Like many tech professionals, I’ve experimented with multiple AI coding agents for both work and personal projects, across different stacks and different types of requirements. Overall, I believe these tools are game changers for productivity, and teams or companies that properly integrate them into their workflows can gain significant advantages in terms of time to market, idea testing, and even solution robustness.
For example, in one of my hobby projects, the LanguageMap, where I’m not particularly proficient with the tech stack, GitHub Copilot helped me implement four new features and resolve multiple small bugs in just an afternoon. This would have taken me at least a week of work and research when I first started. Similarly, Claude Code, though more expensive, has helped me analyze a large codebase along with multiple bug tickets, and then suggested very good fixes for many of them, saving at least a day of boring work.
This might not sound like much, but in general the impact on time to market and idea iteration is immense. Let’s take for example the indie makers’ community. Around March of this year, inspired by a flight simulator game vibe coded and released by Pieter Levels, many developers began creating their own games. These games had different styles and genres, but none of them would have been a trivial project for people without game development experience. Nevertheless, many delivered playable games quickly by simply vibe coding with tools like Cursor or Windsurf and iterating until their games worked. Was their code robust and high quality? Probably not, but they still managed to deliver functional products.
This is the aspect of the current AI state I am most interested in, the ability to quickly tackle projects, problems, and ideas that many of us have had for years. We no longer need to struggle for days or weeks to take the first steps and see some progress. We can develop a working prototype, somewhat patchy, in a fraction of the time, and then build upon it if we believe the concept is worth it.
But of course, there are issues with this “vibe coding” dream. Just prompting the AI without having basic software development knowledge to critically analyze the produced code only leads to solutions that are suboptimal at best and often full of bugs, ignored edge cases, are non-scalable or simple change completely from prompt to prompt. I think of it as code a junior developer with little experience might write if given enough time.
I believe experienced developers can profit the most of these agents, but to do so they need to take control and approach their use in a planned and systematic way, defining rules for the AI to follow, establishing code standards, writing clear and very detailed prompts of what is desired and expected, and reviewing each suggesting and requesting corrections based on their knowledge of the application. This way the generated code is way closer to meet production standards.
In practice, this approach will create a kind of partnership where AI takes care of generating boilerplate code, implementing standard algorithms, converting specifications to initial implementations, refactoring repetitive patterns, and creating basic test cases. And this will leave developers with more bandwidth to focus on the high-level architecture design, algorithm selection, business logic details, and complex test cases.
