Special · Claude Code setup

Claude Code project structure: CLAUDE.md, Skills, Rules and Tools

A visual guide to where Claude Code context, rules, skills, tools, subagents, hooks and portable agent instructions belong.

10 slides · October 4, 2026

1. 

   Anthropic / Claude

   01

   Claude Code has layers

   A strong setup separates always-on project context, reusable workflows, external capabilities, automation and isolated workers.

   How Claude Code works in large codebases ↗
2. 

   Anthropic / Claude

   02

   CLAUDE.md is the briefing

   Keep build commands, architecture, conventions and recurring project facts in CLAUDE.md; nested files can load only when Claude enters that area.

   Steering Claude Code: when to use CLAUDE.md, skills, hooks, and subagents ↗
3. 

   Anthropic / Claude

   03

   Rules vs Skills

   Use .claude/rules/ for constraints that can be path-scoped; use Skills for task-specific knowledge and procedures that should load on demand.

   Extend Claude Code ↗
4. 

   Anthropic

   04

   A Skill starts with SKILL.md

   A project Skill lives under .claude/skills/<name>/ and uses SKILL.md metadata to describe when Claude should load the workflow.

   Equipping agents for the real world with Agent Skills ↗
5. 

   Anthropic

   05

   Skills are folders, not prompts

   A Skill can keep SKILL.md concise and pull in reference files, templates or scripts only when that extra detail is needed.

   Equipping agents for the real world with Agent Skills ↗
6. 

   Anthropic

   06

   Tools give capability

   MCP and built-in tools let Claude act on systems and data; Skills teach Claude the repeatable workflow for using those capabilities well.

   Equipping agents for the real world with Agent Skills ↗
7. 

   Anthropic

   07

   Subagents isolate side work

   Define specialists in .claude/agents/ when a task needs its own prompt, tool access or context window and only the result should return.

   Enabling Claude Code to work more autonomously ↗
8. 

   Anthropic / Claude

   08

   Hooks make behavior deterministic

   Put lifecycle automation in hooks and .claude/settings.json when an action must reliably run, block or validate instead of merely being requested.

   Claude Code power user customization: How to configure hooks ↗
9. 

   Anthropic / Claude

   09

   CLAUDE.md or AGENTS.md?

   Claude Code supports AGENTS.md, but when an ancestor CLAUDE.md exists it uses CLAUDE.md by default; import AGENTS.md or configure both if needed.

   How Claude remembers your project ↗
10. 

    Anthropic

    10

    Use the smallest right layer

    A sane repo starts with CLAUDE.md, then adds .claude/rules/, skills/, agents/, settings.json and .mcp.json only when each solves a repeated need.

    Equipping agents for the real world with Agent Skills ↗

Back to For you