Which tool are you on ...
I will create a series of posts here that outline how to build your own web apps using AI coding. While my focus is on academics as the audience, this walkthrough will help everyone interested in the topic. I am a coding rookie; I do not have extensive experience in software development. So this is a guide for rookies, by a rookie. That will make it necessarily incomplete, but accessible.
I am also a notorious Windows user; most of what I describe here reflects my experience with Windows. Please don't ask me Mac questions (and I'm guessing Linux is out of scope for my objectives here).
Let's start with a basic definition. The process I am describing here is generally called vibe coding. Andrej Karpathy, one of the co-founders of OpenAI, coined this term (https://x.com/karpathy/status/1886192184808149383). It fits. The idea is that you rarely have to code; instead, you provide your AI with plain English prompts to generate and modify your code. For the simple apps we need in the classroom, this works 98% of the time; I have had only one occasion in three months of coding where I had to manually inspect and modify the code.
Since this is about AI coding, maybe the first question is what AI to use. There are different families of LLM foundational models, such as GPT from OpenAI, Claude from Anthropic, and Gemini from Google. Within each family, there are multiple models to choose from. GPT, for example, has Codex 5.2; Claude has Opus 4.5 (which is what Claude Code is mostly using); Gemini 3 Pro is your go-to model for coding in the Google family, although the recent Gemini 3 Flash is more efficient and almost as good for coding. You'll need a Pro subscription for at least one of these models to get far. Pick one, they are all good. Or even better, pick many. I have pro subscriptions to all three, and I often switch between them. The old adage that an ensemble of models outperforms each individual one seems to hold for vibe coding as well.
The next question is how you want to access that model for coding. There are, I believe, three routes for that - Command Line Interfaces (CLI), Integrated Development Environments (IDE), and complete solutions.
Let's start with CLIs. Think of your classic 'DOS' box, if you are as old as me. Click your Windows button, open the search bar, and type cmd. Voila, a simple CLI appears. You can also search Windows for your PowerShell, which is a similar CLI. (PowerShell is a bit more powerful - check out https://netwrix.com/en/resources/blog/powershell-vs-cmd/ if you are interested in the differences). Codex 5.2 and Claude Code have simple command-line interfaces that you can install. I will provide more technical installation advice some other time. For now, the 'vibe' of CLI coding is pretty much you go to the right folder, charge up your AI in the CLI, and enter your prompt. This is fast and efficient, but it gives you less control over what happens. To actually see the code and possibly modify it, you will need an IDE.
Think of an IDE as a fancy text editor for coders. Good examples are the (free) Microsoft VS Code, Google's excellent (and free) Antigravity, as well as Cursor (which has a free version, but generally costs some money). AI comes in extensions. You can get Codex, Gemini, and Claude extensions for VS Code; Antigravity generally only works with Gemini and Claude. Think of these as sidebars in the text editor that let you enter prompts, which then let the AI go into the folder you are in (and its subfolders) to modify the code. It works like a charm. The IDEs also include a Terminal, which lets you remotely access your files and many online resources, e.g., deploy your app to a website. I am using both VS Code and Antigravity - the latter only with Gemini, since it seems to work better in Antigravity. I have not tried Cursor. I heard it's good. Using an IDE can seem a bit intimidating to a non-coder, since you see files, code, dollar signs, etc. But you can ignore most of what you don't understand. And you can do everything you need in an IDE - code, modify, deploy. Nevertheless, if you want a simpler experience, a complete solution might be a better fit.
This leads me to the last way of accessing a model - the complete solutions. At least, that is how I call them. These include sites like lovable.dev, Google's AI Studio or Replit agent. I honestly have less experience with them - I wanted to be a bit more hands-on by using an IDE. But these tools promise to do everything for you through your prompts: From developing the app to deploying it on sites and monitoring usage. Sounds pretty cool. I'll play around with them and report back some other time.
That's a basic overview of the tools you need. I will provide more details next time. Until then, good night and good luck!
