anthropic-architect（Anthropic 架构师）

Anthropic Architect

Expert architectural guidance for Anthropic-based projects. Analyze your requirements and receive tailored recommendations on the optimal architecture using Skills, Agents, Subagents, Prompts, and SDK primitives.

What This Skill Does

Helps you design the right Anthropic architecture for your project by:

• Analyzing project requirements - Understanding complexity, scope, and constraints
• Recommending architectures - Skills vs Agents vs Prompts vs SDK primitives
• Applying decision rubrics - Data-driven architectural choices
• Following best practices - 2025 Anthropic patterns and principles
• Progressive disclosure design - Efficient context management
• Security considerations - Safe, controllable AI systems

Why Architecture Matters

Without proper architecture:

• Inefficient context usage and high costs
• Poor performance and slow responses
• Security vulnerabilities and risks
• Difficult to maintain and scale
• Agents reading entire skill contexts unnecessarily
• Mixed concerns and unclear boundaries

With engineered architecture:

• Optimal context utilization
• Fast, focused responses
• Secure, controlled operations
• Easy to maintain and extend
• Progressive disclosure of information
• Clear separation of concerns
• Scalable and reusable components

Quick Start

Analyze Your Project

Using the anthropic-architect skill, help me determine the best
architecture for: [describe your project]

Requirements:
- [List your key requirements]
- [Complexity level]
- [Reusability needs]
- [Security constraints]

Get Architecture Recommendation

The skill will provide:

1. Recommended architecture - Specific primitives to use
2. Decision reasoning - Why this architecture fits
3. Implementation guidance - How to build it
4. Best practices - What to follow
5. Example patterns - Similar successful architectures

The Four Anthropic Primitives

1. Skills (Prompt-Based Meta-Tools)

What: Organized folders of instructions, scripts, and resources that agents can discover and load dynamically.

When to use:

• ✅ Specialized domain knowledge needed
• ✅ Reusable across multiple projects
• ✅ Complex, multi-step workflows
• ✅ Reference materials required
• ✅ Progressive disclosure beneficial

When NOT to use:

• ❌ Simple, one-off tasks
• ❌ Project-specific logic only
• ❌ No need for reusability

Example use cases:

• Prompt engineering expertise
• Design system generation
• Code review guidelines
• Domain-specific knowledge (finance, medical, legal)

2. Agents/Subagents (Autonomous Task Handlers)

What: Specialized agents with independent system prompts, dedicated context windows, and specific tool permissions.

When to use:

• ✅ Complex, multi-step autonomous tasks
• ✅ Need for isolated context
• ✅ Different tool permissions required
• ✅ Parallel task execution
• ✅ Specialized expertise per task type

When NOT to use:

• ❌ Simple queries or lookups
• ❌ Shared context required
• ❌ Sequential dependencies
• ❌ Resource-constrained environments

Example use cases:

• Code exploration and analysis
• Test generation and execution
• Documentation generation
• Security audits
• Performance optimization

3. Direct Prompts (Simple Instructions)

What: Clear, explicit instructions passed directly to Claude without additional structure.

When to use:

• ✅ Simple, straightforward tasks
• ✅ One-time operations
• ✅ Quick questions or clarifications
• ✅ No need for specialization
• ✅ Minimal context required

When NOT to use:

• ❌ Complex, multi-step processes
• ❌ Need for reusability
• ❌ Requires domain expertise
• ❌ Multiple related operations

Example use cases:

• Code explanations
• Quick refactoring
• Simple bug fixes
• Documentation updates
• Direct questions

4. SDK Primitives (Custom Workflows)

What: Low-level building blocks from the Claude Agent SDK to create custom agent workflows.

When to use:

• ✅ Unique workflow requirements
• ✅ Custom tool integration needed
• ✅ Specific feedback loops required
• ✅ Integration with existing systems
• ✅ Fine-grained control needed

When NOT to use:

• ❌ Standard use cases covered by Skills/Agents
• ❌ Limited development resources
• ❌ Maintenance burden concern
• ❌ Faster time-to-market priority

Example use cases:

• Custom CI/CD integration
• Specialized code analysis pipelines
• Domain-specific automation
• Integration with proprietary systems

Decision Rubric

Use this rubric to determine the right architecture:

Task Complexity Analysis

Low Complexity → Direct Prompts

• Single operation
• Clear input/output
• No dependencies
• < 5 steps

Medium Complexity → Skills

• Multiple related operations
• Reusable patterns
• Reference materials helpful
• 5-20 steps

High Complexity → Agents/Subagents

• Multi-step autonomous workflow
• Needs isolated context
• Different tool permissions
• 20 steps or parallel tasks

Custom Complexity → SDK Primitives

• Unique workflows
• System integration required
• Custom tools needed
• Specific feedback loops

Reusability Assessment

Single Use → Direct Prompts

• One-time task
• Project-specific
• No future reuse

Team Reuse → Skills

• Multiple team members benefit
• Common workflows
• Shareable knowledge

Organization Reuse → Skills + Marketplace

• Cross-team benefit
• Standard patterns
• Company-wide knowledge

Product Feature → SDK Primitives

• End-user facing
• Production deployment
• Custom integration

Context Management Needs

Minimal Context → Direct Prompts

• Self-contained task
• No external references
• Simple instructions

Structured Context → Skills

• Progressive disclosure needed
• Reference materials required
• Organized information

Isolated Context → Agents/Subagents

• Separate concerns
• Avoid context pollution
• Parallel execution

Custom Context → SDK Primitives

• Specific context handling
• Integration requirements
• Fine-grained control

Security & Control Requirements

Basic Safety → Direct Prompts + Skills

• Standard guardrails
• No sensitive operations
• Read-only or low-risk

Controlled Access → Agents with Tool Restrictions

• Specific tool permissions
• Allowlist approach
• Confirmation required

High Security → SDK Primitives + Custom Controls

• Deny-all default
• Explicit confirmations
• Audit logging
• Custom security layers

Architecture Patterns

Pattern 1: Skills-First Architecture

Use when: Building reusable expertise and workflows

Structure:

Project
├── skills/
│   ├── domain-expert/
│   │   ├── SKILL.md
│   │   └── references/
│   │       ├── patterns.md
│   │       ├── best_practices.md
│   │       └── examples.md
│   └── workflow-automation/
│       ├── SKILL.md
│       └── scripts/
│           └── automate.sh
└── .claude/
    └── config

Benefits:

• Reusable across projects
• Progressive disclosure
• Easy to share and maintain
• Clear documentation

Pattern 2: Agent-Based Architecture

Use when: Complex autonomous tasks with isolated concerns

Structure:

Main Agent (orchestrator)
├── Explore Agent (codebase analysis)
├── Plan Agent (task planning)
├── Code Agent (implementation)
└── Review Agent (validation)

Benefits:

• Parallel execution
• Isolated contexts
• Specialized expertise
• Clear responsibilities

Pattern 3: Hybrid Architecture

Use when: Complex projects with varied requirements

Structure:

Main Conversation
├── Direct Prompts (simple tasks)
├── Skills (reusable expertise)
│   ├── code-review-skill
│   └── testing-skill
└── Subagents (complex workflows)
    ├── Explore Agent
    └── Plan Agent

Benefits:

• Right tool for each task
• Optimal resource usage
• Flexible and scalable
• Best of all approaches

Pattern 4: SDK Custom Architecture

Use when: Unique requirements or product features

Structure:

Custom Agent SDK Implementation
├── Custom Tools
├── Specialized Feedback Loops
├── System Integrations
└── Domain-Specific Workflows

Benefits:

• Full control
• Custom integration
• Unique workflows
• Production-ready

Key Principles (2025)

1. Progressive Disclosure

What: Show only what's needed, when it's needed.

Why: Avoids context limits, reduces costs, improves performance.

How: Organize skills with task-based navigation, provide query tools, structure information hierarchically.

2. Context as Resource

What: Treat context window as precious, limited resource.

Why: Every token counts toward limits and costs.

How: Use progressive disclosure, prefer retrieval over dumping, compress aggressively, reset periodically.

3. Clear Instructions

What: Explicit, unambiguous directions.

Why: Claude 4.x responds best to clarity.

How: Be specific, define output format, provide examples, avoid vagueness.

4. Security by Design

What: Deny-all default, allowlist approach.

Why: Safe, controlled AI systems.

How: Limit tool access, require confirmations, audit operations, block dangerous commands.

5. Thinking Capabilities

What: Leverage Claude's extended thinking mode.

Why: Better results for complex reasoning.

How: Request step-by-step thinking, allow reflection after tool use, guide initial thinking.

6. Two-Message Pattern

What: Use meta messages for context without UI clutter.

Why: Clean UX while providing necessary context.

How: Set isMeta: true for system messages, use for skill loading, keep UI focused.

Reference Materials

All architectural patterns, decision frameworks, and examples are in the references/ directory:

• decision_rubric.md - Comprehensive decision framework
• architectural_patterns.md - Detailed pattern catalog
• best_practices.md - 2025 Anthropic best practices
• use_case_examples.md - Real-world architecture examples

Usage Examples

Example 1: Determining Architecture for Content Generation

Input:

Using anthropic-architect, I need to build a system that:
- Generates blog posts from product features
- Ensures brand voice consistency
- Includes SEO optimization
- Reusable across marketing team

Analysis:

• Medium complexity (structured workflow)
• High reusability (team-wide)
• Domain expertise needed (content, SEO, brand)
• Progressive disclosure beneficial

Recommendation: Skills-First Architecture

• Create content-generator skill
• Include brand voice references
• SEO guidelines in references
• Example templates
• Progressive disclosure for different content types

Example 2: Code Refactoring Tool

Input:

Using anthropic-architect, I want to:
- Analyze codebase for refactoring opportunities
- Generate refactoring plan
- Execute refactoring with tests
- Review and validate changes

Analysis:

• High complexity (multi-step, autonomous)
• Different contexts needed (explore, plan, code, review)
• Parallel execution beneficial
• Tool permissions vary by stage

Recommendation: Agent-Based Architecture

• Main orchestrator agent
• Explore subagent (read-only, codebase analysis)
• Plan subagent (planning, no execution)
• Code subagent (write permissions)
• Review subagent (validation, test execution)

Example 3: Simple Code Review

Input:

Using anthropic-architect, I need to:
- Review this PR for bugs
- Check code style
- Suggest improvements

Analysis:

• Low complexity (single operation)
• One-time task
• No reusability needed
• Minimal context

Recommendation: Direct Prompt

• Simple, clear instructions
• No skill/agent overhead
• Fast execution
• Sufficient for task

Example 4: Custom CI/CD Integration

Input:

Using anthropic-architect, I want to:
- Integrate Claude into CI pipeline
- Custom tool for deployment validation
- Specific workflow for our stack
- Production feature

Analysis:

• Custom complexity
• System integration required
• Production deployment
• Unique workflows

Recommendation: SDK Primitives

• Build custom agent with SDK
• Implement custom tools
• Create specialized feedback loops
• Integration with CI system

Best Practices Checklist

When designing your architecture:

• Analyzed task complexity accurately
• Considered reusability requirements
• Evaluated context management needs
• Assessed security requirements
• Applied progressive disclosure where beneficial
• Chose simplest solution that works
• Documented architectural decisions
• Planned for maintenance and updates
• Considered cost implications
• Validated with prototype/POC

Common Anti-Patterns

Anti-Pattern 1: Over-Engineering

Problem: Using Agents/SDK for simple tasks

Solution: Start simple, scale complexity as needed

Anti-Pattern 2: Context Dumping

Problem: Loading entire skills into context

Solution: Use progressive disclosure, query tools

Anti-Pattern 3: Mixed Concerns

Problem: Single skill/agent doing too much

Solution: Separate concerns, use subagents or multiple skills

Anti-Pattern 4: No Security Boundaries

Problem: Full tool access for all agents

Solution: Allowlist approach, minimal permissions

Anti-Pattern 5: Ignoring Reusability

Problem: Recreating same prompts repeatedly

Solution: Extract to skills, share across projects

Getting Started

Step 1: Describe Your Project

Provide clear requirements, complexity level, and constraints.

Step 2: Receive Recommendation

Get tailored architecture with reasoning.

Step 3: Review Patterns

Explore similar successful architectures.

Step 4: Implement

Follow implementation guidance.

Step 5: Iterate

Refine based on results and feedback.

Summary

The Anthropic Architect skill helps you:

• Choose the right primitives for your needs
• Design scalable, maintainable architectures
• Follow 2025 best practices
• Avoid common pitfalls
• Optimize for performance and cost

Key Primitives:

• Skills - Reusable domain expertise
• Agents - Autonomous complex workflows
• Prompts - Simple direct tasks
• SDK - Custom integrations

Core Principles:

• Progressive disclosure
• Context as resource
• Security by design
• Clear instructions
• Right tool for the job

"The best architecture is the simplest one that meets your requirements."