Notes on Shipping When AI Writes Most of the Code
A few things I’ve learned shipping projects when I write far less of the code than I used to.
The first 20 hours still matter — but differently
The critical window is no longer “can I implement the core loop?” It’s now “can I stay in the loop with the AI long enough to reach something that feels real?”
Many projects die in the first few hours not because the AI can’t build them, but because I lose the plot while directing it.
Scope is still a muscle. AI makes it weaker.
The model will happily implement almost anything you describe. This is dangerous.
I’ve started treating “the model made this easy” as a red flag rather than a green one. Easy implementation is no longer a good signal that something should exist.
Distribution still beats product
This hasn’t changed. If anything, it’s more true. An average AI-built product with an audience will still beat a brilliant one with none.
The difference now is that the marginal cost of building the average product has dropped dramatically.
The new “rewrite everything” urge
I used to want to rewrite things in Rust. Now I want to “just prompt it better” or “switch to a stronger model” or “use a better agent framework.”
Same disease, new costume.
The best projects still feel like gifts
This part is unchanged. The projects I actually finish and maintain are still the ones where I personally feel the pain of not having them.
The AI can reduce the activation energy, but it can’t create the desire.
The rules are shifting faster than I can write them down. That’s the whole point of this blog.