Every AI coding assistant I've used can write code that looks great in a demo and then quietly does the wrong thing in production. It picks a bad access pattern, spins up a new client on every request, or writes a query that costs a fortune at scale. After cleaning up enough of this by hand, I got tired of it and went looking for a better way.

This talk is about what I found. I'll walk through three things you can layer onto almost any agent to make it actually understand your stack: Skills that carry your hard-won rules and only load when they're relevant, MCP for giving the agent live access to your schema and data, and plugin manifests so the same knowledge works in Copilot, Claude Code, Cursor, Codex, and Gemini.

I'll show it live against a real backend, with honest before-and-after code, and share the parts that were harder than expected: keeping context small, writing rules that don't backfire, and actually measuring whether any of it helped.