Recently, the field of AI programming has become one of the most successful applications in the AI domain. I tested a few tools and would like to share my experience with friends who haven’t tried them yet:
1. Bolt.new
This tool is an AI-powered web development agent that runs directly in the browser. It provides a complete front-end scaffolding and is particularly suitable for building web projects from scratch. The underlying model is Claude 3.5.
2. Cursor
Based on VS Code, I took a closer look at VS Code’s design philosophy, and its architecture is indeed impressive. The core design revolves around cross-platform compatibility, language-agnostic support, and a community-driven plugin ecosystem. Its architectural innovations and open-source strategy make it the preferred tool for multi-language development—no wonder Cursor chose to build on VS Code.
Using these two tools together seems like a golden combination:
- Use Bolt.new to quickly generate the initial project structure and boilerplate code.
- Then import the generated code into Cursor for more detailed development, refactoring, and implementation of complex features.
There are two other tools on the market: Replit and Windsurf (which is reportedly being acquired by OpenAI). I haven’t had the chance to try them yet—welcome any insights from those who have. Finally, there’s Devin, but at $500/month, it’s beyond my budget. I ended up asking Google Deep Search to generate a comparative summary report.
Initial Impressions:
The strengths are obvious, so I won’t dwell on them. The main drawbacks are the challenges in modular reusability—it’s not easy to manage at the moment. Currently, the only way seems to be abstracting at the prompt level. My current thought is to summarize prompts to achieve modular reuse. Another issue is review and validation—you never know if there might be hidden vulnerabilities. A friend of mine created a diagram to summarize this.
However, I feel these tools should evolve quickly to address current limitations.
Recently, I also took some time to run OpenManus. After downloading it, I was able to get it working—this open-source ecosystem has attracted many developers to build on it, making it a very interesting project. Zhipu’s agent framework has also been open-sourced, enriching the open-source agent ecosystem even further.
Way to data scientist 
Leave a comment