model-selection（选型）

Model Selection Skill

Choose the right model for custom agent tasks based on complexity, cost, and performance requirements.

Interactive Model Selection

Use AskUserQuestion to understand requirements and recommend the optimal model:

# Question 1: Primary Priority (MCP: CLI best practices - tradeoff selection)
question: "What is your primary priority for this agent?"
header: "Priority"
options:
  - label: "Cost Efficiency (Recommended)"
    description: "Minimize API costs, high-volume operations"
  - label: "Balanced Performance"
    description: "Good quality at reasonable cost for most tasks"
  - label: "Maximum Quality"
    description: "Best results regardless of cost, complex reasoning"
  - label: "Lowest Latency"
    description: "Real-time responses, user-facing interactions"

# Question 2: Task Complexity (MCP: Agent SDK model selection)
question: "How complex is the task this agent will perform?"
header: "Complexity"
options:
  - label: "Simple"
    description: "Transformations, extraction, formatting, classification"
  - label: "Moderate"
    description: "Code generation, analysis, planning, most tasks"
  - label: "Complex"
    description: "Architecture decisions, multi-step reasoning, critical code"
  - label: "Variable"
    description: "Mix of simple and complex tasks in one agent"

Use these responses to apply the decision tree and recommend the appropriate model.

Purpose

Guide selection of appropriate Claude model (Haiku, Sonnet, Opus) for custom agent tasks to optimize cost, speed, and quality.

When to Use

• Designing a new custom agent
• Optimizing existing agent performance
• Balancing cost vs quality
• Meeting specific latency requirements

Model Overview

Selection Decision Tree

START
  │
  ├── Is task simple transformation?
  │   └── YES → Haiku
  │
  ├── Is cost the primary concern?
  │   └── YES → Haiku (if adequate) or Sonnet
  │
  ├── Is quality critical (no room for error)?
  │   └── YES → Opus
  │
  ├── Does task require complex reasoning?
  │   └── YES → Opus
  │
  ├── Is latency critical (real-time)?
  │   └── YES → Haiku
  │
  └── DEFAULT → Sonnet (best balance)

Model Selection by Task Type

Haiku Tasks

Best for:

• Text transformations (uppercase, formatting)
• Simple classification
• Data extraction
• High-volume operations
• Real-time processing
• Pattern matching

# Haiku examples
model="claude-3-5-haiku-20241022"

# Echo agent - simple transformation
# Calculator - straightforward math
# Stream processor - high volume, low complexity

Sonnet Tasks

Best for:

• Code generation
• Code review
• Planning and analysis
• Most custom agents
• Balanced performance

# Sonnet examples
model="claude-sonnet-4-20250514"

# QA agent - codebase analysis
# Builder agent - code implementation
# General-purpose agents

Opus Tasks

Best for:

• Strategic planning
• Complex architectural decisions
• Critical code review
• Multi-step reasoning
• Novel problem solving

# Opus examples
model="claude-opus-4-20250514"

# Planner agent - strategic decisions
# Reviewer agent - critical validation
# Architect agent - system design

Cost Considerations

Relative Costs

Cost Optimization Strategies

1. Start with Haiku: Test if simpler model is adequate
2. Use Haiku for preprocessing: Filter/classify before main task
3. Reserve Opus for critical paths: Only where quality is paramount
4. Monitor costs: Track ResultMessage.total_cost_usd

# Cost tracking
async for message in client.receive_response():
    if isinstance(message, ResultMessage):
        print(f"Query cost: ${message.total_cost_usd:.6f}")

Speed Considerations

Latency Profiles

Speed Optimization

1. Real-time needs Haiku: Sub-second response
2. Interactive needs Sonnet: Acceptable latency
3. Batch allows Opus: Latency less critical

Quality Considerations

Capability Differences

Quality Requirements

• Haiku: Acceptable for well-defined, simple tasks
• Sonnet: Good for most development tasks
• Opus: Required for critical decisions

Multi-Model Patterns

Tiered Processing

# Tier 1: Haiku for classification
classification = await classify_task(task, model="haiku")

# Tier 2: Route to appropriate model
if classification == "simple":
    result = await process(task, model="haiku")
elif classification == "complex":
    result = await process(task, model="opus")
else:
    result = await process(task, model="sonnet")

Multi-Agent with Different Models

# Planner: Opus for strategic decisions
planner_options = ClaudeAgentOptions(
    model="claude-opus-4-20250514"
)

# Builder: Sonnet for implementation
builder_options = ClaudeAgentOptions(
    model="claude-sonnet-4-20250514"
)

# Reviewer: Opus for critical review
reviewer_options = ClaudeAgentOptions(
    model="claude-opus-4-20250514"
)

Output Format

When recommending model selection:

## Model Selection

**Task:** [description]
**Recommended Model:** [Haiku/Sonnet/Opus]

### Decision Factors

| Factor | Weight | Assessment |
| --- | --- | --- |
| Complexity | [H/M/L] | [assessment] |
| Cost sensitivity | [H/M/L] | [assessment] |
| Quality requirement | [H/M/L] | [assessment] |
| Latency requirement | [H/M/L] | [assessment] |

### Rationale

[Why this model is appropriate]

### Alternatives

- If cost is concern: [alternative]
- If quality is critical: [alternative]

### Configuration

options = ClaudeAgentOptions(model="[model-id]",...)

Selection Checklist

• Task complexity assessed
• Cost constraints identified
• Quality requirements defined
• Latency requirements considered
• Model selected with rationale
• Alternatives documented

Key Insights

"Choose wisely: Claude Haiku for simple, fast tasks. Claude Sonnet for balanced performance. Claude Opus for complex reasoning."

Model selection directly impacts:

• User experience (latency)
• Operational cost (tokens)
• Output quality (accuracy)

Cross-References

• @core-four-custom.md - Model in Core Four
• @custom-agent-design skill - Agent design workflow
• @agent-deployment-forms.md - Deployment considerations

Version History

• v1.0.0 (2025-12-26): Initial release

Last Updated

Date: 2025-12-26 Model: claude-opus-4-5-20251101