Problem
Architecture knowledge lives in code, wikis, and chat logs. For answers you need to stand behind, similarity search over chunks is hard to anchor to artifacts.
Approach
Documentation as a graph in the repo: Markdown with explicit links, orchestrated via `blueprint.md`, navigation from `entry-point.md`. Agents follow edges. Humans review outputs (Verify in a fresh chat).
Artifact
docs/architecture/ (Markdown link graph)
OKF-like structure (frontmatter, index.md, log.md). Tracks: Build, Evolve, Architect, Domain, Verify. Optional MCP/CLI.
Working thesis: architecture documentation is the API of the AI conversation — when it’s versioned, linked, and reviewable.
Golden path in short: Install → Adopt → Continue → Maintain → Verify. Details in the Assistant and workflow prompts.