Note: This is a continually evolving field, so this article may become outdated within a few weeks.

Using Codex and GPT Together

Nikos Katsikanis - 14 June 2025

Complementary Strengths

OpenAI Codex excels at tackling big-picture features or repetitive tasks. It can refactor an entire project in a targeted way or apply metadata consistently across all files with a single prompt.

GPT models, by contrast, shine in focused, detail-oriented work. For example, GPT is well-suited to upgrading a UI component to use daisyUI. Codex often glosses over the finer details in such upgrades—leading to inconsistencies or only partially applied class changes.

Used together, Codex and GPT cover both broad and narrow development tasks, creating a smoother, more productive workflow.

Changing Perspectives

I've softened my stance on so-called "vibe coding," as discussed in the video below. That said, I still believe AI tools require significant human oversight—you can't expect a single prompt to build a full, reliable app on its own.

Initially, I reacted to these tools with curiosity more than confidence. But after building a complex MVP of a Rails app in just three days, I was shocked at how much my development speed had increased. These days, I often run two AI agents in parallel across different projects.

However, it's helpful/essential to understand all/most of the code that's generated. Blind trust will get you into trouble.

Like this site? I built Vanilla Toolkit (aka Vanilla JS Patterns) -> Fork me on GitHub