Writing

Notes from the two years before ChatGPT — IQ-testing davinci at 2 a.m., whale facts, an Abe Lincoln chatbot, and a machine that told me it worried it was a toy.

A sonnet generator I remembered from a 1995 Mac book turned out to be uncatalogued anywhere online — so I dug it out of my own disk image, decoded how it works, and put it back on the internet twice.

Notes on a compliance-data pipeline — scraping, normalizing, and scoring 10,000+ establishments.

How nelson.love publishes itself from an Obsidian vault, and what it was like to build that with an AI agent reading the notes over your shoulder.

What the reviewer-recommendation engine taught me about reaching for the legible solution first.

Using models to turn messy test output into clean, queryable data — the boring, useful application.

Why you have five note apps: a three-axis model — object type × verb × register — of the whole space personal-information tools occupy, and the orchestration layer that belongs between them.

The most reliable place to put a model is between two systems that already know what they want.

AI coding agents fail when requirements underdetermine the domain model. A case for a machine-readable domain layer — an intermediate representation between domain knowledge and generated code.