Anthropic has published a new guide on Claude Code at scale, outlining how the tool is being used in large monorepos, legacy systems, and distributed codebases. The company’s message is that success appears to depend as much on setup and organization as on the model itself.
The post argues that Claude Code works more like a developer moving through live files than a system relying on a centralized index. That approach, according to Anthropic, avoids some of the stale retrieval problems seen in RAG-based tools, but it also means teams need to give the assistant enough starting context to find its way around. The company suggests that codebase structure and local conventions play a major role in how well the system performs.
Most of the article is devoted to what Anthropic calls the “harness” around the model. That includes familiar building blocks such as CLAUDE.md files, hooks, skills, plugins, MCP servers, LSP integrations, and subagents. Rather than treating model quality as the only variable, the guide presents these pieces as the machinery that determines whether Claude Code stays useful in a large organization.
Anthropic also points to several recurring patterns in stronger deployments: layered context files, scoped commands by subdirectory, exclusions for generated files and build artifacts, and codebase maps for less conventional repository layouts. The company even recommends revisiting that setup regularly, since instructions that once helped one model version can become unnecessary or overly restrictive after later updates.
The final section shifts from tooling to ownership, with Anthropic suggesting that adoption tends to go better when one person or a small team owns configuration, governance, and rollout. It also points to broader questions around approved skills, plugin control, and review processes for AI-generated code. Those interested in the specifics can read the full post here: How Claude Code works in large codebases: Best practices and where to start.
Source: Claude
