Last week I finished a custom e-commerce project on my own. My biggest takeaway was not how fast AI writes code, but how it filled the gaps in what I knew.
I used Claude Opus 4.6.
Before preparing a quote, I turned the client conversations into text. AI reviewed them from the user’s point of view and found the questions nobody had answered yet: products, markets, payments, variants and returns.
I am an introvert, and I can spend far too long worrying over a single reply. Messages drafted with AI kept the project moving at the moments when I would otherwise have overthought the conversation.
Stripe was the clearest example.
I had never integrated payments before. The documentation and the forms made me anxious enough to lose sleep. I asked AI to research the process, explain it to me and list the documents I needed. After a few rounds, the work stopped feeling impossible.
AI also surfaced details that are easy to miss. The client dropped PayPal because it was not widely used in their market, ruled out cash on delivery because of the return risk, and put installments off because of entity and licensing requirements.
The role of code became more subtle.
I initialized the repository on April 8. By April 15 the test site was deployed. The project grew to roughly 48,000 lines of code and 123 commits.
Before any code was written, several AI expert roles turned the requirements into Markdown documents, one for each domain. I reviewed them and then started vibecoding. AI generated most of the code; I checked the business flows and the results.
The first version looked complete while some features were still empty underneath. I treated it as a draft and worked through the site one area and one flow at a time, while AI fixed each problem I found.
Deployment followed the same loop: configure DNS and environment variables, build the Make workflow, push an image, deploy with Compose, inspect, fix, and push again. Fixing problems was not separate from development. It was the development.
The process was iterative. AI produced a rough version, and I kept refining it until the details were right.
Going live was not the end. Users still found small issues. Most were cheap to fix, but they needed a quick response.
Response time, code quality, a low-friction process and being there when something breaks now matter more to me than the first implementation on its own.
For someone building alone, AI can be a strong assistant. Using it well still takes time: enough time to learn how it behaves, where it helps, and where I have to take control.