I’m writing this while supervising Claude desktop as it writes out a backend and frontend for a SaaS tool, all based on a simple concept.
Pasted in 10-ish paragraphs of product design and some rough bullet points about data model, stack, screens, and functionality. Then pressed enter and let it rip.
It’s doing its work by creating directories, running commands (eg “npx create-next-app”) and writing files to the file system.
So far it’s bootstrapped an entire python backend with all the right boilerplate, run create-next-app and filled in a full frontend with dozens of components and every page I asked for. I’m reading every line of code as it’s written. It is surreal watching the app magically appear, file by file, in the vscode sidebar.
Everything it’s doing has been according to best practices as of 2025. In an hour of approving tool usages and typing “Continue” into Claude Desktop, it one-shotted an app that could stand a decent chance at a hackathon.
It didn’t do everything in one shot: I had to ask it to make the test suite and had to ask it to replace mock calls in the frontend with real calls to the backend. It trips over when it generates large files, but happily splits them into smaller ones with the prompt “Continue, but split it into several files”.
Claude makes some really nice choices. This is making me seriously reconsider my 6-week-old Cursor subscription. This was all done using just Claude Desktop on macOS.