Writing
Notes on engineering practice — systems, operations, the work of leading teams that ship.
A few months ago I built shipyard — an autonomous engineering loop — because I was the bottleneck in building my app. Over the past year the whole industry converged on the exact same architecture, and half of what I'd hand-built quietly became native. Here's what that convergence taught me, the one problem it revealed nobody had solved yet, and how I just changed my loop to fix it.
I moved my CI onto my own Mac and it worked — then it started getting slower and flakier by the day. My automated triage blamed resource saturation and proposed three fixes that would each slow CI down or cost money. But the Mac has 64GB of RAM and never drops below 16GB free. The numbers told a different story: high load, idle CPU, zero swap. It was never out of resources.
Last month I killed cloud CI and rebuilt the pipeline as local scripts. It worked, but it was a second copy of my pipeline I had to keep faithful by hand. The better idea: register my Mac as a self-hosted GitHub Actions runner and run the exact same workflows, unchanged. Same checks, same merge gate — for about $0.
My GitHub Actions bill hit $489.85 in a single month — to check a pre-launch app with zero users. So I disabled the workflows and moved the pipeline onto local scripts run by a git hook, with a coding agent as the operator. New marginal cost: about zero.
Two commands before work: /audit to fill the backlog, /do-work to burn it down. By evening, 116 merged pull requests and 121 closed issues. The story of shipyard — an autonomous engineering loop built on Claude Code — and Lightwork, the app it helped ship. Both nights-and-weekends side projects.
A short note on what this site is for, the stack behind it, and what I plan to write about next.