Type: url
Original: https://zenn.dev/sasadango28/articles/claude-code-harness-engineering-20260415
Status: generated
Language: ja
Summary
A data engineer describes applying harness engineering to Claude Code in a personal information-gathering and knowledge-management repository, arranged as five layers: CLAUDE.md as the project’s constitution, .claude/rules/ for domain rules, Skills for reusable workflows, Agents for context-isolated subagents, and settings.json permissions as the safety device. Lessons drawn include stating what the agent should not do, moving deterministic work out of the LLM into scripts, keeping CLAUDE.md small, refining permissions through use, and deferring Hooks until deny rules prove insufficient.