I love cooking, seeing all the ingredients come together and tasting the finished product. When I go out to eat, I like to order something that I can also make because I want to compare it. Does it taste as good? What ingredients did they use? How can my cooking be improved? While I have been on this coding journey, I have also been curious to compare the apps I built myself versus what AI could produce.

Here’s what I have discovered using Lovable.dev, Builder.io, Rocket.new and GitHub Copilot with ChatGPT-5, and Gemini 2.5 Flash (for image generation). I have been looking under the hood to see the code the AI produced, what frameworks were used, and the designs implemented. I’ve also been building some new apps using vibe coding.

1. Overall, I have been impressed by what AI can build. I’ve tested out small app projects than can take hours or days. The AI can do in minutes (5 mins or less for a first version). After I prompt the AI, I go get coffee or tea, and I come back and it’s done!

2. I have discovered that giving the full picture to the AI, then feeding it small instructions at a time has been effective. The AI’s context window where you put in the prompts can only handle so much. Plus I can see the changes happen in real time, test it out, and go ahead or roll back to a prior version. It’s like making small, frequent PRs (pull requests). Try to keep changes to one thing at a time, or small bundles. Its like adding one or several ingredients at a time, and tasting frequently!

3. The generated designs are good. It can turn a design for a simple app into a working app. I uploaded a figma design file into Lovable/Builder.io, which it then converted into code. When I tested out some other app ideas with no guidance on design, I notice that the apps look similar to apps produced by other people using the same AI tool. If you want it to stand out and improve the UI/UX, you do need to give it specific prompts. Your dish will look like everyone else’s unless you add some flair!

4. It generates a lot more code than what I did. It is less efficient, and I’m sure over time and with more specific instructions it will improve. If you are building a POC (proof of concept) or want a simple app, it helps to mention that in the prompt. For POC, it may not matter that there’s more code. The goal is to get it to work to validate something.

5. I’ve mostly seen the prompt to code tools use typescript programming language with HTML, Tailwind CSS, and React. It uses a JAMstack (Javascript, APIs, Markup) pattern and that could also be a function of what I have asked the AI to build. There is minimal backend code in Lovable or Rocket. To use APIs you need to use Supabase, a 3rd party service. There’s also no database so no data is saved. You can’t build a full stack app. The Lovable Cloud update last week addressed this.

By coding and using AI in parallel, it has opened my eyes to what AI is capable of and how humans can be in the loop. Who says you can’t have your cake and eat it too 🙂