Pattern recognition gets you started. It doesn’t get you into production.

In my last post, I talked about how n8n clicked immediately because the mental model matched decades of systems thinking. That’s true. But "seeing the shape" and "making it work" are very different things. The first working workflow took about ten iterations. Maybe more. I lost count.

Here’s what "building with AI" actually looked like: copy JSON from Claude into n8n. Run it. Watch it fail. Copy the input data and error message back to Claude. Get a revised approach. Paste it in. Run it again. Fail differently. Repeat.

I was also, in those early days, trying to code against applicant tracking systems directly — Greenhouse, Lever, Workday. Each one has its own API, its own quirks, its own parameters. I got one working. Then three. And as I progressed I thought about the other dozen or more I’d need to cover.

That’s when I found Apify: a web scraping and automation platform already built, with connectors to over forty ATS platforms. Already maintained. Already handling the edge cases I hadn’t even discovered yet. The relief was immediate. I didn’t need to build the infrastructure. I needed to build the thing that used the infrastructure.

But even with Apify in place, the iteration grind continued. Different problems — data transformation, workflow logic, error handling — but still the same loop. Try, fail, learn, adjust.

There’s a version of the AI story where you vibe code what you want and it just… works. Maybe. That’s not this story. This is the version where AI is a tireless collaborator who doesn’t get frustrated when you come back for the eleventh attempt.

Next up: what it felt like when I stopped building plumbing and started building the actual system.