Post

I feel like over the years, we've sort of given up documenting our mental models of code and just accepted that everyone who reads the code builds their own model from scratch.

Now, with vibe coding we don't even do that. The agent is often the only one with a mental model of the code, and it throws it away after each session.

Only document. Double down on documents. And insist on them being faithful. The code coming with a perfect textbook intro to it, and the code perfectly implementing this textbook.

I mostly agree but Naur is arguing that there’s no such thing as a perfectly faithful spec either. We need to draw the line somewhere. But I’m optimistic that language models help us get a lot more of the details out of our heads and into the specs.

0 reposts1 like

For larger projects I'm a big fan of trying to distill and verbalize the philosophy of the code. But I'm only starting my current project haha, so time will tell if this "philosophy first" approach is reasonable 🤪