Using Codex. Impression? I think it's pretty cool that computers can do this now. Codex successfully automates some of the core tedium of making software: finding and reading docs, writing boilerplate and glue code, using git, navigating the command line, etc.
My other impression was that this probably won't be a game-changer for me. I found myself gravitating towards small, incremental commands whose outcomes I could quickly comprehend. More sophisticated commands generated much larger diffs, and, hot take: I like writing code more than I like reviewing it. The more I asked the model to do, the more of my time shifted from the activity I like (coding) to the activity I tolerate (reviewing).
I have considered that I could be missing the point—the "vibe" in "vibe-coding"—by exhaustively reviewing the model's code. Maybe I'm a control freak about my codebase and can't emotionally tolerate committing code I haven't grokked. Maybe I'm hurting my economic output by not leveraging modern automation to the fullest. I happily accept both of those possibilities as I relegate Codex to the back of my toolbox; still there at hand, but not the first (or fifth...) thing I reach for.
Addendum: at the time of writing, the version of the Vibe coding page on Wikipedia indexed by search engines stated the following (entertaining...) definition:
