Anthropic SDK Skill

Provider: Anthropic Languages: Python, TypeScript Last Updated: 2026-01-01

Overview

Progressive disclosure documentation for Anthropic Claude API development. This skill provides comprehensive patterns for Messages API, Tool Use, Extended Thinking, Agent SDK, Batches API, Computer Use, MCP, Citations, Streaming, Vision, and more.

Architecture

anthropic/
├── README.md              # This file - overview and usage
├── SKILL.md               # Quick reference (<500 lines) - Essential patterns
├── REFERENCE.md           # Comprehensive reference - Advanced patterns
├── VALIDATION.md          # Feature coverage validation matrix
├── templates/             # Code generation templates
│   ├── messages.template.py          # Messages API (Python)
│   ├── messages.template.ts          # Messages API (TypeScript)
│   ├── streaming.template.py         # Streaming patterns
│   ├── tool-use.template.py          # Tool use patterns
│   ├── extended-thinking.template.py # Extended thinking
│   └── README.md                     # Template usage guide
└── examples/              # Real-world implementations
    ├── basic-chat.example.py         # Basic message completion
    ├── tool-use.example.py           # Tool use patterns
    ├── streaming.example.py          # Streaming responses
    ├── vision.example.py             # Image analysis
    └── README.md                     # Examples index

Progressive Disclosure Pattern

SKILL.md (Quick Reference)

Size: <500 lines target
Use Case: Fast lookups during active development
Content: Essential patterns, model specs, common operations
Load Time: <100ms

REFERENCE.md (Full Reference)

Size: ~1000 lines
Use Case: Advanced patterns and detailed implementations
Content: Agent SDK, Batches API, Computer Use, MCP, Citations, etc.

When to Use

The backend-developer agent loads this skill when:

package.json contains "@anthropic-ai/sdk" dependency
requirements.txt or pyproject.toml contains anthropic
.env contains ANTHROPIC_API_KEY
User explicitly mentions "Anthropic", "Claude", or specific Claude models

Framework Detection

Primary Signals (Confidence: 0.5 each):

package.json -> dependencies.@anthropic-ai/sdk
requirements.txt contains anthropic
pyproject.toml contains anthropic

Secondary Signals (Confidence: 0.3 each):

.env contains ANTHROPIC_API_KEY
Import statements: from anthropic import or import anthropic

Boost Factors (+0.1 each):

Files matching *claude*.py pattern
Files matching *anthropic*.py pattern

Minimum Confidence: 0.8 (80%) required for automatic detection

Core Capabilities

1. Messages API

Basic message patterns
System prompts
Multi-turn conversations
Temperature and parameters

2. Agent SDK (NEW)

Build custom agents powered by Claude
Tool access (Task, Read, Write, Edit, Bash, WebSearch)
Permission modes (default, plan, acceptEdits)
Session management and conversation continuation
Context management and automatic compaction

3. Message Batches API (NEW)

50% cost savings for async processing
Bulk document processing
Up to 10,000 requests per batch
24-hour processing window

4. Streaming

Server-sent events (SSE)
Async generators
Partial responses
Content block deltas

5. Tool Use

Tool definitions
Tool execution loops
Parallel tool calls
Result handling

6. Extended Thinking

Budget tokens configuration
Thinking block handling
Streaming with thinking
Cost optimization

7. Vision/Multimodal

Image inputs (base64, URL)
PDF document analysis
Multiple images
Detail analysis

8. Computer Use (Beta) (NEW)

Browser/desktop automation
Screenshot capture and analysis
Mouse and keyboard actions
Sandboxed execution patterns

9. MCP Integration (NEW)

Model Context Protocol overview
Tool exposure patterns
Server implementation
Client libraries

10. Citations (NEW)

Grounded responses from documents
Multi-document citations
Quote extraction
Source verification

11. Prompt Caching

Cache control headers
Ephemeral caching
Cost optimization
Cache hit tracking

Platform Differentiators

Feature	Anthropic	Description
Extended Thinking	Native	Budget-controlled deep reasoning
Agent SDK	Full SDK	Build agents with multiple tools
Claude Code	Built-in	Terminal-based coding assistant
Constitutional AI	Core	Safety-focused training
200K Context	Standard	All models support 200K tokens
MCP	Open Protocol	Extensible tool protocol
Prompt Caching	Native	90% cost savings
Computer Use	Beta	Browser/desktop automation
Message Batches	Native	50% cost savings
Citations	Native	Grounded responses

Model Coverage

Claude 4 Family (Latest)

Model	Context	Max Output	Extended Thinking	Computer Use
claude-opus-4-5-20251101	200K	32K	Yes	Yes
claude-sonnet-4-20250514	200K	64K	Yes	Yes

Claude 3.5 Family

Model	Context	Max Output	Extended Thinking	Computer Use
claude-3-5-sonnet-20241022	200K	8K	No	Yes
claude-3-5-haiku-20241022	200K	8K	No	No

Claude 3 Family

Model	Context	Max Output	Extended Thinking	Computer Use
claude-3-opus-20240229	200K	4K	No	Yes
claude-3-sonnet-20240229	200K	4K	No	No
claude-3-haiku-20240307	200K	4K	No	No

Quick Start

Loading the Skill

// Embedded in backend-developer.yaml
const skill = await skillLoader.loadSkill('anthropic', 'quick');
// Returns SKILL.md content for fast reference

Using Templates

# Generate messages handler
cp templates/messages.template.py src/chat.py
# Replace placeholders

# Generate streaming handler
cp templates/streaming.template.py src/stream.py
# Replace placeholders

File Size Guidelines

SKILL.md: Target <=100KB (Quick reference should be fast)
Templates: 50-200 lines each (Focused, single responsibility)
Examples: 100-300 lines each (Real-world, production-ready)

Integration with Backend-Developer

The backend-developer agent uses this skill by:

Detection Phase: Runs framework-detector to identify Anthropic
Loading Phase: Loads SKILL.md for quick patterns
Code Generation: Uses templates for boilerplate reduction
Learning Phase: References examples for best practices

Maintenance

Updating Content

When Anthropic releases new features:

Update SKILL.md with new APIs
Add templates for new patterns if commonly used
Update model coverage tables
Increment skill version (semantic versioning)

Validation

Before releasing updates:

SKILL.md file size <=100KB
All templates pass linting
Examples demonstrate production-ready code
Model information is current

Support

For issues or improvements:

Review existing examples for patterns
Reference official Anthropic documentation: https://docs.anthropic.com/

Performance Metrics

Target Metrics:

Skill load time: <100ms (SKILL.md)
Template generation: <50ms per file
Code generation success rate: >=95%
User satisfaction: >=90%

Part of Skills-Based Framework Architecture Related: @fortium/ensemble-ai

using-anthropic-platform