Most of my building happens here now rather than in an editor, which is not a sentence I expected to write.
The thing that actually changed my workflow was not code generation. It was that the unit of work got bigger. I stopped asking for functions and started handing over outcomes, and the useful skill turned out to be decomposition rather than prompting. Say what done looks like, say what must not break, then get out of the way.
Two things I would tell someone starting:
Ask it how it would do the thing before you ask it to do the thing. The plan is cheap to read and expensive to skip. Most of my bad runs were ones where I skipped straight to execution.
Give it a memory. Out of the box every session starts from zero and you re-explain the same context forever. A docs folder it reads on every run was the single biggest improvement I made, and it cost nothing.
Where it bites: long sessions quietly compact and you lose reasoning you did not know you needed. And it is confidently wrong in exactly the register it is confidently right, so the review burden never really goes away, it just moves.
Still, it replaced the thing I used to call my job.
Codex 3.0 by OpenAI