Vibe Coding using Antigravity, part 2

Using Antigravity for a few minutes feels like magic. With just a few words in a prompt, you can have "working code" in minutes that would have taken a few hours before. I had specific expectations for this experience; I thought it would be a much more enhanced scaffolding of code; no different than "npm-ing" a starter template and then modifying what you need. Let’s just say it’s slightly more than that.

I vibe coded in my spare time, averaging probably two to three hours a day after work. I used Gemini 3.1 Pro the majority of the time, except for one weekend where I used it for about five to six hours and was throttled for a few days. I had to drop down to Gemini Flash for that one weekend, though I haven't been throttled since.

My plan was to update my website. It was running Wagtail CMS 4 on the backend, and I wanted to start from scratch with the latest version. My first prompts were something along the lines of: "I want a blogging website using the latest Wagtail CMS. It will use Docker for production and a ready-made YAML file for my CI/CD repository of choice." I hit enter, and it created an implementation plan for what it wanted to do. On the right side of the plan is a plus sign you can click to add comments or revise the steps. I was good with the plan, and after hitting enter again, it started to do its thing. Within a couple of minutes, it built the folder structure and files from an empty project.

Once it finished, it provided a walkthrough plan. It told me exactly what it did and gave specific steps to get it running; like how to create a Python virtual environment and the commands needed to launch the site. The default Wagtail site ran locally in Docker with no errors right out of the box. I was really impressed, I had only been vibe coding for a few minutes. To get to this point it might take a couple of hours writing by hand.

Since I was using a subscription, I was prompting away on new features whether I thought they made sense or not. I dumped most of it. I just wanted to see what the model can do. I didn’t need to worry about how many tokens I was using. I was as verbose as possible with my prompts, expletives and everything.

You have to treat vibe coding as if you are talking to another developer; do not treat it like a search engine. Let the model figure things out. You’ll be surprised at the solutions it comes up with; some good and some questionable, which I’ll go into detail about in the next post.

For instance, I wanted a way to minify the JavaScript in my project. Instead of looking for a single answer, I asked the model to provide multiple solutions, list the pros and cons for each, and then give a recommendation. It gave me exactly what I was looking for. It initially recommended the Django Compressor package, but I opted for its second suggestion: Vite via npm. The model actually provided a detailed explanation of why Django Compressor was technically the 'better' fit for a Python project, but when I mentioned I might eventually want to use TypeScript, it pivoted. It agreed that while Vite might be overkill for now, it was the best path forward for that direction.

This is exactly what I want from AI, not me poring over Stack Overflow or hunting through random tech blogs, but having a high-level architectural discussion.

Invariably, bugs and errors will creep in. When they did, I simply copied and pasted the error into the prompt and told it to fix it. I gave the agent permission to run commands as it worked through issues; it would read the stdout to look for errors and iterate. It even created temporary Python scripts to add users or dummy blog posts to fix a bug it created when the "next" and "previous" buttons weren't linking to the correct posts.

I was absolutely amazed. I can see why people get caught up in the hype. You say in plain English what you want, and you get "more or less" exactly what you asked for.

It’s all good... until you get into a debug error death loop, which happened a few hours into my experience and it happened a few more times after that.

Stay tuned for the next post, where I’ll dive into some of the negatives of vibe coding.