From Cursor to Claude Code: What I built in my first week
Victor is a developer at Gear Train. This post is part of our ongoing series on learning and building with AI.
I had never used Claude before. I was a Cursor user. That changed when I completed four courses on building with AI, and by the end of the first week, I was running sub-agents, writing custom hooks, and building a real project from scratch.
Here is what I learned, and what I built.
The course that changed how I think
The course that stuck with me most was Building with the Claude API. It did not just teach me how to use a tool. It helped me understand the first principles of large language models: how they work, why they behave as they do, and how the pieces fit together.
I learned about tooling functions, hooks, and MCPs (Model Context Protocols). More importantly, I learned how these components connect to create AI-assisted workflows that actually do useful work.
Before this, I was using Cursor - a good tool, but I did not understand what was happening under the hood. Now I do.
I started building before I finished learning
The best way to test what you have learned is to build something real. So I started a church project, a management system where I am using Claude Code as a full development teammate, not just a code completion tool.
I began by using it to understand requirements, build a project roadmap, and design the architecture. Then I moved into implementation, refining reusable skills, creating architecture-specific workflows, and writing custom commands.
That is where it got interesting.
The problem I hit: and how I solved it
As the project grew, the generated code began to introduce errors. Not every time, but often enough to slow me down. I needed a way to catch and fix issues automatically, without interrupting my flow.
So I implemented a postToolUse hook. Every time Claude generates code, the hook automatically triggers a linting skill to detect and fix issues on the spot. The errors stopped accumulating. My feedback loop got tighter.
I also set up sub-agents to generate and run unit tests. This has worked better than I expected. The tests run automatically, the results come back clearly, and I can see exactly what is passing and what needs attention. It has significantly improved the quality of what I am shipping.
What this actually feels like
Working this way is different from anything I have done before. Claude Code is not a tool I use; it is a teammate I direct. I define the work, set the constraints, and let it handle the execution while I focus on the decisions that actually require judgment.
What excites me most is that I am only getting started. My next goal is to build MCP integrations as proof-of-concept projects, tools that connect Claude to external systems and demonstrate what is possible when AI has access to real data and real workflows.
Why you should try this
If you have access to these courses, take them. Do not just read the material — start a project alongside. A small one. Something with a real goal.
The best insights come from applying concepts to something that matters to you. The gaps in your understanding show up quickly, and filling them feels much more rewarding when there is something concrete to show for it.
I went from never using Claude to building with sub-agents and custom hooks in a week. I am not an expert. I am just someone who started building and kept going.
Gear Train is a technology transformation company based in the UK and Kenya. Follow us on LinkedIn for more content:
