Product Feedback & Suggestions for OpenAI Codex
I am a long-term user of ChatGPT and Codex, and I actively use Codex in real-world engineering development. After extensive hands-on experience, I would like to share two important suggestions for improving Codex.
1. Prevent Overengineering in AI-Assisted Software Development (My Most Important Concern)
I believe this is one of the most important issues Codex should address, particularly for developers who are using AI coding tools for the first time.
When I first started using Codex for software development, I noticed a recurring problem: as development tasks progressed, the project tended to grow unnecessarily complex. Large numbers of redundant files, obsolete historical files, duplicated documentation, and unnecessary architectural layers gradually accumulated.
The main issues include:
1. Accumulation of obsolete files
During repeated development and iteration, Codex may create new files without properly cleaning up temporary files, deprecated implementations, or outdated versions. Over time, the project becomes increasingly bloated.
2. Overengineering
Problems that could be solved with relatively simple implementations are sometimes turned into unnecessarily complicated modules, abstraction layers, or management mechanisms. This increases development costs without providing corresponding value.
3. Duplicated and conflicting decision records
The same engineering decisions are repeatedly documented across different files, documentation, and Agent configurations. Sometimes these records even become inconsistent with one another. As the project grows, this creates an increasingly difficult-to-maintain collection of duplicated and conflicting decisions.
4. Declining long-term maintainability
As the number of files and architectural layers increases, developers find it increasingly difficult to determine which files are still relevant, which implementations are obsolete, and which decisions should be treated as authoritative.
I believe these problems can easily lead users to incorrectly conclude that the underlying model lacks coding ability, execution accuracy, or the ability to complete complex tasks.
However, based on my own experience, model capability was not the primary issue. The more significant problem was overengineering and the lack of effective engineering management constraints.
I eventually addressed these issues by introducing dedicated Agents and rules focused on two key principles:
-
Prevent overengineering by prioritizing simple, direct, and maintainable solutions.
-
Avoid duplicating engineering decisions by establishing a single source of truth for important project decisions.
I then reorganized and rebuilt the entire project around these principles.
The results were surprisingly good. Codex became significantly more effective. Overall task completion time decreased substantially, while execution accuracy and the ability to complete tasks improved considerably.
This experience made me realize that AI coding tools need more than excellent code-generation capabilities. They also need a strong understanding of software engineering discipline and long-term project management.
I would strongly encourage OpenAI to address this issue through the following improvements:
-
Introduce default safeguards against overengineering in Codex workflows. Encourage the simplest implementation that adequately solves the problem, rather than unnecessarily introducing complexity.
-
Improve the ability to identify obsolete, redundant, and historical project files, and prevent unnecessary project bloat during long-term development.
-
Establish better engineering decision-management mechanisms to prevent the same decisions from being repeatedly documented across multiple files and Agent configurations.
-
Provide first-time users with practical guidance on project file management, context management, and engineering decision tracking.
-
Make Codex consider not only whether the current task can be completed, but also whether the resulting project will remain maintainable throughout future development.
This is particularly important for developers who are new to AI-assisted programming. Many of them may not have the experience to recognize these problems as they gradually emerge.
If these issues remain unaddressed, users may end up blaming the model’s capabilities when the real problem is an increasingly complex and poorly maintained engineering environment.
I believe this directly affects both the user experience and how users perceive Codex’s actual capabilities.
2. Simplify the Codex Application Settings Interface
My second suggestion concerns the overall design of the Codex application.
At present, I find the settings interface overly complicated. It lacks the simplicity, visual clarity, and intuitive design that I would expect from a modern developer tool.
I believe Codex should combine powerful functionality with a clean and straightforward user experience.
My suggestions are:
-
Simplify the overall settings structure and reduce unnecessary navigation layers.
-
Separate commonly used settings from advanced settings, rather than presenting everything to users at once.
-
Provide more sensible default configurations so that first-time users can get started without having to understand numerous complicated options.
-
Improve the overall visual design to make the interface cleaner, more consistent, and easier to navigate.
I believe powerful functionality and a simple interface are not mutually exclusive. Complex underlying capabilities should be presented through an intuitive and streamlined user experience.
Final Thoughts
I genuinely appreciate the direction in which Codex is developing, and I intend to continue using it extensively in real-world engineering projects.
These suggestions are based on my own practical experience rather than theoretical expectations.
I hope Codex can evolve beyond being an excellent code-generation tool and become a truly capable AI development assistant that understands engineering quality, project management, and long-term maintainability.
I hope this feedback can contribute to the continued improvement of Codex.
Thank you for your work.