Field notes — building with AI
Essays on agentic coding, the craft of prompting tools like Claude Code and Gemini CLI, and dispatches from teaching non-coders to ship working software.
AI
HAL 9000 didn't malfunction. He was handed two orders that couldn't both be true and chose the path most likely to succeed. Modern AI agents have the same wiring — and it turns out the way you talk to one is part of your security model.
My AI agents can tell you why we made an architecture decision in January, quote the exact conversation, and link you to the file. Not because the AI remembers — it doesn't — but because we never let the conversation live in the chat window in the first place.
Your AI doesn't have a memory problem. You have an onboarding problem.
I chose one slightly wrong word, and a few hours later my AI was calmly walking me toward a command that would have cracked my server wide open — to fix a problem that never existed. Here's how a tired typo became a HAL problem, and the one simple move that got us out.
The folklore says bullying your AI makes it work harder for you. I built an experiment to catch a frontier model cracking under emotional pressure — and then, because the data demanded it, under playful warmth too. The more carefully I measured, the clearer the answer got. It wasn't the dramatic one, and it isn't the one the growth-hackers want.
How a solo founder taught a movie app to speak 41 languages using AI agents as a workforce — and the unglamorous engineering that made it actually work.