Writing
Long-form posts on AI workflow engineering, taste in AI-assisted development, and the craft of shipping real systems.
Published
What Persists When the Model Forgets
I asked Claude to find every hole in how I was trying to make truth mechanical between us. It found seven. Then I asked if it remembered the conversation that led there. It didn't, and its answer to that was the actual proof the work had held.
7 min readThe Asymmetric Cost of Being Wrong
Every model-routing system treats over-spending and under-spending as symmetric failures. Neurotoken treats them as fundamentally different.
6 min readThe Operator Problem
What breaks at agent N+1, and why every senior engineer wiring up more than one AI agent has felt the gap without naming it.
8 min readPheromones for AI Agents
Agent context is the scarcest thing I work with. So I taught my agents to read pheromones: a 1-token code that decodes into a full protocol, the way an ant's antenna turns one molecule into a fixed action. Here is the research behind GBuild, including where it works and where it breaks.
8 min readPuber: Built for the Long Miles
An app that connects 100-mile trail runners with pacers and crew, plus a per-race coordination board. The whole thing's built: mobile app, ops console, backend. Next stop: iOS and Android.
9 min readJohnson Power Rankings: Building a Ranking System With My Dad
There's no credible ranking for 200-plus-mile ultrarunners. My dad knows the sport; I know the stack. We built one in a weekend. It's in preview now.
10 min readGBuild: One Orchestrator, Many Agents
The build environment I've been reaching for: a macOS app that runs a fleet of AI coding agents and lets me switch models per task. Marketing site's live; the desktop app is in private development.
9 min readBuilding What I Keep Reaching For
An update on the multi-model app I've started building, and three goals memorialized for later accountability.
3 min read