Claude Code — Field Guide
What is Claude Code?
Claude Code is Anthropic's command-line tool. You run it in your terminal, point it at a task, and it reads your files, edits them, runs shell commands, and builds software on your machine — in a loop, until the job is done.
It's different from pasting code into a chat window. Claude Code is agentic: it decides what to read, what to run, when to stop and ask. Used well, it's less "AI autocomplete" and more "a smart engineer who lives in your terminal."
Out of the box it works fine. Configured properly, it's the most productive setup I've ever run. This guide covers the configuration part.
What my setup actually looks like
I run Claude Code Max. On any given day it's managing 6–8 active projects simultaneously, each with their own CLAUDE.md, rules, and context. The stack:
- ~/.claude/CLAUDE.md — global rules applied across every project
- ~/.claude/rules/ — modular rule files loaded on demand via
@import - ~/.claude/agents/ — specialist agents for frontend, backend, security, database, git
- ~/.claude/skills/ — 200+ skill files for repeatable workflows (TDD, debugging, design review)
- MCP servers — context7 (docs), lean-ctx (token compression), memory system, Notion, Playwright
- Hooks — PreToolUse guards, PostToolUse triggers, session-start context loading
- Obsidian vault — long-term memory that survives sessions; agents read it directly
This guide documents all of it, in the order that makes sense to build it up.
Who this is for
- You've installed Claude Code and want to set it up right the first time.
- You've been using it for a while and suspect you're leaving most of it on the table.
- You've heard terms like
CLAUDE.md, MCP, hooks, skills, agents — and want to know what they actually do.
No prior knowledge assumed. Each section explains the concept first, then shows exactly how I run it.
How to read this
Sixteen sections, grouped by topic. Read in order the first time. After that, the sections you'll return to most are 01, 06, 08, and 09.
Fundamentals
Start here. These three anchor everything else.
- 01 — Philosophy & CLAUDE.md — how Claude Code actually thinks, and how to shape it with instructions it actually follows.
- 13 — Bootstrap & Install — zero to a working setup on a fresh machine.
- 14 — Secrets & Env — where API keys live, what gets committed, what gets rotated.
Configuration
- 02 — Settings & Permissions —
settings.json, permission rules, deny lists. How to stop getting asked for the same permission 40 times a day. - 03 — The rules/ Directory — modular rule files and
@import. Split your CLAUDE.md before it becomes unmanageable.
Extensions
- 04 — MCP Servers — what MCPs are, which ones are worth installing, and what to check before adding one.
- 05 — Plugins — the plugin system and how to evaluate one.
- 17 — Plugin Deep Dive — an evaluation rubric so you stop installing things that don't earn their keep.
Automation
This is where Claude Code earns its keep. If you skip everything else, read these.
- 06 — Custom Agents — specialist agents, model routing, when to dispatch vs. handle inline.
- 07 — Slash Commands — wrappers for repeatable workflows. The commands that save the most time.
- 08 — Skills — skills, gating, the set that pays for itself.
- 09 — Hooks & Scripts — PreToolUse, PostToolUse, Stop. Automated behaviors that run without you asking.
Memory & Context
- 10 — Memory System — context that survives sessions. The difference between starting fresh every time and picking up where you left off.
- 11 — Obsidian Vault — a vault layout that agents can read. Project notes, decisions, handoffs — all accessible to Claude.
- 12 — lean-ctx Layer — a context engineering layer that cuts token usage by 80–90% on reads. The single highest-leverage install in this list.
- 16 — Workflow Playbooks — composing agents, commands, and skills into repeatable chains.
A note on honesty
Some of this is overkill for most people. Some of it is essential. Where I've kept something running because it earns its keep, I say so. Where I've stopped using something, I say that too.
This is what running 8+ projects in parallel with Claude Code looks like from the inside.