Research and scientific computing rely on notebooks because they often consist of complex sequences of one-off variations, and demand both flexibility and reproducibility. But my experience using Cursor and Claude Code with Jupyter has been quite poor.

What I've found here is that creating a "notebook", naively, as a parameterized markdown file that is instantiated upon each execution, worked much better than I had expected it to, and I'm interested in figuring out what the limits of this approach are. I would welcome any ideas or feedback!