1. Note 06 · 6 Oct 2026Everything Screenshots Can't Show Is Still My JobEverything I ship for someone else gets audited before launch. My own site, built with AI assistance, waited longer than I'd normally let it, so this week I gave it the same audit. It found thirteen issues, one of which could have run script on the page, and all of them are now fixed. This note covers what I found and the checklist I'd give anyone who builds this way.Read →
2. Note 05 · 2 Oct 2026Memory Needs a Way OutFor months, every assistant I work with was framing advice around a move that wasn't happening, because the memory that feeds them had no way to forget. This post covers the one-night fix: what was broken, how bad it was, and how a small yes/no model, a link graph, and a few keyboard shortcuts turned "stale" from a feeling into a queue.Read →
3. Note 04 · 1 Oct 2026Sovereignty Is an Architecture DecisionDigital sovereignty gets talked about as a stance. I treat it as a set of boundaries you can point to in a diff: which data never leaves the machine, which jobs go to a rented GPU, and which actions need my hand on them. What follows are the boundaries I've drawn for my own setup.Read →
4. Note 03 · 24 Sep 2026Memory Is a Pipeline, Not a DatabaseThe obvious way to give an assistant memory is a database it writes to as it learns. I built the opposite: a nightly pipeline that distills my own notes into a layer the assistant can read but never write. Below is why I made that call and what the system looks like.Read →
5. Note 02 · 17 Sep 2026Procedures Are the New SoftwareMost of what I ship now is a written procedure that a model executes, with scripts at exactly the points where judgment isn’t allowed. This post covers how I build them and where they fail.Read →
6. Note 01 · 9 Sep 2026The Model Is the Least Important PartI added a second model to the environment I work in every day. The change touched two folders. Everything that makes the system useful lives outside the model, and that's on purpose.Read →