From Idea to Application with AI in a Flash

However, with the arrival of AI to the world of IT, the time investment needed to create a working prototype of an application (Proof of Concept) based on an idea written down in "normal" human language has been radically reduced.

Developers could start flying. The airlines? OpenAI, Anthropic, Microsoft, Google, Meta…

From my perspective, the time needed to create a PoC has dropped to about 10–20% thanks to AI. No more hunting for tiny code blocks on StackOverflow and stitching them together. Suddenly, you can operate at a higher level of abstraction and describe multi-module, multi-component solutions in plain words. With major time leverage. Usually 😁.

The 40-hour PoC web app (FastAPI Python backend + React frontend) became a single-afternoon task. And yes, that gives you wings. Well, they could be taken away just as quickly—when ChatGPT was overloaded in its early days, you instantly felt how helpless you'd become. On the flip side, what looked like a 20-minute bash script sometimes turned into a three-hour debugging marathon. Maybe you’ve also gotten stuck in a loop where the AI apologises for taking the wrong approach and assertively suggests solution A or B—neither of which works. That’s when you need to just step out and start over. Still, we started talking to AI, and it (mostly) gave us working code—along with tips for improvement.

What many of us have been doing for over two years is now called Vibe Coding. Here we are, just talking to an AI and as an almost magical side-effect an app comes out. Easy and to be honest… a little weird. A bunch of tools and workflows have sprung up to make this smoother, especially by integrating AI into IDEs so we don't wear out Ctrl+C and Ctrl+V keys on our keyboards. At one point, things really took off, and I was updating my dev tool setup almost every other week. But let’s not forget ChatGPT was here first. At least, for me. I still talk to it. It’s a buddy. Held me up in the beginning. It was—and still is—my all-purpose senior IT consultant and consigliere. Sure, there are arguably better tools and workflows for software dev out there now, but ChatGPT is close to my heart. Like Wolfenstein 3D. Plus, it’s mobile. I can build something while waiting for the bus—or in the tub.

Thanks to AI, we can bring our IT ideas to life faster and easier. It streamlines creativity. Let’s take a look at a few little projects that came about because of the ease AI brings to a dev’s working life.

Controlling a Slideshow with Gestures

Idea:

An app that lets you control your slides and zoom your webcam feed from picture-in-picture to fullscreen using gestures. I wanted to try out some open-source gesture recognition library in Python. Also, with all these online meetings and presentations, it just feels right to maintain visual connection with your audience—especially when you want them to focus on something important.

Execution:

About an hour of ChatGPT banter (can't even remember which version it was, ten weeks ago feels like a decade), then a bit of tweaking the gesture recognition math—specifically which region of the webcam feed gets interpreted as a valid gesture, taking into account different distances between the presenter and the camera. Three free ChatGPT image generations later, and voilà— we have digital kittens for the demo presentation.

Push-up Counter App

Idea:

I once heard you need to do something for 10,000 hours to get good at it. Or maybe just repeat it 10,000 times? Tai chi, kung fu, kung pao, and whatever else. How long would it take to do 10,000 push-ups? Sounds like I might need a push-up counter app. 😊

A single HTML file app, super portable—because it runs in any browser. I already knew the name of the library that tracks 33 points on the body (MediaPipe Pose Landmarker), and ChatGPT told me it works with JavaScript too. So, abracadabra—let’s fire up the webcam and make it count my push-ups!

https://ai.google.dev/edge/mediapipe/solutions/vision/pose_landmarker

Implementation:

Roughly 2 hours of conversation with ChatGPT, followed by some refining of the click recognition math—which it almost got right on its own.

How do I talk to ChatGPT?

When we code together, I try to keep the entire code in one file—and keep asking for the full thing every time. If a response misses the mark, I’d rather rework the prompt and ask again than request a fix. After ten back-and-forths max, I grab the code and make a run for it—start a new convo. Same if it drops a "Put your implementation here" in the middle of the script.

And most importantly—be polite! “Please,” “thank you,” a bit of praise now and then. Warm politeness might seem wasted on a cold language model, but hey—you never know. If AI ever takes over, it might help to have a solid track record—I mean, database—of respectful behavior. 😊