How do you stop AI from making bad architectural decisions in your codebase?

The problem I keep hitting: AI assistants are making architectural decisions — what to import, which layer to touch, which module to put things in — without any understanding of the architecture they're working in. It's not just wrong paths. It's more fundamental. The AI doesn't know your module ownership. It can't trace the blast radius of a change. It has no idea which constraints your team has defined or why. It gets files and hallucinates the rest. I built Mikk to fix this — a codebase context engine that parses your TypeScript project and gives your AI the architecture it's missing. Dependency graph, module boundaries, blast radius analysis, architectural contracts, and a live MCP server so the AI can query your real architecture mid-task. Curious how others are handling this. CLAUDE.md written by hand? RAG pipelines over your source? Context window tricks? What's actually working for non-trivial codebases? (Mikk is open source: https://github.com/Ansh-dhanani/mikk)