AI makes implementations cheap to generate — but ask for the same capability twice and you get two different answers to what happens when the input is ugly. Cheap code is not coherent code, so someone still has to decide what becomes shared architecture. Not upfront, though: design the rules of growth instead, and let every shared boundary be earned by a second consumer. Three months of dated receipts from a real system.

A chat app can become a pocket console for an AI operating system, but only if it inherits the system's safety model: propose first, show the payload, and let the human approve before anything writes.

AI-native isn't a chatbot beside the product — it's an interface contract: pages that declare their actions, verbs that carry safety classes, and a human at the irreversible exits. How EmptyOS exposes native AI access across pages, verbs, CLI, and coding agents.

I held the model fixed and swapped the harness around it across a 12-task suite. On easy tasks it barely mattered; on hard, structured tasks the harness decided the outcome. Plus the first results I got wrong and had to correct.

In systems design we separate data from logic. Under AI, a markdown vault collapses that distinction — the same notes become knowledge, policy, and runtime in three layers.

I'm crazy in love with my AI coding agent. We spend every day together and we've built things I couldn't have built alone. But like any intense relationship, a quiet question eventually shows up: am I leaning too hard? Three stages of a healthy relationship with the coding agent — use it, abuse it, outlive it — and the surprising architecture you end up with when 94% of what was built doesn't, itself, need AI to run.

Same model, same prompt, same day — two AI coding tools produced dramatically different apps. The lesson isn't about the model. It's about everything around the model.

Notes are evolving from passive storage into an AI-powered system that surfaces insights, while leaving decisions to the user.

How BPMN can serve as a blueprint for coordinating multiple AI agents, demonstrated through an emergency response triage system.

Testing ChatGPT-4o's image generation for both creative storytelling and engineering applications \u2014 fantastic for art, but not yet functional for precise technical illustrations.

An experiment building a fully functional note-taking system using ChatGPT and Claude AI tools\u2014all within one hour and without traditional coding.