. Catching it after ten files changed costs you a cleanup.
Give it one job at a time
A vague, sprawling instruction ("fix everything wrong with this folder") produces sprawling, hard-to-check results. A specific one ("rename these three files to lowercase") produces a result you can look at and judge in seconds. When a task is big, break it into named steps and hand them over one at a time. You stay in control, and you can stop early if something looks off.
Watch the loop, don't walk away
The tool runs its loop and pauses for your approvals. That pause only protects you if you're there to read it. Letting a session run while you're not watching turns the approval prompt into a button you mash. Stay with the session while it works. These are short bursts, not hour-long jobs you leave running.
These are not Codex tricks
Everything above is just as true of the Claude CLI. That's deliberate. The point of learning a second tool is partly to see that the skill of directing an AI tool sits above any one product. Explain-before- acting, one-job-at-a-time, watch-the-loop: that's the craft, and it travels with you to every tool you'll ever use.
Three habits worth carrying into every session
You can drive Codex now. Here are three working habits that make any terminal AI tool, Codex or Claude, go better. None of them are features to memorize. They're ways of working.
Ask it to explain before it changes anything
On a real task, before you let the tool start editing, you can say: "Don't change anything yet. First tell me your plan." You read the plan, and if it's heading the wrong way, you correct it before a single file moves. Catching a wrong direction at the plan stage costs nothing. Catching it after ten files changed costs you a cleanup.
Give it one job at a time
A vague, sprawling instruction ("fix everything wrong with this folder") produces sprawling, hard-to-check results. A specific one ("rename these three files to lowercase") produces a result you can look at and judge in seconds. When a task is big, break it into named steps and hand them over one at a time. You stay in control, and you can stop early if something looks off.
Watch the loop, don't walk away
The tool runs its loop and pauses for your approvals. That pause only protects you if you're there to read it. Letting a session run while you're not watching turns the approval prompt into a button you mash. Stay with the session while it works. These are short bursts, not hour-long jobs you leave running.
These are not Codex tricks
Everything above is just as true of the Claude CLI. That's deliberate. The point of learning a second tool is partly to see that the skill of directing an AI tool sits above any one product. Explain-before- acting, one-job-at-a-time, watch-the-loop: that's the craft, and it travels with you to every tool you'll ever use.