In the beginning, working with AI went really well. The codebase was small, the AI delivered, everything worked. But because nobody told me anything about maintainability and refactoring, I ended up with 3 files containing 2,000 to 4,000 lines of code each.
For a human, that’s slow. For an AI, that’s poison.
Because when it has to read that much code and can’t find its way around, the context is gone. Bug hunting starts from scratch. Then it goes in circles, trying to fix something the same way it already tried. And again. And again.
That was the moment I understood: the AI is only as good as the structure you give it.
So I built a system.
Notion is my brain. Everything lives there. Roadmaps, feature specs, architecture decisions. Before I touch a single line of code, the plan exists. I develop it with Claude in chat. It thinks along, asks questions, and writes out the tasks I then work through.
Claude Code does the implementation. And here’s what I learned: just talk normally. Plain language, like explaining something to a colleague. No elaborate prompt. 95% of the time you don’t need to steer further. And if something’s unclear, I just write: “Ask me.” And it does.
For bigger things: turn on plan mode. Let it plan first, then implement.
For nasty bugs: run multiple agents in parallel. And sometimes you have to stop it. Because otherwise it keeps putting band-aids on a bug instead of solving the actual problem. Then it’s: stop. 2 steps back. Start over.
And the most important thing: .md files in the project that work like skills. They describe how the project is structured, what the known issues are, which patterns we use. Claude Code reads these before every task. That changed everything. Because the context no longer gets lost.
The image I have in my head: I’m the conductor. Claude Code is the musician. Claude Chat is the advisor. Notion is the score.
But the conductor needs to know what’s supposed to be played. Otherwise it’s chaos.
By the way: I don’t use ChatGPT for coding anymore. My wife uses it every day for family planning and everyday stuff and it’s great for that. But building software productively? Different level entirely.
Martin