A field manual for terminal AI that actually explains the knobs

What it does

This is a living guide to Anthropic’s Claude Code CLI, organized as a choose-your-own-path manual rather than a linear tutorial. It covers setup, prompt engineering, reusable slash commands (Skills), event-driven automation (Hooks), isolated sub-agents, and MCP server integration. The repo includes its own .claude/ directory as a working reference implementation with 8 custom skills, 5 subagent definitions, and Python hook scripts.

The interesting bit

The README treats Claude Code’s “effort levels” as a behavioral dial, not a token budget—a distinction most users miss. It also documents a genuine footgun: setting "effortLevel": "max" in settings.json silently downgrades, and only the CLAUDE_CODE_EFFORT_LEVEL=max environment variable actually persists the top tier. That’s the kind of operational detail you usually find in issue threads, not documentation.

Key highlights

• Four extension points mapped side-by-side: Skills, Hooks, Subagents, MCP servers—with clear “use when / skip if” guidance
• Working examples in-repo: .claude/commands/, .claude/agents/, .claude/hooks/
• Explicit workflow patterns: Explore→Plan→Code→Commit, TDD loop, visual iteration with screenshots
• Model selection guide through the Claude 4 family (Opus 4.7, Sonnet 4.6, Haiku 4.5) with context-window specs
• “Plan with Opus / execute with Sonnet” cost-optimization pattern

Caveats

• Some sections reference future-dated model versions (“May 2026”) and features labeled “experimental—2026”; timeline clarity is uncertain
• The “BMAD method” and “Super Claude framework” are named but not substantively described in the provided excerpts
• Heavy on mental models and tables; light on raw benchmark data or quantitative comparisons

Verdict

Worth bookmarking if you’re past the “hello world” stage with Claude Code and want to systematize your usage. Skip if you’re looking for API internals, alternative tool comparisons, or a quick reference card—this is a strategy guide, not a cheat sheet.