Claude Code

Claude Code

Claude Code has the most comprehensive configuration system among AI coding agents.

File Overview

Category	File	Scope
Instructions	CLAUDE.md	Project-wide context and rules
Instructions	CLAUDE.local.md	Personal overrides (gitignored)
Instructions	.claude/CLAUDE.md	Alternative project instructions
Rules	.claude/rules/*.md	Topic-scoped or path-gated rules
Rules	~/.claude/rules/*.md	User-level global rules
Skills	.claude/skills/<name>/SKILL.md	Reusable workflows
Skills	~/.claude/skills/<name>/SKILL.md	User-level global skills
Commands	.claude/commands/*.md	Legacy custom commands (integrated into skills)
Subagents	.claude/agents/*.md	Project-scoped custom subagents
Subagents	~/.claude/agents/*.md	User-level custom subagents
Settings	.claude/settings.json	Permissions, hooks, MCP servers
Settings	.claude/settings.local.json	Personal settings overrides
Settings	~/.claude/settings.json	User-level global settings

Instruction Hierarchy

Claude Code merges instructions from multiple scopes (global > project > subdirectory):

1. ~/.claude/CLAUDE.md (user-level)
2. CLAUDE.md (project root)
3. Subdirectory CLAUDE.md files (deeper = more specific)

Rules

.claude/rules/*.md files provide topic-specific or path-gated instructions. They supplement CLAUDE.md without duplicating its content.

Rules have no required frontmatter. They are plain Markdown files that Claude reads automatically.

Skills

Skills are the primary mechanism for reusable workflows. Custom commands (.claude/commands/*.md) are integrated into skills — both create slash commands, but skills offer richer configuration.

Skill Frontmatter

---
name: deploy                        # Optional. Creates /deploy slash command
description: Deploy to production   # Recommended. Used for auto-invocation matching
argument-hint: [env]                # Optional. Shown in completion UI
disable-model-invocation: true      # Optional. Prevents auto-invocation (manual /name only)
user-invocable: true                # Optional. Whether it appears in slash command menu
allowed-tools:                      # Optional. Restricts available tools during skill execution
  - Read
  - Edit
  - Bash(make *)
model: opus                         # Optional. Override model for this skill
effort: high                        # Optional. Reasoning effort level
context: fork                       # Optional. Run in forked subagent context
---

Claude has the most feature-rich skill frontmatter of all tools.

Subagents

Custom subagents are defined as Markdown files in .claude/agents/. Each file defines a specialized agent with its own system prompt, available tools, and model configuration.

Settings

.claude/settings.json controls:

• Permissions and approval policies
• Hook definitions (pre/post tool execution)
• MCP server connections
• Environment variables
• Model defaults