Skill patterns you can drop into Claude Code today.
Battle-tested. Documented.

What it does

Hooks are shell commands that fire on Claude Code events: PreToolUse, PostToolUse, Notification, Stop, SubagentStop. The hook runs OUTSIDE the AI's context — meaning rules are enforced without burning tokens or relying on the model "remembering."

When to use

• Auto-format / lint on every write — never forget to run Prettier again
• Audit log every Bash call — append to a file Claude can't tamper with
• Brand enforcement on Write — scan for forbidden patterns (emojis, wrong casing) before commit
• Block dangerous commands — refuse git push --force via PreToolUse
• Notification routing — pipe to terminal-notifier, Slack, or your DND

Install

Hooks live in .claude/settings.json at user level (~/.claude/) or project level. Project-level wins on conflicts.

// .claude/settings.json — auto-lint HTML after every Write/Edit
{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Write|Edit",
      "hooks": [{
        "type": "command",
        "command": "~/scripts/lint-html.sh \"$CLAUDE_FILE_PATH\""
      }]
    }]
  }
}

Available variables in hook commands

• $CLAUDE_FILE_PATH — the file being touched (Write/Edit)
• $CLAUDE_TOOL_NAME — name of the tool that triggered
• $CLAUDE_PROJECT_DIR — the project root

Real example: brand enforcement

// reject the Write if file contains emojis or "AIOPSFORGE" lowercase
{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Write|Edit",
      "hooks": [{
        "type": "command",
        "command": "~/scripts/brand-check.sh",
        "exitOnError": true
      }]
    }]
  }
}

What it does

MCP servers are standalone programs that expose tools, resources, and prompts to Claude over a structured protocol. Anthropic ships official servers (filesystem, brave-search, fetch); the community builds anything else. Claude Code connects via claude mcp add and the tools appear in its tool list.

When to use

• Bring private data into Claude — your Notion, Obsidian vault, internal docs
• Connect to APIs — Gmail, Linear, Stripe, Vercel, Whop, custom internal
• Add specialized capabilities — browser automation (Playwright), local file ops, DB queries
• Avoid prompt-bloating Claude — MCP serves data on demand, not pre-loaded

Install an existing server

# Add via Claude CLI
$ claude mcp add brave-search -- npx @modelcontextprotocol/server-brave-search

# List configured
$ claude mcp list

# Remove
$ claude mcp remove brave-search

Build a custom MCP server (Node.js · stdio)

// my-mcp/server.js
import { McpServer } from '@modelcontextprotocol/sdk/server/mcp';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio';

const server = new McpServer({ name: 'my-tools', version: '1.0.0' });

server.tool('get_today_stats', 'Return today\'s ops stats', {}, async () => {
  const data = await fetch('https://api.internal/stats/today').then(r => r.json());
  return { content: [{ type: 'text', text: JSON.stringify(data) }] };
});

const transport = new StdioServerTransport();
await server.connect(transport);

Then register: claude mcp add my-tools -- node /abs/path/to/my-mcp/server.js