I handed our intern a piece of work yesterday that, a year ago, would have taken me two weeks to build on my own. He handed it back to me this morning. I spent an hour reviewing it, made a couple of small cosmetic changes, and deployed it. The code looked as if I had written it. Everything was well-named, properly typed, and fully tested. Code so clean you could eat off of it.
Darius isn't a savant who's picked up 20 years' worth of coding experience in two months. He certainly didn't vibe code it. Yesterday, I handed him a fully formed plan for exactly what I wanted built, complete with all of the standards and conventions I wanted used, and he supervised a set of Claude agents as they implemented the plan almost perfectly.
The actual time it took Claude to generate the code was a couple of hours, but that's not how long it took to build the feature. I spent a full week creating the specification. It was a painstaking, slow process that involved me exporting Markdown to PDF and sending it to my Kindle 6 times. By the time I was done, there weren't any known unknowns or unknown unknowns. There were no more decisions to be made. My intent was completely and clearly defined in the documentation, broken down into atomic tasks.
When we were responsible for actually typing all of the code into the IDE to bring our vision to life, we didn't have the time to shape that vision in advance to the level of specificity and completeness that we can today. This feature would have taken me two weeks to write myself a year ago, and there's no way I'd have spent the whole first week planning out and documenting every single detail and decision. It would have worked and been as well-tested as what Darius handed me today, but all those little decisions that I would have made along the way while typing the code into PHPStorm would have been lost to time. Now, they're all recorded and ready to reference the next time we write code that touches this feature.
This is the velocity paradox in the age of AI. Our velocity is way up, but we're all slowing down and taking our time. I've been saying for a long time that code generation isn't the interesting or important part of the job, and now that it's been mostly automated, we can spend much more of our time doing the work that delivers real value to our users and clients.