Stop Tuning Prompts. Build a Harness.

Anthropic's piece "How Claude Code Works in Large Codebases" emphasizes that the surrounding ecosystem, or "harness," is more crucial to the performance of coding agents like Claude Code than the model itself. The harness, comprised of seven components, provides context and tools that tailor the agent’s operation within a specific code repository, eliminating the need for a codebase index. This setup allows the agent to navigate using basic shell tools, though it initially lacks a comprehensive understanding without contextual guidance. Key elements include a foundational CLAUDE.md file, hooks for self-improvement, path-scoped skills, and symbol-level search capabilities via language and MCP servers, which enhance navigation and reduce brute-force searching. Additionally, subagents are employed to separate exploration from editing, preserving context and efficiency, while periodic reviews and updates to the harness ensure ongoing relevance and performance improvement. The success of implementing Claude Code at scale requires dedicated ownership and strategic integration to prevent fragmentation and ensure consistent use across teams.