I'm giving up ipynb files 🙅‍♂️

The era is over, AI-first development dictates its own rules, and old habits are starting to cost too much.

Feeding raw notebook JSON to an LLM is expensive. The context window gets clogged with junk: metadata, execution_count, cell_type, and endless brackets. The model wastes tokens parsing the structure instead of thinking about the logic.
And generating a valid .ipynb from scratch is much harder for most models.

And versioning — the old pain that hasn't gone away. One accidental cell execution changes metadata, and suddenly you have 500 lines of changes in the diff, even though you just fixed a typo.

Starting today, I'm getting used to the jupytext format, so that regular py files are displayed as notebooks and, if necessary, synced with ipynb, which I will only use for demonstrations.