Anthropic Tool Mastery

# Anthropic Tool Mastery

Part of [Agent Skills™](https://github.com/itallstartedwithaidea/agent-skills) by [googleadsagent.ai™](https://googleadsagent.ai)

## Description

Anthropic Tool Mastery is the definitive skill for leveraging Claude's native tool_use capability to its fullest potential. Claude's tool system is not merely a function-calling interface — it is a structured reasoning protocol that enables the model to decompose problems, dispatch parallel operations, handle streaming results, and chain tool outputs through extended thinking. Mastering these patterns is the difference between an agent that awkwardly calls one tool at a time and one that orchestrates complex multi-tool workflows with the fluency of a senior engineer.

This skill codifies patterns proven across the Buddy™ agent at [googleadsagent.ai™](https://googleadsagent.ai), where tool orchestration handles concurrent Google Ads API calls, web searches, file operations, and analysis computations within single reasoning turns. The techniques cover tool definition design (schema quality directly affects call accuracy), parallel dispatch (multiple independent tool calls in a single turn), result composition (combining outputs from parallel calls), and error recovery (handling partial failures in multi-tool batches).

Advanced patterns include streaming tool results for real-time feedback, extended thinking integration (using thinking blocks to plan tool sequences before execution), and tool result caching to avoid redundant calls. These techniques apply directly to Claude Code's built-in tools and extend to custom MCP server tools.

## Use When

- Building agents that need to call multiple tools per reasoning turn
- Tool call accuracy is below acceptable thresholds (wrong parameters, wrong tool selected)
- You need real-time streaming feedback from long-running tool operations
- Extended thinking should inform tool selection and parameter construction
- Multi-tool workflows require coordination and result composition
- Custom tools need to be designed for maximum model compatibility

## How It Works

```mermaid
sequenceDiagram
    participant U as User/System
    participant C as Claude (Extended Thinking)
    participant T1 as Tool A
    participant T2 as Tool B
    participant T3 as Tool C

    U->>C: Task + Tool Definitions
    Note over C: Extended Thinking:<br/>Plan tool sequence
    C->>T1: Parallel Call 1
    C->>T2: Parallel Call 2
    T1-->>C: Result A (streamed)
    T2-->>C: Result B (streamed)
    Note over C: Extended Thinking:<br/>Compose results, plan next step
    C->>T3: Sequential Call (depends on A+B)
    T3-->>C: Result C
    C->>U: Final composed response
```

Claude's tool execution model supports parallel dispatch of independent tool calls within a single assistant turn. The model uses extended thinking to plan tool sequences, identifying which calls are independent (and can be parallelized) and which are dependent (and must be sequenced). Streaming enables real-time progress feedback during long-running tools. After all tool results arrive, extended thinking composes the results into a coherent response or plans the next round of tool calls.

## Implementation

**Optimal Tool Definition Pattern:**

```json
{
  "name": "search_knowledge_base",
  "description": "Search the domain knowledge base for patterns matching a query. Returns ranked results with confidence scores. Use this BEFORE generating recommendations to ground them in verified patterns.",
  "input_schema": {
    "type": "object",
    "properties": {
      "query": {
        "type": "string",
        "description": "Natural language search query describing the pattern or information needed"
      },
      "category": {
        "type": "string",
        "enum": ["bidding", "targeting", "creative", "