How I work with AI

Working well with AI is now a basic part of software development. I was lucky to learn to code just before AI took off. It would have been easy to skip that and go straight to AI coding, without seeing how much the foundation matters.

A lot of what makes a good developer is not only the code. It is talking to users, planning, and thinking about what comes next. Even as AI takes over more of the coding, those parts will likely stay.

Much of the real work is finding the balance between plain code and AI models. A model has to earn its place. Often plain, testable code is the better tool.

I think a lot of what comes next is smaller, specialised models running locally, and narrow agents that each do one job well. A lean agent that only loads what it needs tends to beat a big generalist. I build my own agents that way.

Models make mistakes, so tests, monitoring and guardrails matter as much as the features. I have started running changes to my own projects through automated checks and an AI review before they are merged. Agents only get the access they need, and they ask before anything risky.

Clients often expect an AI project to be quick and simple. In practice it rarely is. AI helps a lot during development, but a robust result still needs planning, specs, tests and the right model for the job. Testing by people matters as much as it ever did.

For teams, the hard part is not the models. It is shared ways of working with coding agents, such as shared memory, notes and setup, handled the same way by everyone.

It is tempting to chase every new model and tool. Better models do matter, and it is worth keeping up with what happens in the field. But the real foundation is understanding how AI works, what it can and cannot do, and how to use the tools well. That knowledge takes you a long way.

Some people say AI kills creativity. For me it is the opposite. It makes me think in new directions. The real challenge is to not get overwhelmed by everything new.