Writing
Field notes
I gave seventeen model configurations the same 620 words to write and scored them all on one rubric. Thinking mode made them measurably worse at prose, six minutes of reasoning lost to twenty-two seconds, and rewriting the prompt beat every jump in model size. The full scores, failures included.
read it →First principles
My eval said 100 percent. The honest version said 70.8, and the ranking inverted. Six principles, one loop, and a worked example for choosing the models that drive your agents - plus what scoring state exams taught me about trusting an LLM judge.
read it →Field notes
A product team walks in with an idea and walks out, three hours later, with market research, a PRFAQ, a requirements doc, and a clickable prototype. The toolkit drafts each one; the room makes it right.
read it →Essay
Twenty years building software, four times a CTO, and I don’t write the code anymore. I notice friction, ask an AI to solve it, throw most of it away, and keep the few tools worth reusing.
read the essay →Field notes
A real Mac app that thinks like vim, renders like a technical journal, and exports like it works in an office. Nobody sells that, so Claude and I built it in six days - and the interesting parts are the failures.
read the build →Essay
The life and death of software subscriptions, seen from one house: a hand-me-down Mac mini, an AI companion, and the return of fixable things.
read the essay →Technical companion
Headscale on a $10 server, a fleet-native fork of Blink, mosh sessions that survive elevators, and claude -p as the whole RPC protocol.