How my way of working with AI changed within years

I'm not sure yet if this is a bigger shift in the industry, but I guess it is. I can clearly see a difference in how I work with AI now vs how my work looked 1-2 years ago, not even saying about pre-AI era.

The first attempt — assembling pieces

I can recall first Generative AI things, poor texts and images. I think we all know that. However, I can also remember the first thing I built with AI - it was a small "fun" app for generating random calendar events (I called it Random Nap Generator). I wrote it in Flutter, which I didn't really know at the time. And I was amazed, because it allowed me to use technology I used only once before.

The "workflow" looked like this:

• I described what I wanted in ChatGPT in browser
• it generated me some code
• copied it into VS Code
• something didn't work → back to chat

Repeat. Copy and paste. Repeat.

It worked, but it felt more like assembling pieces than actually building a system. It was interesting, it was cool at this time, but it was also very frustrating at the moment.

The second attempt — less copying, more talking

Later I built a portfolio website. The same one you are currently visiting. It looked different on first try, but this is the same page.

This time I used a CLI tool — Claude Code, Codex, I don't remember now which one exactly. That was the point where:

• I spent less time copying code from chat to IDE
• It felt more like "talking" to AI
• the agent started modifying code directly which was easier to process but was also easier to lose track

But it was still not the perfect, most wanted way. There were still:

• a lot of corrections
• a lot of "no, not like this"
• a lot of iteration

However, I must admit that at this stage I felt that I can ship things faster. Like this portfolio. I always wanted to have one but never found time to do it. And when AI became more popular I was finally able to create my own website. That was nice, I felt like my ideas can become true faster and I can learn new things faster than before... but there was also one problem, and it wasn't code. It was that I didn't know how to tell AI what I want without spending a lot of time fixing its hallucinations.

The shift — SwedenApp

The biggest change happened with my SwedenApp project — an app for organizing a trip to Kungsleden with friends.

For the first time, I didn't start with code.

Knowing that AI can bring ideas to life faster I sat and started thinking how to organize 3 other people to share all info with them without repeating things on messenger. I guess you all know how it usually looks. There is always one person asking about prices, plan and where you all are going. I wanted to avoid it and I wanted to deploy something on Render. And this time I decided to go with agent differently. I decided to make a proper plan to see how it will differ from my previous experiences with creating something from scratch with AI.

Instead:

• I wrote an .md file describing the app
• what it should do
• assumptions
• problems it solves

Then I asked the agent to generate a plan, I reviewed and corrected it, and only then said: "ok, build it."

That's where something changed.

It's not even about speed. It's more that I stopped fighting with individual code pieces and started correcting direction.

When something was wrong, it was usually bad planning or wrong assumptions — not a missing semicolon. It was clear that if the plan was good and everything was described properly the code generated "itself".

What actually shifted

Before:

• I started from code
• reacted to errors

Now:

• I start from description
• try to anticipate problems

The work didn't disappear. It moved. It's still surprising to me, because when I started in 2018 every code line I commited was written by me. Debugging was sometimes hell. And there were a lot of boring things like writing simple tests, fixing CSS styles keeping that one pixel or generating simple CRUD. Now I write even more than before but it is not a code. I write more instructions. I spend much more time describing a problem, discussing with the chat about edge cases and trying to understand system before anything is created. It is sometimes scary, sometimes unreal. Looking back I can see huge difference.

... but I also see the weak spots.

The uncomfortable part

It's easy to lose control if you don't manage context or rush too much. AI can produce something technically correct but completely wrong for the system. The output looks fine. The direction is off.

I had a few situations when I decided to go too fast and then I had to scrap everything, or one decision made too fast broke a lot of forms in this little app for trip planning. I found issues when I was doing manual testing. Of course I am not happy to say that, but that's true. Every time I put too much trust into AI I quickly regret it.

That's a different kind of problem than a broken test. It's harder to catch late.

I don't know yet

I'm still not sure what to think about all this. On one hand — it's more convenient, less mechanical work. On the other — it demands more precision upfront. More thinking before typing. I don't know if this is a permanent shift or just a phase.

And that AI didn't just make me faster. It moved the work somewhere else.

Still figuring this out. If your workflow changed in a similar (or completely different) way — drop a comment, I'm curious.