The AI is a very capable builder. It is not a systems thinker. That line is real.

The project is FWD Path, a home services automation platform I'm building solo.

The core is a suite of n8n workflows covering the full business development cycle: discovering prospects through Google Places, scoring their online presence, generating outreach emails, auditing Google Business Profiles, sending proposals, following up. Nineteen production workflows running right now. I built all of them using Claude Code, Anthropic's CLI tool that lets you work with an AI directly in your terminal alongside your actual codebase.

AI handled more than I expected. Writing the workflow JSON, wiring up API calls, building error handling, structuring data transforms: I'd describe the task and it would produce working code. For self-contained logic, the output was often right on the first pass.

What that means for me

What I still had to own completely was the architecture. How data flows between systems, what lives in Notion versus GitHub, when to use a webhook versus a form trigger, how to structure the CRM so it doesn't create duplicate records: none of that came from the AI. I had to make those calls, document them, brief the AI on the decisions, and then it could execute.

The thing that changed the scale was Ralph Loop, an autonomous agent mode where Claude Code picks up a task list, implements changes, opens a pull request, watches CI, and merges when it passes. I hand it a phase of work and it runs. Not perfectly, but enough that I can be doing something else while it's executing. That pattern didn't exist two years ago. You would have needed a developer sitting at a keyboard doing all of that.

The reality of the work

19 integrated workflows, a live discovery pipeline, a content engine, a CRM with automated scoring and outreach: that's not a solo project under normal conditions. That's a small team. Without AI I would have built maybe three or four of these and stopped because the maintenance overhead alone would have killed it.

But none of it runs without structure. Every change goes through a branch, a pull request, and CI validation. There's a linter that checks workflow JSON before it deploys. There are skills files: markdown documents that tell Claude exactly how to approach specific types of work. The AI isn't improvising. It's operating inside a defined process.