Notes
Shorter pieces on product judgment, methods, and concepts.
You're not using AI. You're organizing it.
The default metaphor for AI is a tool — you pick it up, use it, put it down. But that breaks when you're coordinating multiple agents across days, each with different context, and the real work is keeping state coherent between them.
Why multiple AI roles still produce one opinion
Naming four different reviewer roles in a prompt doesn't create four independent perspectives. They paraphrase each other under different names. What actually changed the output was context isolation — and that took three versions to figure out.
What a failed RAG experiment taught me about AI memory
I built a full RAG pipeline for my personal AI system. The engineering ran perfectly. The answers got worse. The problem wasn't the retrieval quality — it was what "relevance" meant in the first place.
The better AI knows you, the harder it is to surprise you
AI personalization has a shadow: the more a model understands how you think, the more it reinforces how you already think. The thing you actually need from AI — a way out of your own frame — gets harder, not easier, as personalization improves.
Agent autonomy is not the same as good collaboration
Most AI collaboration tools optimize for making the agent more autonomous. The harder problem is designing how human and agent share state and exchange control when both sides can act continuously — and making sure the human's edits are never silently overwritten.