Vibe coding was pretty revolutionary by Codex
#SRG(Service Reliability Group) mainly provides cross-sectional support for the infrastructure of our media services, improving existing services, launching new ones, contributing to OSS, etc.
This article is an account of my experience with vibe coding using Codex.
Introduction
In recent years, development support tools that utilize AI have been attracting attention.
Among these, a new development style called "vibe coding" is gaining attention.
I've been using Cursor's Agent mode, an AI code editor, in my work.
Recently, the OpenAI Codex became available to our team, so we tried out vibe coding using it.
In this article, I will introduce what vibe coding is and my experience with vibe coding using Codex.
What is vibe coding?
First, let me explain vibe coding.
Vibe coding is a programming paradigm proposed by computer scientist Andrej Karpathy.
In this approach, a programmer provides a few sentences of instructions (prompts) in natural language to a large-scale language model (LLM), which then generates the software.
In this approach, the programmer's role changes from writing code manually to guiding, testing, and refining the AI-generated source code.
According to IBM, vibe coding's goal is to create an AI-powered development environment where AI provides real-time suggestions, automates tedious processes, and generates a standard code base structure.
The idea of "write the code first, refine it later" is said to be in line with agile principles such as rapid prototyping and iterative development.
Vibe coding in Codex
This time, I used OpenAI's Codex to experience vibe coding.
Codex is a cloud-based software engineering agent developed by OpenAI.
They can perform many tasks in parallel, including feature development, answering questions about the code base, fixing bugs, and proposing pull requests.
I don't think there is any clear difference from Cursor's Agent mode.
(Autonomously explore the code base, plan, and modify complex code bases)
By the way, I have the following written in my Cursor Rule to ensure that I search my code properly.
I feel that the difference between having "Search for related files if necessary" and not having it is quite big.
What was particularly impressive about using Codex was how easy it was to create a pull request (PR) with the push of a button, without having to set up any special instructions, such as Cursor's "Rules."
Cursor will not create PRs using GitHub MCP unless you explicitly tell it to do so.
The changes we tried this time were modifications to the sitemap.xml of the English version of the portal site that you are currently viewing.
All the code was created with Cursor Agent, mostly vibe coding.
The details are as follows, and it is so complex that you would give up on implementing it without AI.
Instructions
It goes against the principles of vibe coding to take the time to read through the code output by vibe coding, but I did look it over this time while writing this article, and I think the output is easy to read and of good quality.
Although it's not stated in the instructions, it's nice that they have also updated the README.md.
In the case of Cursor, it is often not written (if you write the Cursor Rule mentioned above, it will be done).
Conclusion
I personally felt that vibe coding using Codex was "pretty good" and a very innovative experience.
With AI taking on much of the coding, developers could be freed up to focus on higher-level design and problem-solving.
Of course, reviewing and testing the code generated by AI is essential, but we felt that vibe coding would be a powerful tool in the early stages of development and prototyping.
I hope that AI development support tools like Codex will continue to evolve and that vibe coding will become a more common development style in the future.
SRG is looking for people to work with us. If you are interested, please contact us here.