Architect Uses Codex to Develop Plugins and Websites with Zero Coding: Six Months of Hands-On Experience Shared

A frontline architect at a major state-owned enterprise began teaching himself AI programming during Chinese New Year this year. In just six months, he used GPT Codex to develop multiple professional plugins and a personal website. He doesn't know any programming language—all development work was accomplished entirely through natural language conversation. This case perfectly demonstrates: In the Vibe Coding era, everyone can become a product manager.

From Architect to AI Developer: Why Learn Zero-Code Programming?

The video author is a frontline architect at China Energy Engineering Group, working on digitalization in architectural design. He has been exploring AI applications in the architecture field for over two and a half years. His core philosophy is highly pragmatic—only tools that align with production and solve real problems are good tools.

In his daily work, he frequently encounters specific production challenges: extracting public land auction information from websites, developing custom Blender or ComfyUI plugin nodes, and building a personal website that integrates commonly used tools.

It's worth explaining the roles these tools play in the architecture industry. Blender is an open-source 3D modeling and rendering software widely used in architectural visualization, supporting custom plugin development through Python scripting. ComfyUI is a node-based workflow interface for AI image generation that allows users to build complex AI image generation pipelines by connecting different processing nodes. In the architectural design industry, these two tools are commonly used for concept rendering, scene visualization, and rapid design scheme visualization. Developing custom plugin nodes means automating specific design workflows, such as batch-generating architectural renders from different angles.

As for land auctions—referring to the three public land transfer methods of "bidding, auction, and listing"—this represents the source information for architectural design projects. Architects need to obtain plot specifications like land use type, floor area ratio, and building height limits from local natural resources bureau websites. This information directly determines the basic framework of design proposals. However, this data is scattered across hundreds of government websites nationwide, in varying formats, making manual collection extremely time-consuming. Using AI programming to develop web scraping tools that automatically extract and organize this information is a classic example of "small tools solving big pain points."

If he waited for others to open-source solutions or hired professional programmers, the efficiency would be too low and the cycle too long.

So starting from Chinese New Year 2025, he embarked on the path of AI programming (Vibe Coding). Vibe Coding is a concept proposed by OpenAI co-founder Andrej Karpathy in early 2025, referring to a programming approach where developers describe requirements entirely in natural language, and AI automatically generates the code. Unlike traditional low-code/no-code platforms, Vibe Coding doesn't require dragging components or configuring parameters—instead, you describe the functionality you want as if talking to a person, and AI independently handles code writing, debugging, and even deployment. The core philosophy is: you only need to "vibe" what you want, without needing to know precisely how to implement it.

What surprised him was that AI programming capabilities far exceeded expectations—an architect with absolutely no coding knowledge could independently develop plugins and build websites.

AI Programming Tool Selection: The Journey from Domestic Models to GPT Codex

Three Generations of Tool Iterations

The author's AI programming tool journey went through three phases:

1. VSCode + GLM5 (domestic solution): Used during the beginner phase, capable of completing basic tasks
2. Claude Code (Sonnet 4.5): Noticeably improved programming capability
3. GPT Codex Desktop: The final choice, offering the best overall experience

GPT Codex is OpenAI's AI product focused on code generation, initially well-known as the underlying engine for GitHub Copilot. The Codex desktop version launched in 2025 is a standalone application integrating a code editor, terminal, and built-in browser, allowing users to complete the entire development workflow from writing to previewing within a single interface. Its core capability comes from GPT-series large language models trained on massive code datasets, enabling it to understand natural language instructions and convert them into executable program code.

He has firsthand experience with the gap between domestic and international models: solving a small problem with domestic models might require ten or more rounds of conversation to reach a usable state, while Claude Code or Codex typically achieves the goal in two or three rounds.

A specific example was developing a Gaussian-driven camera plugin that needed an intermediate control point to fix camera rotation. This involves 3D Gaussian Splatting technology—a major breakthrough in the 3D reconstruction field since 2023. This technology represents scenes as large numbers of 3D Gaussian ellipsoids to achieve high-quality real-time rendering, improving rendering speed by dozens of times compared to traditional NeRF (Neural Radiance Fields) methods. In architectural design, this technology can quickly reconstruct photographed architectural scenes into interactive 3D models. The camera control plugin enables smooth camera path animation in such 3D scenes, with control points allowing precise control over camera movement trajectory and rotation angles. This feature could never be implemented with domestic models and was ultimately accomplished with Claude Code and Codex.

Why Codex Desktop Was the Final Choice

Codex Desktop has several key advantages:

• Easy to obtain: Simply search and install from the Microsoft Store, supporting both Windows and Mac
• Intuitive interaction: Built-in browser for real-time preview of modifications
• Annotation feature: You can click directly on the page to mark positions that need modification, which is far more precise than taking screenshots in VSCode
• Latest model support: Currently recommends using the GPT 5.5 model for the strongest programming capability

API Relay Station Money-Saving Guide: The Key to 90% Cost Reduction

Plus Membership vs. API Calls: Cost Comparison

Codex supports two login methods: GPT Plus membership account and API key. The author strongly recommends the API approach for one straightforward reason—the cost difference is enormous.

Using a Plus membership account, generating just a few images consumes about 20% of the quota, and you quickly have to wait for the rate limit to reset. Through an API relay station, generating four images costs only 0.14 RMB, working out to less than five cents per image.

An API relay station (also called API proxy or forwarding service) is a third-party service that bulk-purchases API call quotas from AI companies like OpenAI, then resells them to individual users at lower unit prices. Its business model is similar to "group buying"—obtaining better prices through bulk purchasing, then reselling with a reasonable profit margin. Users receive a compatible API key through the relay station and simply replace the official key in their tools. Note that using a relay station means your request data passes through third-party servers, so security risks should be evaluated when dealing with sensitive information.

The author did the math: GPT's official Plus membership costs about 150 RMB per month, while the same amount deposited to a relay station lasts two to three months. For non-full-time programmers, the overall daily expense is only about 10-20 RMB.

Codex API Configuration Recommendations

The author recommends configuring the Codex CLI API first. Once configured, it can be used universally across the entire Windows system without needing repeated configuration in different tools. Desktop configuration also has detailed tutorials—just follow the steps.

Hands-On Demo: Redesigning a Personal Website with Natural Language

Project Launch Process

The author demonstrated how to use Codex to modify his self-built personal website. This website integrates commonly used AI image generation tools, and the entire site was built through Vibe Coding.

The workflow is remarkably simple: type "Launch this project in the Codex browser, I want to continue modifying it" in Codex's prompt input box, then wait for the project to load. Since the project is large, the startup process took about 8 minutes, during which Codex automatically handled various minor issues that arose.

Precise Modifications Driven by Natural Language

During actual modifications, the author noticed that preview images in the image generation module were too long, requiring scrolling to see completely. He used Codex's annotation feature—clicking on the page position that needed modification and directly typing his requirement:

"Make the output image preview fit within one window, don't make it too long. The preview frame should be able to display one complete image thumbnail."

Codex immediately understood the requirement and modified the code. After completing modifications, it also explained its reasoning, which itself serves as a learning opportunity. This "what you see is what you get" interaction essentially compresses the traditional multi-step software development process of "requirement description → code implementation → compile and run → check results" into a single step, dramatically lowering the cognitive barrier for non-technical people to participate in development.

The Multi-Round Iterative Optimization Process

After the first modification, the image preview area met requirements, but the "Recently Generated" section below didn't automatically snap into place. The author continued by selecting that area and describing the issue in natural language, and Codex made another precise fix.

This process perfectly demonstrates the Vibe Coding workflow: you don't need to know how to write code, you just need to be clear about what you want. Each modification is like conversing with an exceptionally perceptive programmer, iteratively approaching the ideal result. This iterative development approach is called "agile development" in professional software engineering. Traditional teams typically iterate on weekly cycles, while in Vibe Coding mode, iteration cycles are compressed to minutes.

Four Tips for Zero-Experience AI Programming with Codex

Based on six months of hands-on experience, the author summarized several key insights:

1. Clear requirements matter more than technical ability. As the product manager, you need to clearly describe "what you want," not "how to do it." AI handles implementation; you handle definition. This is actually the ultimate manifestation of the "separation of concerns" principle in software engineering—domain experts focus on business logic and user experience, while AI handles all technical implementation details.

2. Prioritize first-tier AI programming tools. For programming scenarios, GPT Codex and Claude Code are far more efficient than domestic alternatives. The time savings far outweigh any difference in tool costs.

3. Start from real work pain points. Don't blindly develop tools unrelated to production. Building small tools and plugins targeted at your own work pain points yields the most significant efficiency gains.

4. Embrace iterative thinking. AI programming isn't a one-shot process—it's about continuous optimization through multiple rounds of conversation. Each modification is a learning opportunity, gradually building understanding of the development process.

Vibe Coding is lowering the barrier to software development. When an architect can independently develop professional plugins and websites, it means professionals in every field have the opportunity to become developers of their own workflows. The people who best understand the requirements can now implement them with their own hands. This perhaps signals the arrival of a new era: software development is no longer the exclusive domain of programmers, but a capability accessible to every professional with clear requirements.