Summary

Claude Code is Anthropic’s coding agent CLI. Its power comes from combining features into workflows, not using them individually. The feature set maps directly to harness engineering concepts: Memory/CLAUDE.md = feedforward guides, Hooks/Tests = feedback sensors.

Key Points

  • 10 core features: Slash Commands, Memory, Skills, Subagents, MCP, Hooks, Plugins, Checkpoints, Advanced Features (planning/thinking), CLI
  • The real leverage is composition: e.g., code review = Slash Commands + Subagents + Memory + MCP
  • Harness engineering mapping:
    • CLAUDE.md / Memory = feedforward (guides agent before acting)
    • Hooks / Tests = feedback (sensors after action)
    • Skills = reusable harness components
    • Checkpoints = rollback capability (safe experimentation)
  • Feature persistence spectrum: Session-only (slash commands) → Cross-session (memory) → Real-time (MCP)
  • Learning path: 11-13 hours total, but 15 min for immediate value

Feature Quick Reference

FeatureInvocationPersistenceHarness Role
Memory/CLAUDE.mdAuto-loadedCross-sessionFeedforward guide
SkillsAuto-invokedFilesystemReusable harness
HooksEvent-triggeredConfiguredFeedback sensor
SubagentsAuto-delegatedIsolatedTask decomposition
MCPAuto-queriedReal-timeExternal data access
CheckpointsManual/AutoSessionRollback safety net

Open Questions

  • What’s the optimal CLAUDE.md structure for different project types?
  • How to version-control and share Claude Code configurations across teams?
  • Best practices for hook complexity vs. maintainability?

Evidence Timeline

  • 2026-04-07: Compiled from claude-howto guide (5,900+ stars, luongnv89)

相关页面

luongnv89, claude-code-architecture