Turns out it's not often you can point to a single markdown file and say "this made the agent 27% better at resolving real tasks." That's what we saw on our production monorepo. I imagine this as a way for teams to actually make their agents write better code instead of guessing.