The article was written in August 2025, so my experiments are relevant for this period. At that time, gpt-5 was released.

At the beginning of August, I arrived in Aktau and stayed with friends. They are also IT specialists and often use GPT in their work. Naturally, I was inspired by their inspiring stories about the wonders of working with neural networks and decided to try it myself.

I'll write my conclusion right away, if you're too lazy to read this article, my conclusion is this: neural networks can really do very cool things in your work, but only if you yourself are very good in your field:
1. You know what you want and understand how it should work;
2. You can formulate your wishes and wrap them in a promt (technical specifications for a social network);
3. You know how to test the work of a neural network, and are also ready to refine your technical specifications and achieve the desired result from the neural network.

My experiments - how to use ChatGPT in work

1. Oddly enough, but the first experiment was with the neural network manus.im - this neural network can make websites. We tried to print a promt and launch it. Manus made a simple landing page for services with an imperfect, but acceptable design. Then I asked her to make a site similar to my own (which is very different in design). Imagine my surprise when the network eventually issued a site with exactly the same design as the very first one. A request to change it did not change the situation dramatically. This option does not suit me. Failure.

2. I asked ChatGPT to expand the image by drawing the background so that the square image would turn into a 3:2 format without losing information in the center. The picture expanded, but all the texts on it turned into mysterious hieroglyphs. Failure.

3. I had semantics for 1200 keywords for contextual advertising. I asked to group words by type and issue a text file as a result.
Attempt-1: 1200 words turned into only 200, some of which were invented by gpt;
Attempt-2: when I asked not to add or delete anything, he issued a list of 400 queries, also adding my own;
Attempt-3: led to the same result;
In addition, when downloading the file with the result, an error appeared, I asked to redo it several times, but in the end I could not download it. Failure.

4. We were making a new site to replace the old one, working on Wix. Or rather, it is no longer working in the Russian Federation. Accordingly, it was necessary to transfer content from the old to the new one. There were large and bulky text articles with a "dirty" wix layout that needed to be reformatted for a classic layout. On the first attempt, gpt lost half of the article, for some reason deciding that it needed to be redone. After several iterations of requests, it did as it should, without losing anything. But at the time of checking and publishing the article on the new site, I found errors - extra tags and unclosed tags. Partial success with modifications.

5. The most difficult task, surprisingly, was solved well. Now it was necessary to download images from the same wix and number them in the order of publication on the old site. The problem is that on wix images are not published explicitly and files in a specific format, and in the control panel they were scattered in different folders and it was also impossible to download them from there in some simple way. In short - we struggled for a long time to create a script for the browser console, which will search, convert, download and rename files. Several hours and iterations of prompts did not give the desired result - ChatGPT could not understand the wix code and extract images from there. At some point, I sent him a part of the html markup, in which the path to the image was encoded, and lo and behold - it worked. Gpt remade the script and I downloaded all the images to the *.jpg format, which is what was required. Success.

So is it worth using artificial intelligence and does ChatGPT really help save time?

Conclusion: I think that neural networks can be very useful, but you can't blindly trust them and all the results need to be double-checked, and for this you need to be in the know, as they say. If you don't treat them as a 100% truth and a miracle, but use them wisely, understanding that the results may not be accurate, then you can quite use them in your work and save your time.