Honest notes on AI, product, and learning — mostly PM and building, occasionally Chinese.
7 notes
In July I built a dashboard to run my week from. Seven weeks later it has seven tabs and 564 tests, and I closed exactly three weeks with it. An honest follow-up on the stack, the lessons, and what a trip did to the ritual.
Grant Sanderson's 3Blue1Brown series on neural networks and transformers, translated into product language: why evals matter, why long context costs money, and why facts aren't editable.
Andrej Karpathy's two YouTube lectures on LLMs, translated into product management terms: why hallucinations are structural, why context engineering is a PM discipline, and more.
Eight Claude Code subagents that critique my specs, decks and posts before anyone real does — what they're good for, where they fail, and why loops are the natural next step.
Most writing about LLM evaluation is aimed at engineers. But defining what 'good' means for an AI product is product work — and it's the skill I'd bet on hardest right now.
My todos lived in Google Calendar, my study plan in markdown, my surf forecast in a separate app, and my weekly reviews in my head. So I built Week Board — a local-first dashboard that puts my entire week on one screen.
From a blank chat to a live site on my own domain: brainstorming with Claude, building with Claude Code in VS Code, and shipping with GitHub, Vercel, and a Namecheap domain.