In Part 3, I described my early workflow: copy JSON from Claude into n8n, watch it fail, copy the error message back, get a revised approach, paste it in, fail differently, repeat.

It worked. It was also exhausting.

Claude and ChatGPT lived in separate windows. n8n lived in a browser tab. My terminal sat in the background, mostly unused. Every iteration meant switching windows, selecting text, copying, pasting, context-switching. The AI was helpful, but it felt like passing notes in class.

Somewhere in the middle of this project, that changed.

I’d known about Claude Code and MCP connectors for months, dabbling in tutorials. Model Context Protocol is a way to give Claude direct access to external tools and data sources. Sounded interesting. Never had the right use case. Building RoleFinder was the right use case.

Now my setup is three windows. First: a terminal for shell commands, git, the occasional emacs session. Second: Claude Code with the n8n MCP connector, giving it direct visibility into my workflows, data tables, node definitions, and build functions. Third: the n8n web interface for visual editing and watching executions run.

The difference isn’t just efficiency. It’s the nature of the interaction.

Before, I was translating. Describe the problem in English. Get code back. Paste it somewhere else. Hope the context carried over. Every round trip lost something.

Now, Claude can see what I’m working on. When I ask about a node that’s misbehaving, it can look at the node. When I want to update documentation, it can read the current version. The conversation has shared context. Less translation, more collaboration.

And the muscle memory came back. I hadn’t done serious command-line work in years. But git branching, diffing, merging — it’s all still there. Rusty, but there.

I still approve every change Claude suggests. No auto-accept. If I don’t understand it, it doesn’t land. But the changes are better now, because Claude understands more of what I’m actually building.

Next up: what building taught me that workshops couldn’t.