Six hours. That is how long AI needed to build my whole side project. A Node.js API with 150 endpoints, unit tests, end-to-end tests, a Flutter app for iOS and Android, an admin panel, and a marketing page. More than 100k lines of code. I was able to run everything.
For a long time now I have been seeing companies and people use AI more and more. They say AI writes entire projects, burning a huge amount of tokens, and that they do not need developers anymore. I have been using AI every day for the last three years or more, and I am not sure anymore if that is true, or if it is the right way to code. So I decided to test it. This was not a real product for anyone, only a test, because I was curious if it can really work the way people say.
The plan I gave it
I have a lot of ideas for side projects. I picked one of them and wrote down everything that was in my head: all the ideas, the entire plan, everything. I used Claude Code with Opus 4.8, and Codex 5.6 with multiple models via OpenRouter, to review it. Out of that came a huge detailed plan and a long list of tasks for AI to implement step by step. Every task was fully described, with examples. I created designs for everything with the Claude Design tool. I also wrote rules and skills to enforce the shape of the final project.
I will not lie: I did not review all of it. I tried, but it was too much to read, so I went through it briefly and then asked AI some questions to verify that everything was correct.
When I decided the plan and the tasks were ready, I ran Claude Opus 4.8 on a server and told it to implement everything. Then I waited.
Six hours of my weekends
At first I was happy with it. If I wanted to write this manually, it would take me one or two years, or even more, because I usually code on weekends and only for two to four hours.
That is everything I liked.
What I found when I started testing
Every screen of the mobile app had UI/UX issues and bugs. On some screens the content was rendered outside of the screen. In some places the app sent incorrect data to the API. It also added things I never asked for, like extra authentication methods that were not in my plan at all, and it built forms where the user has to type the country and the city by hand instead of picking them from a list.
It is kind of a black box to me now. I know how it should work, but I do not know the code, and I am not sure it works the way I wanted.
Code I would not have written
The more I read, the more I saw how overcomplicated the solutions and the code are. Sure, it may work, but a lot of it could be simplified, improved, or replaced with an open source library. It wrote the OpenAPI specification by hand instead of generating it from the code. It also used Tailwind for the frontend, because I forgot to write in the plan that I wanted to use pure CSS.
And maybe, if I wrote this project from scratch, it would not be 100k lines at all. Maybe I would build it in a more efficient way, reuse existing tools instead of writing my own, and make different decisions along the way to keep the whole app simpler. When you write the code yourself, you see those chances while you work. AI does not do that. It just writes what I asked for, and it writes a lot of it.
So now I wonder how much time I will need to review, test, and fix everything. If it takes me one to three months, then maybe that is OK, because I could not write this project in that time. But what if it takes me longer? Six months? A year? What if there are issues I do not see yet?
Too many decisions I did not make
The bigger the task, the worse the results. Now I wonder if all those articles, blog posts, and social media posts are just marketing and clickbait.
The more I trust AI and do not review, the more decisions it makes for me, and those decisions are not always what I want. That is the part that bothers me the most. Giving AI too much power over the decisions makes the app too complex. Every decision I did not make is one more thing I have to read later and probably simplify. And it does not always use the rules I wrote. I have them in the agent files, but sometimes AI still goes its own way.
Maybe it is my skill issue. Maybe I have the wrong AI skills or rules. Maybe I do not have access to the best models. Maybe I am just doing something wrong.
Small tasks and real reviews
I am starting to think that maybe I should still review the code, split the work into small tasks, and approve every AI decision. Like before the AI era, but now I can use AI agents instead of real people in my team.
My plan now is to build a workflow I trust, where multiple AI models work as agents to plan, review, and code, and the final review and the merge decision stay with me. I will share my results.
AI can still help me deliver work faster. But after six hours and 100k lines, I am not sure anymore that I can trust it.