Stop Tuning Prompts. Build a Harness.

A practical guide to building a robust 'harness' around Claude Code for large codebases, arguing that the tooling layer matters more than the model itself. Covers seven components: keeping CLAUDE.md lean and layered across subdirectories, using hooks for self-improving context (SessionStart for orientation, Stop for auto-updating CLAUDE.md), path-scoped skills for progressive disclosure, LSP-backed MCP servers for symbol-level navigation instead of grep, and subagents to isolate exploration from editing. Includes concrete code examples for hooks and subagent definitions, plus organizational advice on assigning a harness owner and using plugins for team-wide distribution.