The same loop, the same habit

When you press enter, Codex runs the agent loop. You know this loop. Look, plan, act, check, repeat. Here it is, in your practice folder:

1. Codex looks. It checks the folder. It's empty, so there's little to read this time, but on a real project this is where it reads your files.
2. It plans the next concrete step: create toolkit.txt with the two lines.
3. At the act step, because creating a file changes your machine, it stops and shows you the action before doing it, then waits for you to approve.
4. You read the proposed action. It's what you asked for. You approve.
5. The file gets created. Codex checks that it worked, sees it did, and stops, since the job is done.

Read the line. Still your job.

The wording of Codex's approval prompt is a little different from the Claude CLI's. The job it does is identical: it shows you what's about to happen and waits for your yes.

So the habit is identical too. Read the proposed action before you approve it. Not because this first action is risky, it isn't, but because you're keeping the muscle in shape for when an action is risky, on a real project, where reading the line is the only thing between "not what I meant" and "done anyway." Same muscle as rm in the terminal chapter, same muscle as the Claude CLI. One habit, every tool.

Check it yourself

When Codex says it's done, verify it. Leave the session with /exit, then in the plain terminal:

cat toolkit.txt

There are your two lines. You instructed a second AI tool, in plain English, and it did the work. Your toolkit now has two tools in it, and they work the same way.