This article was originally written in Chinese and translated into English by AI.
Preface
Before I knew it, a full year had passed since I last wrote about my experience using AI. The changes in AI over that year have been extraordinary.
On the application side, the Doubao app surpassed 300 million monthly active users. While visiting museums, I saw quite a few people using Doubao as a tour guide: they would point their phones at an exhibit and ask it to explain the background.
On the tools side, the AI agent OpenClaw became a sensation in early 2026 before gradually fading from the spotlight. Codex and Claude Code were both released in mid-2025 and grew rapidly in 2026, with their user bases rising exponentially. They have gradually become valuable assistants for research and work.
On the large-model side, OpenAI released GPT-5, Anthropic released Fable 5, and new Chinese models continued to emerge. GLM 5, DeepSeek V4, Kimi K3, and many others have created a flourishing ecosystem.
This article briefly introduces the general AI-coding workflow and shares some thoughts from my recent experience using AI to build small projects.
Vibe Coding
In 2025, “vibe coding” became increasingly popular. The phrase essentially means programming by feel: users do not need to understand the technical details at all. By simply chatting with AI, they can ask it to generate, modify, and debug code.
Over the past year, I have gradually built several small projects through vibe coding, including applications for personal finance management, an English dictionary, and health tracking. I also completely redesigned my personal blog as a stylish desktop-inspired website.
Yet I know almost nothing about the technology stack required to build this website—React, Next.js, and so on. If I had written these projects myself, I would have needed to spend weeks learning the technologies. My first demo would certainly have progressed slowly and looked terrible. I would then have needed several more months of practice and refinement before the project became usable. In the age of AI, however, I only need to describe my requirements verbally. Within half an hour, AI can assemble a website that already looks reasonably polished. No wonder people say front-end development—the work of building website interfaces—is dead. This kind of work can now be replaced by AI with remarkable ease.
I have since developed a habit of using AI before bed each night, letting it turn the ideas in my head into reality.
AI Coding Tools
Mainstream AI coding tools currently include Codex, developed by OpenAI; Claude Code, developed by Anthropic; and Cursor.
Two concepts need to be distinguished here: AI coding tools and AI large language models. Codex, for example, is an AI coding tool created by OpenAI and can be used for free. The company providing the model that Codex calls is the one that charges for its use. Users can choose OpenAI's models or connect models from other companies, such as DeepSeek.
The same applies to Claude Code. It is an AI coding tool developed by Anthropic, and the tool itself can be used for free. It can connect to third-party model APIs or use Anthropic's own models. Anthropic's models are widely regarded as the best AI models for programming at present. However, the company often engages in public-relations behavior that I find off-putting, so I personally prefer OpenAI's Codex.
Other companies offer similar AI coding tools as well. Kimi is one example; its recently released K3 model ranked first globally in front-end code design.
The most visible change in AI coding from 2025 to 2026 has been the default interface of coding tools.
In the first half of 2025, when I was still using Cursor to build Goldflow Hourglass, AI coding tools still opened as code editors. Code occupied most of the screen, while an AI chat panel appeared at the side as an assistant.
Starting in the second half of 2025, the concept of agentic programming began to gain momentum. AI was no longer merely a question-and-answer tool. It became an agent capable of using plugins and Skills while planning and coordinating the work as a whole. The default home screen of AI coding tools changed from a code editor to a chat interface. The latest version of Cursor now displays only the chat panel by default; users must click IDE in the upper-right corner to switch to the code editor.
Similarly, the first question I had when I tried Codex was: where is the code editor? In fact, Codex does not include one at all. To edit code manually, users need to open a separate editor such as VS Code or Cursor.
Besides graphical interfaces, these AI coding tools usually support a command-line mode as well. With Claude Code, for example, you can simply enter claude in a terminal and ask it to modify the project in the current directory.
About Code Plans
Large-model companies generally offer subscription packages for coding. Using OpenAI as an example, monthly plans range from $8 to $200. More expensive plans provide more tokens.
When running a task, users can choose the model, reasoning effort, and processing speed. A more capable model, higher reasoning effort, and faster response speed generally consume more tokens. Code plans are usually billed monthly, but their usage allowances are calculated weekly. Once the weekly allowance is exhausted, users must either wait until the following week for it to reset or pay separately for API calls, which are usually more expensive. Previously, there was also a five-hour usage limit in addition to the weekly cap. Once that limit was reached, users had to wait until the next five-hour period. That restriction was recently removed, leaving only the weekly limit.
Codex and Claude Code also frequently give away additional usage or temporarily reset limits as they compete for users. This has been especially noticeable over the past month. Codex grants a quota reset whenever it gains another million users. In my experience, this means that even with heavy daily use, it is almost impossible to exhaust a Code Plan allowance.
Vibe Coding in Practice
Vibe coding has almost no barrier to entry. You describe what you want to AI, wait for a while, and it can generate an application that is already online and accessible.
For example, AI built this English dictionary in seven minutes and forty seconds. The result was a demo that anyone could access and that was already deployed to the cloud.
Of course, AI is not perfect. The first draft had some minor issues and small bugs. Through several more rounds of conversation, I reported the problems, added features such as word collections and API calls, and created a logo for the website.
In the end, I spent no more than two hours on the entire project. The tokens it consumed amounted to roughly 20% of the weekly allowance on my Code Plan, which costs $100 per month.
There used to be an online joke about someone who used AI to generate a website and then shared it with other people. Because the person knew nothing about computers, however, the address they shared was
localhost:3000, which was accessible only on their own device. That situation no longer occurs today, because AI can now deploy a website to the cloud for you, making it directly accessible to anyone.
Reflections
From April through July this year, I experienced the full progression from GPT-5.4 to GPT-5.5 and now GPT-5.6.
I can clearly feel that AI's visual taste is becoming stronger. When I first used GPT-5.4 and GPT-5.5, the websites GPT built were extremely ugly and had serious aesthetic problems. At the time, I would ask Claude Design to establish an art direction first, then have ChatGPT build from that style. With the recent release of GPT-5.6, however, the first draft of a website designed by ChatGPT from scratch is already attractive enough—at least for a design novice like me.
Starting with ChatGPT 5.6, AI will inspect the website after writing the code, create tests, simulate mouse clicks, and continue working until the site's behavior matches the original expectations. This kind of QA process either did not exist one or two minor versions ago or happened only rarely.
The internet is still full of people questioning whether AI can truly transform productivity. This reminds me of a story about the electrical revolution that I recently heard in a Bilibili video by the creator Meitou Talks U.S. Stocks.
Electricity is essential to human technology, yet productivity did not rise dramatically at the beginning of the electrical revolution.
When electricity was first introduced into industrial production, factories used electric motors to replace coal-powered steam engines. At first, the results were disappointing. Replacing the equipment required enormous investment, while the gains in efficiency and cost savings fell short of expectations. The reason was that production lines had been designed around the characteristics of steam engines. Simply replacing a steam engine with an electric motor could not make full use of electricity's advantages. After several decades of development, however, production lines were redesigned entirely around the characteristics of electrical equipment. Only then did factory productivity rise far beyond what had been possible with the old coal-powered machinery.
I believe AI is in a similar position today. We are at the stage when “electric motors have only just replaced steam engines”: AI has only begun to replace simple forms of knowledge work. Replacing individual steps in an existing workflow with AI may not deliver as much efficiency as expected. But I believe that over the next several years or decades, once people redesign entire systems of production around AI and make its capabilities the foundation of new working methods, productivity will undergo a truly dramatic transformation.