UNIT.City — місце, де люди працюють... КРАЩЕ! Обирай свій простір просто зараз 👉

"This leads to severe degradation and laziness." Ukrainian programmers shared the secrets of effective vibe coding and the "side effects" of using AI when creating code

80% of IT professionals reading dev.ua admit that they use artificial intelligence to write code. The so-called vibecoders thus simplify and speed up their work.

We conducted a small survey and found out in which cases Vibe coding is appropriate and useful, whether IT professionals are satisfied with the quality of code written by artificial intelligence, and how to correctly formulate a prompt for AI so that the result satisfies the requester as much as possible.

Leave a comment
"This leads to severe degradation and laziness." Ukrainian programmers shared the secrets of effective vibe coding and the "side effects" of using AI when creating code

80% of IT professionals reading dev.ua admit that they use artificial intelligence to write code. The so-called vibecoders thus simplify and speed up their work.

We conducted a small survey and found out in which cases Vibe coding is appropriate and useful, whether IT professionals are satisfied with the quality of code written by artificial intelligence, and how to correctly formulate a prompt for AI so that the result satisfies the requester as much as possible.

When IT people use AI to write code

dev.ua asked IT professionals to share when they think it’s appropriate to use artificial intelligence in their work. Here are 10 answers that illustrate the most common uses of Vibe coding.

When you need to perform secondary duties using unfamiliar technologies. For example, edits in ci/cd scripts from a developer. When you need to write a lot of boilerplate code (but code generation is better).
Simple routine tasks. On more difficult tasks, he doesn’t understand the end result, and writing a step-by-step TOR is a different story.
Working with well-structured projects, simple updates in api circles, such as adding new fields or filters. Also writing tests.
Something you are absolutely sure you have already done, and that you can verify. It should be something not very important (not security, for sure).
Frontend tests are written perfectly on it, as are backend queries, but writing code directly is quite risky.
Generating boilerplate code, creating JSON (etc.) strings. You can delegate only very simple and typical operations to AI. It is not worth delegating complex problems, because it often takes longer to rework it than to write it yourself.
Writing code (coding) when there is already a template, solution, architecture, but you are too lazy to write code.

When the task is really small (limited to a few lines or one function), ideally, when you know how to solve it yourself and can clearly say what needs to be done.

It also helps to start with something small and gradually add functionality to it (some edge cases, new conditions, etc.).

AI also helps to optimize a piece of existing code. It is also quite useful for understanding what a piece of code does and how it works.

As a technical writer, I don’t know anything about programming. But I’ve written a few scripts that communicate with AI via API and process data (generate documentation) based on my prompts.

Prototype writing. Propose solutions. Automation tasks. Routine tasks.

"Programmers are becoming process managers, not routine task performers." AI enthusiasts discuss AI-powered vibe coding in Threads — most are skeptical
«Programmers are becoming process managers, not routine task performers.» AIT people discuss AI-powered vibe coding in Threads — most are skeptical
On the topic
«Programmers are becoming process managers, not routine task performers.» AIT people discuss AI-powered vibe coding in Threads — most are skeptical

What’s wrong with Vibe coding?

We also asked whether IT specialists were satisfied with the received AI code, what «sides» does vibe coding have, and what in the generated code should always be checked by a human.

Here is a list of the most interesting verdicts:

For AI code to meet your needs, you need to understand how language models work and their limitations. This is the main «side» of coding vibes, because to write high-quality code using AI, you need to be an experienced developer, which in turn means that AI will not solve many work problems, but about 80% of them it will. It is faster to write it yourself. For me personally, AI is useful only for generating boilerplate code and converting data from one format to another (and this is if such conversion should take place without complex nuances).

Dima Maleev recently wrote a post on LinkedIn that AI does not adhere to any architecture or code style, and also does not know how to generate the same code twice, and I agree with him. I do not believe that AI can write all the code in a project. It is an assistant, but it will not help if you do not fiddle around yourself.

For example, I often come across cases where AI generates outdated code that no longer works with the new version of the library, and cannot fix it — you have to read the changelogs and update it yourself. Also, sooner or later AI will run into some bug or error that it cannot fix and will generate 2-3 different solutions in a circle that will not work — in this case, you have to take the last state in which the code worked and start another chat.

In general, if you generate something with AI, you should understand the code you generated well, test it in different scenarios. I had a case where the generated code could have caused twice as much extra costs.

AI generally does a good job of everyday things, but when it is given something that is not even completely unique, it writes crookedly. As a side note, I notice right away that I forget even elementary things just by doing things with AI. If I don’t know something, I turn to artificial intelligence instead of figuring out how it works. This leads to severe degradation and laziness. I believe that in general, a person should use AI as little as possible when coding, and try to learn the material independently as much as possible.

You always have to redo it, and sometimes this process takes longer than if you were to just tap the satisfaction level with your hands.

In my opinion, 5/10 vibe coding is needed more as a tool to expand existing tests (code), perhaps to embellish the code, to adjust it (cosmetic and at least functional).

This is suitable for a beginner who uses AI as a guide, a sandbox, or, like the same Copilot, but for experienced players who want to improvise a little.

Overall, yes, I’m satisfied. «Bonuses» — if the dialogue is large, the AI ​​is more likely to make mistakes and forget something from the chat, even when it’s still far from the maximum size of the context window.

If there is no clear request, it almost always ends up being a mess.

Satisfied. This made the work 20 times faster.

Dima Maleev on the impact of AI on IT: "In a few weeks we will see more and more "vibe coders" launching startups and receiving millions in funding"
Dima Maleev on the impact of AI on IT: «In a few weeks we will see more and more „vibe coders“ launching startups and receiving millions in funding»
On the topic
Dima Maleev on the impact of AI on IT: «In a few weeks we will see more and more „vibe coders“ launching startups and receiving millions in funding»

How to write AI prompts correctly to make Vibe coding effective

We also asked what prompts for writing code, from the experience of programmers, the AI ​​perceives best and executes most accurately. And here is what they answered us (from the most useful):

  • Chain-of-thought, few-shot prompting, prompt chaining are our everything. Few-shot prompting is especially helpful in coding, if you need to write some algorithm — you describe the algorithm itself and provide a few examples of input and output. Prompt chaining and CoT — when you develop something with the help of AI gradually.
  • It is better to break everything down into small tasks and be sure to check the results with large-context models. And also check it yourself. Testing should be twice as rigorous.
  • The prompt should be precise and exclude the possibility of misinterpretation. You also need to prepare the AI ​​to show all the options (so that you can choose the effective/working one). You also shouldn’t expect the AI ​​to «understand» your context even if you show it to it, so the problem delegated to the AI ​​should be previously isolated by you. And AI doesn’t mess around with math, and you need to remember that.
  • Prompts are a very specific topic. It is important that AI has the context of your application. Because if it doesn’t, you either have to decompose the tasks very much or write huge step-by-step prompts, which already negates all the advantages of code generation.
  • Provide references to the code you want to see in the end result.
"Code generated by AI in its pure form does not make it into the project." CTO SharksCode on Vibe coding, AI assistance and IT development prospects
«Code generated by AI does not enter the project in its pure form.» SharksCode CTO on Vibe coding, AI assistance, and IT development prospects
On the topic
«Code generated by AI does not enter the project in its pure form.» SharksCode CTO on Vibe coding, AI assistance, and IT development prospects
In Ukraine, they are looking for a powerful vibe coder with a salary of up to $12,000 who will be 5–10 times faster than a developer. What do IT professionals think about this “fabulous” vacancy?
In Ukraine, they are looking for a powerful vibe coder with a salary of up to $12,000, who will be 5–10 times faster than a developer. What do IT professionals think about this «fabulous» vacancy?
On the topic
In Ukraine, they are looking for a powerful vibe coder with a salary of up to $12,000, who will be 5–10 times faster than a developer. What do IT professionals think about this «fabulous» vacancy?
“The speed you can get compared to programming yourself is just crazy.” 3 vibe coding tips from top software engineers
«The speed you can get compared to programming yourself is just crazy.» 3 vibe coding tips from top software engineers
On the topic
«The speed you can get compared to programming yourself is just crazy.» 3 vibe coding tips from top software engineers
Read the country's main IT news in our Telegram
Read the country’s main IT news in our Telegram
On the topic
Read the country’s main IT news in our Telegram

Have important news to share? Message our Telegram bot

Key events and useful links in our Telegram channel

Discussion
No comments yet.