I've spent more late nights than I'd like working through AI automations, building systems and trying to turn ideas into something that actually works.
And one thing has become increasingly clear to me.
The quality of an AI build often comes down to the planning that happens before anything gets built.
It's incredibly easy now to open ChatGPT or Codex, describe what you want, and watch something start taking shape within minutes.
But getting AI to produce code is one thing. Getting it to build a reliable system that actually does what you intended is something else entirely.
Product thinking before implementation
Something I've found particularly interesting is how much my background in UI/UX and product design has helped me with this.
I've spent years working with complicated products, understanding how different components and processes connect, and then finding ways to simplify that complexity into something people can actually understand and use.
And I've realised that much of that same thinking translates directly into building with AI.
Before worrying about implementation, I try to break an idea down into smaller, manageable parts. What are we trying to achieve? How should everything connect? What does each component need to do? What happens when something goes wrong?
Because a system can be incredibly complex behind the scenes without needing to feel complicated to the person using it.
That's always been an important part of good product design for me.
And increasingly, I'm finding it's just as important when building AI-powered systems.
Give the tools clear responsibilities
Even the way different AI tools work together needs thought.
When I'm working with ChatGPT and Codex, for example, I want clear responsibilities. Planning, implementation, review and validation should complement each other, rather than constantly pulling the project in different directions.
Otherwise, it's surprisingly easy to find yourself in a cycle where one fix introduces another problem, and you're spending more time correcting work than moving forward.
Security is another consideration. Not every project needs enterprise-level complexity, but understanding what data a system can access, what permissions it has and where the risks sit should be part of the process from the beginning.
Plan enough, then test
Of course, none of this means planning every tiny detail before starting. Experimentation is still important, and sometimes you only discover the right approach once you begin testing.
But I think that's where professional experience becomes particularly valuable.
It's not just about knowing how to use the tools. It's about understanding the problem, connecting the different moving parts and knowing what a successful outcome should actually look like.
AI has made execution incredibly accessible.
But it hasn't removed the need for planning, technical understanding, product thinking or sound judgement.
If anything, I think those skills are becoming more valuable as building becomes easier.
And perhaps that's the biggest lesson I've taken from working more deeply with AI.
The better you understand what you're building and why, the less time you tend to spend untangling it afterwards.
For those building with AI, where are you finding the real challenge now: getting something built, or getting it to work reliably as part of a larger system?