It's actually very rare to have to innovate, starting from scratch. So whatever problems I need to solve, my first step is always looking for information. And more often than not, the solution is right there.

I don't spend much time at the keyboard. Most of my time is spent perusing books, articles, blog posts, and source code. Or talking to a domain expert. That's what will gave me the system design/architecture for what I want to build. Once that's done, it's pretty much a straight line, or better a spiral iteratively building and testing, going deep only when things matters.

This approach clashes with AI (at least from my POV), because they can't give you a coherent overview of a domain. It's always piecewise elements that are probably wrong in the problem's context, or just plain wrong.

I much prefer to go from general/introductory materials to very specific/advanced ones that matters with sources I trust.

As for the code, it's starting small, get it to work well, and then start on the next iteration. And thanks to the kindness of people, there's so many opensource code from which to learn from.