Meet Codex, the Contractor Who Works While You Sleep

I gave Codex a task before dinner.

By dessert, there were changes waiting for me to review.

I hadn't even finished my dal bhat.

The contractor with your house keys

When you renovate a room, you don't stand next to the contractor all day.

You agree on the job.

"Replace these tiles. Fix the leaking tap. Don't touch the kitchen."

You hand over the keys.

You go to work.

In the evening, you come back and inspect.

If something is wrong, you point at it.

That's how I think about OpenAI Codex.

It can take a coding task, work on it in its own environment, and come back with proposed changes for you to review.

Where it shines

Codex works best with tasks that are clear and bounded:

• "Add tests for this module."
• "Fix this specific bug, here's the error."
• "Update this dependency and fix what breaks."
• "Refactor this file to use the new helper."

These are the "replace these tiles" jobs.

Where it struggles

It's less useful for:

• "Make the app better."
• "Redesign our architecture."

No contractor can "make the house better" without a plan.

A good work order

Bad:
Fix the login.
 
Good:
Users with emails containing a "+" can't log in.
The validation is in src/auth/validate.ts.
Fix the regex, add tests for "+" and uppercase
emails, and don't change the API response format.

Notice the "don't change" part.

That's the "don't touch the kitchen" instruction.

Always inspect the work

Treat what Codex gives you exactly like a teammate's pull request.

Read the diff.

Run the tests.

Ask "why did it change this file?"

The contractor did the work.

You still own the house.

Final thoughts

The magic isn't that Codex codes while you eat.

The magic is that you learned to write a clear work order.

Do that, and you come back to a finished job.

Skip it, and you come back to new tiles in the kitchen.