Outdoors during the holiday, it suddenly hit me—I’ve been using AI-assisted coding for almost three years! So let me record my experiences and takeaways.
In 2023, I first used Devv. Compared to later tools, AI-assisted coding back then was pretty primitive. You had to manually copy your code and requirements into the chat box, hit send, wait for the AI to generate code, then manually copy it back into the editor. This back-and-forth cost a lot of time. The upside was that I remembered every line of code. When an issue came up in production, I could usually spot the problem right away. The downside was all the time wasted on tedious, repetitive work.
After using Devv for a year, my proudest moment was designing and implementing CodeGod Challenge (https://1024.codefather.cn/) for the 2024 1024 Programmer’s Day. This marketing campaign drove a huge number of users to the paid product. Without that AI coding tool, I honestly couldn’t have delivered it on time. That experience convinced me that AI would only become more important.
Then I tried Cursor, which at the time was the best solution for eliminating that copy-paste dance. It was like magic autocomplete—fast, accurate, and even supported cross-file autocomplete. Smashing Tab all day felt amazing. Later, the Composer feature arrived, letting me complete a large feature in one go. I used it to build a full-stack Nest.js + Next.js project called SVG Show (https://github.com/coderPerseus/blog/issues/48) from scratch. It leverages Claude’s powerful SVG generation to stream the drawing process in real time and export images. Ruan Yifeng even featured it on his blog. But the cost turned out to be too high — I wasn’t making any money — so I gave up. That gave me a fresh perspective on product development.
Then Claude Code launched. AI-assisted coding had evolved from a copilot to an agent. I’ve been using Claude Code for development ever since. I’ve lost count of how many features I’ve built with it. At first, I treated it as just a coding agent, but then I realized it could operate the computer programmatically. So I started delegating more tasks to it, acting only as the person who reviews the results, ensuring the AI stays within my predefined boundaries and delivers the expected functionality.
Now I’m deeply using both Codex and Claude Code to develop my own tool, popMind — an Electron-based selected-text translation and AI explanation tool. With AI, I had a working prototype running in no time. Of course, a key lesson is to find existing products to use as templates — AI is best at painting by numbers.
In conversations with friends, what struck me most is that AI’s arrival is tough on people looking to break into programming. In the past, a programmer would spend an afternoon implementing a feature, tracking down and fixing a bug, or designing a solution. Now all of that gets handed to AI, which does it faster and better. AI replaces human thinking, but the understanding and experience behind that thinking are invaluable to a programmer.
With AI, programmers no longer write code; we all drive development with natural language. Unsurprisingly, coding ability is no longer the core competitive edge.