For years my head was home to an excellent PaaS. Next door lived a menu bar utility that keeps an eye on dev daemons. Both were flawless. Neither of them existed. This summer they became real: Krill and DevDeck. In late August Base13 moved in next door too, a remake of Battle City, the old NES tank game.
That was the pattern for years. An idea arrives in the evening, picks up a name and a stack on the walk home, survives until Saturday. And on Saturday it turns out that forty boring things stand between me and the first interesting line of code. Scaffold the project, work out an unfamiliar framework, remember how macOS apps get signed. The idea died on those forty things.
I wanted different things from them. Sometimes I just wanted to write code with no goal at all, the way other people whittle stools in the garage. Sometimes the opposite: a product that somebody other than me would use. Neither survived a Tuesday evening.
The part about not having time was a lie I told myself. I had the same hours then as I do now. What I lacked was the reserve to open an unfamiliar API after a full work day and not quit halfway. My inspiration usually ran out at the first bit of boilerplate.
What changed
Numbers as of November. Krill: 519 commits in its first six weeks, a deploy control plane in a single Go binary. DevDeck: 21,000 lines of Swift in the app itself and over a thousand tests. Base13 has already grown a 35-level campaign and 710 tests, and in network co-op the two machines trade nothing but button presses. All of it solo, on evenings and weekends.
It’s not that “AI writes the code for me”. I’ve got years of backend work behind me and enough pattern matching to tell a design that holds up from one that falls over on the second user. That used to count for little. Testing any hypothesis cost a weekend, so I preferred not to test. The agent killed the price of checking. The experience is still mine, the speed is the agent’s, and together they got things moving that never used to get started.
The caution didn’t go anywhere. An agent multiplies care and the lack of it with equal enthusiasm. The most tortured test in Krill went through 38 rewrites: the agent typed it, I made the call every time. Hence the sticker at the top of this blog: “No slop, just AI”. Slop starts where generated code goes unread.
There will be fewer breakthroughs
The bar for being impressed moved up. When a weekend produces what used to take a quarter, pulling off a real breakthrough only got harder. I’m fine with that. I sit down with a project to understand how it works on the inside. While writing Krill I worked out why “deploy converged” is the most dishonest status in existence. While writing DevDeck I learned far more about orphaned processes and PTY races than I’d planned to. Base13 made me take determinism seriously: the browser and the desktop have to compute the world down to the same hash, or a two-player match falls apart.
There’s a work reason too. Five days a week I’m Head of Backend, and I need to see where the industry is heading before it turns into a question in planning. Reading other people’s threads doesn’t cover it. The only honest way to learn what a new tool can do is to build something with it. So side projects doubled as a proving ground: I try things on my own code first, then decide whether to bring them to the team.
And I got back the first thing that job takes from you. The higher you climb, the further you sit from production: you think in terms of the company and the quarter, the product is somewhere down below, and you find out about a release from someone else’s message in a channel. The joy of shipping quietly disappeared. In my own projects it’s still there: I hit deploy myself and watch whether it lands.
This blog runs on Astro and is deployed by Krill, which I wrote myself. I’ll write up what worked and what broke, one post per project. Let’s go!
■ this post was deployed by the thing this blog writes about
Comments