Injection order: Global CLAUDE.md → Global rules → Project CLAUDE.md → Memory

This ~12KB payload is re-sent with every request. A 500-line CLAUDE.md quietly burns tokens on every API call. Keep it lean.

2. MCP tools are lazy-loaded — watch tools[] grow

Built-in tools (27) ship their full JSON schemas every request. MCP tools don't — they start as names only.

Watch the count change in real-time:

| Step | What happens | tools[] count | |------|-------------|-----------------| | Initial request | 27 built-in tools loaded | 27 | | Model calls ToolSearch("context7") | Full schema for 2 MCP tools returned | 29 | | Model calls ToolSearch("til") | 6 more MCP tool schemas added | 35 |

Unused MCP tools never consume tokens. The Inspector lets you watch tools[] grow as the model discovers what it needs.

3. Images are base64-encoded inline

When Claude Code reads a screenshot or image file, the image is base64-encoded and embedded directly in the JSON body:

{
  "type": "image",
  "source": {
    "type": "base64",
    "media_type": "image/png",
    "data": "iVBORw0KGgo..."
  }
}

A single screenshot can add hundreds of KB to the request payload. The Inspector shows you the exact size.

4. Skill ≠ Command — completely different injection paths

Typing /something triggers one of three completely different mechanisms:

| | Local Command | User Sk...