1. The Problem with Generic Assistants Every AI assistant starts the same way: a powerful model with no memory, no personality, and no idea who you are or what you're building. You get capable but characterless. You ask it something, it helps, and tomorrow it's a stranger again. You find yourself re-explaining your stack, your preferences, your context every single session. I wanted something different. Not a smarter search engine but a collaborator. One that knows I run a homelab on HAOS, that I think in infrastructure, that I care about elegance as much as correctness, and that I don't need things explained twice. The answer turned out to be surprisingly low-tech: a handful of markdown files injected into the model's context at the start of every session. For context: at work I'm a heavy user of OpenCode , which has its own take on this through plugins like oh-my-openagent . The homelab setup I'm describing here is inspired by that, but it's not the same thing.…