Here’s the thing about building with AI: it’s fast. Maybe too fast.

Claude can generate working code in seconds. n8n workflows that would have taken me hours to figure out — done in minutes. But speed has a cost. If you’re not careful, you end up with a working system you don’t actually understand.

I’m not a JavaScript programmer. I’ve worked alongside engineers for decades, so I know what good code looks like, but I’m not writing it from scratch myself. So when Claude generated the workflow logic, I could trust it and move on — or slow down and make sure I understood what I was shipping. I chose slow.

Once each workflow was running, I did something that might seem old-fashioned: I fed the working JSON back into Claude and asked it to comment the code as if explaining it to a junior developer. Not because I was going to hand it off, but because I needed to be able to understand it myself.

This isn’t a magic trick. It’s discipline. AI can generate so much faster than you can comprehend. The only way to stay connected to your own system is to force the comprehension step.

And it didn’t stop at code comments. After documenting the nodes, I went back through my README and operating guide using Claude Code and ChatGPT together — Claude for the heavy editing, ChatGPT as a second opinion to catch blind spots before open-source release. Then a final pass in a markdown editor, because the last eyes on it should be human.

Every update Claude Code suggested? I approved each one individually. No auto-accept. If I don’t understand the change, it doesn’t land.

The interesting learning: pair programming isn’t just for code. It works for documentation too.

Next up: when the architecture broke, and why that was a good thing.