Quick impressions: A week of using Codex more than Claude

The author is conducting an informal comparison of AI coding assistants, having used Codex more than Claude this week, with a full analysis planned. They note that while they tried to keep both tools on par with the same plugins and skills, Claude ended up with more skills because some sessions were ported, and they suggest pointing Codex at the Claude skills folder to convert them.

On the positive side, Codex produced fewer comments in Ruby/Rails code (which the author liked), gave more technical output that feels like 'Data from Star Trek' compared to Claude's collaborative colleague vibe, encouraged more focused sessions rather than large monolithic ones, and made changes faster initially. However, Codex took much longer to finish pull requests due to extensive test reruns and reviews, so the time difference evened out. Codex also created simpler code architecture with fewer abstractions than Claude, but this meant it sometimes missed edge cases that Claude's more complex code would have handled.

There were notable issues: Codex made mistakes with branching, rebasing against main instead of the target branch and creating a 4000+ addition PR, and required explicit instructions to rebase correctly. Working with Jira/Atlassian was a hassle (browser/CLI switching), while Claude was more adaptable based on previous sessions. On the other hand, the author preferred Codex's MCP login flow, which opens the correct authentication easily each time.

Overall, these are subjective impressions highlighting trade-offs in speed, thoroughness, simplicity, and tooling. The author still felt more 'at home' with Claude in urgent debugging contexts, but found Codex's approach appealing enough to explore further.