Original video in Chinese.
Key Takeaway
- Cursor is a powerful AI coding IDE that natively supports AI features, making pure chat-based coding with AI possible.
- By using Cursor for 10 minutes without writing a single line of code and only talking to AI, the author built a Chrome browser extension, showing how efficient AI coding can be.
- Cursor’s “Apply” feature can automatically locate where code changes need to be made, improving debugging efficiency.
- The value of AI coding tools lies in giving users new skills they did not have, rather than simply replacing existing work.
- The article emphasizes that AI can amplify individual capabilities and create more value, and looks ahead to the broad prospects of AI in product development.
Three years from now, what will the most popular programming language in the world be?
Chances are, it won’t be Python, and it won’t be JavaScript. It’ll be English.
Recently, there’s been a wildly popular IDE and coding tool in the AI world called Cursor. Once you’ve used it, you’ll probably think the same thing.
I see Cursor as the ultimate evolution of VS Code. It’s a fork of VS Code, with basically all the same features, and you can import everything over seamlessly. But on the AI side, Cursor does a much better job. It doesn’t work like GitHub Copilot, which is inserted as a plugin; instead, it supports AI natively and inherently — and that’s very important, because it makes a qualitative difference.
As for models, Cursor is very generous: you can use its model for a monthly subscription fee of $20, or you can use your own model by simply entering an API Key. It even lets you use GitHub Copilot, though it warns you that it’s not recommended because its performance isn’t as good as Cursor’s own.
Last weekend, I spent some time testing Cursor intensively. My best run took less than 10 minutes. I didn’t write a single line of code; I just chatted with AI, and I developed a Chrome browser extension that can call GPT to summarize webpage content.
The first 3 minutes of those 10 minutes were spent writing requirements:
I want to develop a browser extension whose main function is to summarize webpages using a large language model. It will first do Scraping, meaning it will pull down all the content on the page. Then it will send that content to the model and output it in the format of Summary and Key Facts.
As for specific features, the extension will have three buttons: Summarize means summarize. Settings means settings, and it will ask the user to enter an OpenAI API Key. The extension will first verify whether this Key is usable. If it is, it will fetch all available models as a list so the user can choose and save one. Clear means clear the previous summary result, or interrupt the current summarization task.
After I finished writing all this in Word, I copied it, opened Cursor, opened the chat interface through the control panel, and pasted everything in. The experience after that was basically the same as when we use ChatGPT or Claude:
Cursor responded very quickly, understanding and breaking down the requirements in just a few seconds. It told us which files to create, and gave us the file names and code for each one. At that point, all we had to do was create the files as instructed, keep the corresponding files open, and click Apply; the AI would fill in the code.
The Apply feature is especially convenient. Because during debugging, code will definitely need to be modified. Cursor doesn’t generate the entire code again — that would be too slow and consume too many tokens. So it only outputs the few lines that need to be changed. At that point, the burden shifts to the user, because finding the lines to change among hundreds of lines of code is still pretty taxing. So the Apply feature automatically locates the edit position, marks the original code in red and the suggested code in green, and after the user confirms, it automatically replaces it.
Once we had pasted all the code into the files, we could test it.
Open the Chrome browser’s extensions page, enter Developer Mode, and open the folder containing the code, and the extension can be loaded.
The first test is bound to have bugs. In our case, we found that the Settings button didn’t respond. Simple enough: go back into Cursor, tell the AI about the problem, and then replace the newly generated code using the Apply feature I just mentioned.
Refresh the extension, and now the settings page opens. Enter the API Key, and just like the requirement said, the extension first verifies it, then pulls up the Model List. At this point, we found that the list was incomplete; there were only two models. I’m guessing Cursor preset them on its own and didn’t actually fetch them. Also, clicking Summarize didn’t work properly — it only showed a demo.
So I communicated with Cursor again and reported these two issues. Debugging like this happens often. But this time I was pretty lucky; it only took two back-and-forth rounds.
Refresh the extension again and test it once more. This time, the model list displays correctly. Click Summarize, and the extension starts working. After waiting a few seconds, the summary is successfully generated. To make sure it really worked, I opened two more webpages, and both were summarized successfully too.
I checked the time. From writing the requirements to successful testing, it took about 10 minutes. This wasn’t the first time I’d done something like this. I had actually tried it the day before too, but it wasn’t very successful, and the process left me a bit broken.
Every code change brought new bugs. I watched helplessly as code that had originally been only 30 lines ballooned tenfold into more than 300 lines, and the problem still wasn’t solved.
Later I thought, why not switch to another implementation approach and use multimodal input instead? First take a screenshot of the whole page, then give it to GPT to recognize and extract the content. That still didn’t work. Anyway, I spent almost an hour messing with it.
On the second day, I reflected on it. In fact, this extension’s workflow is just two steps: first Scrape, then Summarize. That’s how I handled it when I wrote Agent Workflow myself before. So I adjusted the requirements a bit, did the product manager’s job myself, and let Cursor focus on coding. Sure enough, the effect was immediate. The rush at the moment of success was just like clearing a game level.
This summarization extension is still very rough. If I wanted it to reach the point where it could be shipped in an app store, I’d probably need to spend some more time:
First, its Scraping is fairly basic and could be strengthened further so it can handle more webpages.
Second, its summarization isn’t good enough yet. This is easy to adjust; I don’t even need AI to modify it. I can just make the prompt requirements more detailed myself.
Third, it only supports OpenAI right now. Google, Anthropic, and others could all be added.
Fourth, make the UI prettier. That’s also easy. I can find someone else’s product, take screenshots, and give them to AI — it can definitely reproduce it.
If everything goes smoothly, I estimate these four improvements could be finished with another half hour. Once everything is done, it can be submitted for Google review and then released.
I don’t like saying things like “the future is already here,” but Cursor shocked me far more than ChatGPT did two years ago. Maybe that’s because the things ChatGPT did were all things I could already do — wasn’t it just generating some text? And not even as well as I could. So what it gave me was only freshness and surprise.
But Cursor is different. Something like developing a browser extension is something I absolutely cannot do, let alone get a prototype working in 10 minutes. So this was a kind of shock — it gave me skills I didn’t have at all. I think that is AI’s real value.
Today, many people, especially many bosses in China, think AI is for cutting costs, replacing employees, and serving as an excuse for layoffs. They view labor as a cost. The world in their eyes is limited — there’s only so much room, so they have to save wherever they can.
But in fact, the world doesn’t have to be a “finite game”; it can become an “infinite game.” AI can give individuals skills they’ve never had before, or amplify their existing skills several times over. To satisfy more needs, to create more — isn’t that better?
For me, if I want to build a small tool, I can now complete it directly with Cursor’s help. If something is a little more complex, I can first build a prototype myself, then pay front-end and back-end people to help me refine it.
Suddenly, I feel like my whole world has become so much wider. Maybe by next year, what I give everyone won’t just be videos and articles like this, but more forms — it could be a website, it could be an app, or even a small model. I’m so excited!
OK, that’s it for this episode. I’m going to keep dragging Cursor along to develop things. I’ll post more discoveries in the newtype community later. If you haven’t joined yet, hurry up — the community already has more than 400 people. It will definitely surpass 500 this year, and next year I’m aiming for 1,000. The other day I thought of a slogan that fits my channel and the community perfectly: “In the AI era, cross the river by feeling for Old Huang.” Alright, enough rambling — see you next time!