Why the setup around Claude Code matters as much as the model

Anthropic's Applied AI team published a guide on running Claude Code in large codebases, and its main claim is counterintuitive: success depends less on raw model capability than on the scaffolding you build around it. Claude Code reads a repository by exploring it like an agent, not through embedding-based search, so how you structure context changes the output a lot.

Some of the advice is easy to miss. Keep the root CLAUDE.md to pointers and critical warnings only. Start Claude inside a subdirectory rather than at the top of the repo. The team also says to revisit your configuration every three to six months, because rules tuned for an older model can quietly hold back a newer one. That last point catches teams still running a long CLAUDE.md they wrote a year ago.

They lay out a layered setup, roughly in order: CLAUDE.md, hooks, skills, plugins, LSP integration, MCP servers, then subagents to split exploration from editing. For company-wide use they suggest one owner or team for governance, with a three-phase rollout that moves from infrastructure to a small pilot and only then to a wide release. The full guide is on Anthropic's blog.

Why it matters

If your team already uses Claude Code on a big repo, the actionable part is the maintenance habit. A CLAUDE.md written for last year's model may be slowing you down right now, and the fix is a review you can put on the calendar, not a rewrite you have to invent.