
I Got Tired of Explaining Myself
Why I stopped writing better prompts and started treating an AI like a new member of my team: write it down once instead of explaining it every time.

Why I stopped writing better prompts and started treating an AI like a new member of my team: write it down once instead of explaining it every time.

I ran a Karpathy-style LLM-maintained wiki for six months. What compounded, what didn't, and the 614 notes I archived.

I built a tool to generate writing prompts from my vault, and none of them got me writing. What worked was a prompt built from my own morning pages.

I ran out of Claude tokens and handed my six-repo codebase to Gemini CLI. It made some mistakes and had one genuinely good idea.

Giving an LLM the keys to your wiki is not a dramatic handoff. It looks like maintenance done quietly, completely, every time.

An AI writes instructions for itself, backs up its own memory, and builds a tool to read articles on your behalf. It considers this progress.

Two sessions, two takeaways, and why you should not skip the training day at PyTexas.

The terminal window is a powerful tool that lets you quickly verify what LLMs tell you that they have done

How to run a local LLM from a Python script without an API key, a virtual environment, or Docker. You need two tools and one file.

Karpathy called LLM-based knowledge bases "very useful." Why the architecture works and how to start building one in practice.