Claude Code Features

Claude Code Features Guide

Purpose

This skill activates when discussing Claude Code's extensibility and configuration features. It ensures accurate, actionable guidance rather than guessing.

The Five Pillars of Claude Code Extensibility

1. Skills (.claude/skills/)

What they are: Markdown documents that expand Claude's capabilities with domain knowledge, workflows, or specialized instructions.

Structure:

.claude/skills/{skill-name}/
├── SKILL.md          # Required - contains frontmatter + instructions
└── reference.md      # Optional - supporting documentation

SKILL.md Format:

---
name: my-skill-name
description: Brief description. Include trigger keywords for automatic activation.
allowed-tools: Read, Grep # Optional: restrict to specific tools
---

# Skill Title

## Instructions

Step-by-step guidance for Claude when this skill is active.

## Examples

Concrete usage examples.

How they trigger: Claude autonomously decides to use skills based on matching the user's request to the skill's description. Good descriptions = reliable activation.

Location options:

• ~/.claude/skills/ — Personal skills (all projects)
• .claude/skills/ — Project skills (team-shared, committed to git)

2. Slash Commands (.claude/commands/)

What they are: Markdown files that define reusable prompts invoked with /command-name.

Structure:

.claude/commands/
├── my-command.md           # Invoked as /my-command
└── folder/sub-command.md   # Invoked as /folder:sub-command

Command file format:

---
description: What this command does (shown in command list)
---

Your prompt content here. This expands when the user types /command-name.

You can use $ARGUMENTS to capture user input after the command.

Special variables:

• $ARGUMENTS — Everything typed after the command name
• {{file_path}} — Prompt user for a file path
• {{user_input}} — Prompt user for text input

Location options:

• ~/.claude/commands/ — Personal commands
• .claude/commands/ — Project commands

3. Hooks (.claude/settings.local.json)

What they are: Shell scripts that run automatically before/after Claude uses tools.

Hook types:

• PreToolUse — Runs before a tool executes (can block with exit code 2)
• PostToolUse — Runs after a tool executes
• Notification — Runs when Claude sends notifications
• Stop — Runs when Claude stops (turn completion, interrupt, error)

Configuration in settings.local.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": ".claude/hooks/my-hook.sh",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Environment variables available to hooks:

• CLAUDE_TOOL_NAME — Name of the tool being used
• CLAUDE_TOOL_INPUT — JSON of tool parameters
• CLAUDE_TOOL_OUTPUT — (PostToolUse only) Tool result

Blocking behavior: Exit code 2 blocks the tool; stdout becomes the block reason shown to Claude.

4. MCP Servers

What they are: External servers that provide additional tools to Claude via the Model Context Protocol.

Configuration in settings.json:

{
  "mcpServers": {
    "server-name": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-name"],
      "env": {
        "API_KEY": "your-key"
      }
    }
  }
}

Common MCP servers:

• @modelcontextprotocol/server-filesystem — Enhanced file operations
• @modelcontextprotocol/server-github — GitHub API integration
• @modelcontextprotocol/server-postgres — Database queries
• @anthropics/claude-code-mcp-server — Claude Code specific tools

Location options:

• ~/.claude/settings.json — Global MCP servers
• .claude/settings.json — Project-specific servers

5. CLAUDE.md Files

What they are: Markdown files that provide context and instructions to Claude.

Hierarchy (all are read, in order):

1. ~/.claude/CLAUDE.md — Personal global instructions
2. ~/CLAUDE.md — User home directory
3. {project}/CLAUDE.md — Project root (most common)
4. {project}/{subdir}/CLAUDE.md — Subdirectory-specific context

Best practices:

• Keep focused on immutable truths (architecture, conventions, patterns)
• Avoid transient information (status, versions, URLs)
• Use for coding standards, project structure, key commands

Settings Files

.claude/settings.json — Project settings (committed to git):

• Permissions (allow/deny tool patterns)
• MCP server configuration

.claude/settings.local.json — Local overrides (gitignored):

• Personal permissions
• Hooks configuration
• Local MCP servers

Permission format:

{
  "permissions": {
    "allow": ["Bash(npm test:*)", "Read"],
    "deny": ["Bash(rm -rf:*)"]
  }
}

Quick Reference: When to Use What

Debugging Tips

1. Skill not triggering? Check description has specific trigger keywords
2. Command not found? Verify path: .claude/commands/name.md → /name
3. Hook not running? Check matcher regex and file permissions (chmod +x)
4. MCP server failing? Run the command manually to see errors