This week, I tried a bit of vibe coding. I really enjoyed the process, it was actually pretty fun. When I get into a good rhythm and vibe well with the right AI agent tool, I usually can’t stop myself from diving deeper into my ideas.
Let’s be straightforward. What I found is that I really like using pair programming, working together with Google Antigravity. I think this might be the right way to do things.
For something simple, where high-quality code is not strictly required, and the codebase is not huge (like Unreal), it becomes a perfect target for vibing. You can let the AI write most parts of the code, while you focus on reviewing and thinking. It feels very similar to a pair programming workflow.
Alternatively, you can also let the AI act as the reviewer. This can work even in very complex projects such as Unreal. Usually, AI doesn’t find very deep architectural issues during code review, but it can spot some funny bugs, like spelling issues, lots of OO_1, OO_2, and suddenly there is an oO_3, or 0O_4. Humans can hardly notice these ridiculous problems, but AI can catch them within seconds.
Using AI to generate documentation is also a good idea. Combined with Doxygen, it can become a kind of documentation automation.
Recently, I was vibing on a graphics programmer’s image viewer. It’s not exactly simple, but not very complex either. I started with the free AI included in JetBrains’ bundled subscription, but I quickly ran out of capacity because I vibed a lot.
Then I tried Claude Code. Honestly, it was terrible. It’s fast, no pause, no planning, but extremely easy to fall into traps and never get out. My usage limit ran out in half an hour, and I had to wait five hours to continue. That really made me mad. I think they have good models, but terrible productions.
After two days of struggling with usage limits, I gave up, requested a refund, and then found that Google actually has a very beautiful all-in-one subscription, which I was already paying for.
It works pretty similarly to Cursor. I chat with the agent to get a plan, review it, and then let the agent execute.
I only dive into the code when the AI falls into traps, and yes, it still has traps. But overall, it’s a very convenient coding experience for a simple tool like this.
I want to ship it soon and don’t really want to get lost in details. So I just give ideas, review the results, and let the AI do most of the coding. For this project, it feels like a perfect workflow.
This work is still in progress. I’ll probably have more to share as I keep exploring the possibilities of AI-assisted coding.
Recent Comments