November 18, 2025
I’m on a plane to New York, with Claude Code open on one side of the iPad split screen and Jane Austen on the other (finally reading the Vintage Classics edition of Persuasion with the Brandon Taylor introduction, which … go read the introduction and then reread the novel; I’ll wait).
WiFI is spotty. ssh is happy on the iPad at the moment but my laptop knoweth not the Internet. So I’m thinking about this jagged frontier of AI code generation while reading about Captain Wentworth and Anne Elliott meeting again after eight years, all hopes uncertain.
This is both a more productive and less productive way to write code than ever before. I am finding that many things are much easier to do in a codebase than I expect, with the help of my little LLM friends. Rewrite that test right now during a port, rather than temporarily mark it as “skip?” Sure! Tackle a big project while doing other things in parallel? Oh, yes.
But some things are much harder than I expect. Consistently applying common patterns across a large test suite, for example. And these unexpectedly hard things eat up hours and hours, because by the time I realize we are going in circles, the LLM and I, we are well and truly lost in the woods. I don’t know the code the LLM has written well enough to pick apart the hallucinations and misapplied patterns and properly reset the context. So we stumble around until I can start to separate the confidently wrong guesses from the actual architecture explanations.
I don’t even have a real editor on this iPad, so I’m really living in the future, in this tin can 30,000 feet up in the air. But the basic process is pretty much the same even when I do have Windsurf and Emacs both open (yes, really) next to the multiple Claude Code terminals on my desktop machine at home.
Some of this is a skill issue on my part (I’m slowly getting better at using these tools). Some, a skill issue on the LLM side (the models and harnesses will continue to improve).
It’s also, though, a new normal. Pair programming with my little LLM friends is not the same as writing the code myself. I have to learn a bunch of new things and unlearn a few old ones. Like Anne and Wentworth.
