"The harness matters as much as the model." That's the central claim in Anthropic's new guide on deploying Claudee Code](/tools/claude-code/) across large codebases - and after reading it, the point holds up.
The guide covers real deployments: multi-million-line monorepos, decades-old legacy systems, dozens of microservices, and codebases in languages most people don't associate with AI coding tools - C, C++, Java, PHP. The headline finding is that raw model capability isn't the bottleneck for most teams. Configuration is.
The Seven Pieces That Determine How Well It Works
Claude Code's performance in a large codebase depends on what Anthropic calls "the harness" - seven components that together shape what the model can see and do:
- CLAUDE.md files - Context loaded automatically at session start. The advice is to keep these lean and layered: a root file for the high-level overview, subdirectory files for local conventions. Bloated root files are a common mistake.
- Hooks - Scripts that trigger at specific moments, useful for capturing session learnings or automating repetitive steps.
- Skills - Packaged instruction sets loaded on demand. Loading instructions only when relevant prevents the context window (the amount of text Claude can hold in memory at once) from filling up with irrelevant noise.
- Plugins - Bundled configurations distributed across an organization. One retail company in the guide shipped an internal analytics plugin to teams before any broad rollout.
- Language Server Protocol (LSP) - This is the technical one worth understanding. LSP gives Claude symbol-level navigation rather than raw text search. Instead of grep returning thousands of false matches across a massive codebase, LSP lets Claude follow the actual definition of a function or class. One enterprise deployed LSP organization-wide specifically to fix reliability issues with C/C++ codebases.
- MCP Servers - Connections to internal tools and APIs that Claude can't otherwise reach.
- Subagents - Separate Claude instances used to split exploration from editing, so the model doing the reading doesn't pollute the context of the model doing the writing.
The Maintenance Problem Nobody Plans For
The most useful advice in the guide is also the easiest to skip: review your Claude Code configuration every 3-6 months.
This matters because models change. Instructions written to compensate for a specific model limitation become dead weight - or actively harmful constraints - once that limitation is fixed. Hooks that paper over a bug the model no longer has are now just overhead. Teams that wrote their CLAUDE.md files 18 months ago and never touched them are probably running on outdated assumptions.
On the organizational side, the guide recommends assigning a directly responsible individual for configuration management before any broad rollout, not after. In regulated industries, that also means defining approval processes for new skills and plugins up front - trying to retrofit governance after thousands of engineers are already using custom configurations is a documented failure pattern.
How Claude Code Actually Navigates a Codebase
One detail worth understanding: Claude Code doesn't use a centralized index of your codebase. It reads files, runs grep searches, and follows references the same way a new engineer would - traversing the file system in real time rather than querying a pre-built knowledge base. The advantage is that it never serves stale results from an index that hasn't caught up with recent commits. The disadvantage is that without good LSP integration and a sensible .claudeignore file (which tells Claude what to skip), it can waste time in generated files and build artifacts.
For teams evaluating Claude Code for serious engineering work, the guide's advice comes down to one preparation step that pays off: invest in codebase legibility before giving anyone broad access. Clean up your CLAUDE.md files, exclude the noise with .claudeignore, and get LSP running if your codebase is in a compiled language. The model can only work with what it can see.