One week living in Codex instead of Claude: faster edits, leaner code, blunter tone

A developer spent a week defaulting to OpenAI’s Codex instead of Anthropic’s Claude for real Ruby and Rails work, and came away with a portrait of two agents with distinct temperaments. Codex tended to produce simpler architectures with fewer abstractions, less boilerplate, and notably fewer code comments, while Claude leaned toward richer scaffolding — extra concepts, Sorbet signatures, type aliases — that handled more edge cases at the cost of complexity. Codex also felt quicker to make the core change, though that speed evaporated once it insisted on rerunning tests, reviewing, and polishing the pull request, leaving no real wall-clock advantage.

The rougher edges were mostly around workflow and integration. Codex mishandled a stacked-branch setup, rebasing against main instead of the intended target branch and generating a pull request with over 4,000 additions until it was told explicitly what to do. CLI-based Jira and Atlassian work was clumsy, bouncing between browser logins and the terminal, whereas Claude drew on prior sessions to infer intent and push harder to get the job done. Codex did win on MCP authentication, where its explicit login flow proved more reliable than Claude’s habit of trying to auto-run credentials mid-turn and occasionally hanging.

The author’s throughline is a difference in personality rather than raw capability. Claude behaves like an eager colleague that tries to anticipate what you actually want and act on it, sometimes overbuilding in the process; Codex reads more like a literal-minded machine that does exactly what it’s told and stops at the first sign the task is complete. The takeaway isn’t a clear winner but a reminder that tool familiarity and default behavior shape developer trust as much as output quality — the author still reached for Claude under pressure simply because it felt like home.