A week of using Codex more than Claude

During the past week, I had the opportunity to use both Codex and Claude more extensively than before. Here are some personal impressions I've gathered:

Firstly, while I attempted to keep Claude and Codex on equal footing in terms of plugins and skills, Claude had access to more skills due to the skills created during previous sessions not being ported over to Codex. A simple solution is to point Codex at the Claude skills folder and instruct it to adapt them for use with Claude.

When faced with urgent tasks, such as debugging issues, I found myself instinctively opening Claude. Although I'm not suggesting it was objectively better, it felt more familiar and comforting in those moments. Given that I rely on tools I'm already comfortable with during debugging, using Claude proved advantageous.

Another notable difference was the number of comments left by Codex in Ruby/Ruby on Rails code compared to Claude. I appreciated the concise nature of Codex's comments and will be sharing some of the experiments I conducted on this topic soon.

The output from Codex's agent harness appeared to be more technically oriented than that produced by Claude. To me, Claude feels like a colleague collaborating with you in a Tuple session, while Codex feels like an advanced version of Data from Star Trek.

I noticed that I wanted to open multiple focused Codex sessions rather than a single, large Claude session. This change might not be unique to Codex, but it was something I observed while working with it.

It seemed to me that Codex made changes more swiftly than Claude. However, once the primary modifications were made, the process of completing the pull request took considerable time. I appreciated the thoroughness of Claude's work, but the overall time difference was not significant in the end.

Codex seemed to generate a simpler code architecture in comparison to Claude. While Claude tends to introduce various abstractions, concepts, Sorbet signatures, and type aliases, Codex remained more contained, resulting in less complexity in its solutions.

During the course of this week, I experimented with an improved flow of code research involving design changes, review changes, implementation, and verification. Interestingly, both Codex and Claude implemented the same requirement using the same documents. However, while Claude's code was slightly more complex, it was also able to handle additional edge cases.

There were instances where Codex made mistakes, whereas Claude was able to better understand my intention to branch out from other work and maintain synchronization between branches. On one occasion, Codex unintentionally created a PR with over 4000 additions by rebasing with the main branch instead of the target branch, requiring explicit instructions to prevent this from happening.

In terms of working with Jira and Atlassian in my current environment, using the CLI tool instead of the MCP, Codex proved to be more troublesome. It would open JIRA for login, switch to the CLI, and then return to the browser, whereas Claude was more attentive to understanding my specific needs and executing tasks accordingly.

Lastly, I found that working with MCPs was more efficient using Codex's CLI approach. Codex would prompt me to execute commands like "codex mcp login" and guide me through the authentication and authorization process, while Claude sometimes attempted to run these operations automatically, which could lead to potential issues and delays.

Overall, I perceive Claude as more inclined to exceed expectations and make assumptions, while Codex behaves more like a dependable companion that carries out tasks precisely as instructed, without overstepping boundaries.