// dynaum.blog · notes on building with AI
Building things with AI,
written down.
Field notes from shipping software and games where the model is a real collaborator, not a demo reel. The specs, the decisions, and what actually held up.
kubeside: The Refusal Is the ProductA Kubernetes client scoped to the developer who ships the app, not the operator who runs the cluster. It answers four questions and refuses everything else, and the refusing is the part worth talking about.
read →
-
How the Day Actually Runs
A real day building and running software with AI: predefined user personas aim every agent at the hardest person, parallel agents in worktrees do the building, and loops walk production all night as those same personas.
-
The Model Was Never the Bottleneck
You gave everyone Claude and Copilot and the team ships at the same speed. The tool did its job. It made writing code nearly free and left every slower cost, the ones you never timed, sitting right where they always were.
-
tmux Still Works. I Switched Anyway.
Nothing broke. My work changed shape, the tool did not, and after years of the same setup I wanted to find out if I was still able to move.
-
When Do I Wake Up, and When Am I Done?
Every unattended run answers two questions. /loop answers one, /goal answers the other, and the trouble starts when you make one fake the job of both.
-
She Gave Me a Paper Form. I Built an App.
A doctor asked me to track my blood pressure on paper for a few weeks. I built the app instead. It took an afternoon, and it does the one job paper had, better.
-
Training Is Over Before You Arrive
The model never learns from your chats. It arrived finished. What feels like learning inside a session is context, and it dies when you close the tab.
