agent-model-selection
Guidelines for selecting appropriate AI model (Sonnet vs Haiku) based on task complexity, ensuring cost efficiency while maintaining quality. Use when assigning work.
When & Why to Use This Skill
This Claude skill provides a strategic framework for optimizing AI agent workflows by intelligently selecting between Claude 3.5 Sonnet and Claude 3 Haiku based on task complexity. It helps developers and businesses maximize cost efficiency without compromising on quality by routing high-reasoning tasks to Sonnet and high-volume, repetitive tasks to Haiku, effectively balancing performance with operational expenditure.
Use Cases
- Cost Optimization: Automatically routing high-volume, repetitive tasks like data tagging or simple CRUD operations to Claude Haiku to reduce API costs by approximately 95%.
- Quality Assurance for Complex Tasks: Ensuring that critical architecture design, security analysis, and multi-step logical reasoning are handled by Claude Sonnet to maintain high output standards.
- Dynamic Task Orchestration: Implementing a router agent that evaluates incoming requests against specific complexity criteria (e.g., intent analysis vs. rule following) to delegate work to the most appropriate model.
- Efficiency Benchmarking: Using the provided decision flowchart to standardize how a development team assigns AI resources across different stages of a software development lifecycle.
| name | agent-model-selection |
|---|---|
| description | Guidelines for selecting appropriate AI model (Sonnet vs Haiku) based on task complexity, ensuring cost efficiency while maintaining quality. Use when assigning work. |
Agent Model Selection
Instructions
Core decision
Sonnet: Complex reasoning, architecture, security (2+ criteria) Haiku: Defined rules, repetitive tasks, simple commands (~95% cheaper)
Selection criteria
Use Sonnet if 2+ apply:
- Logical reasoning and trade-off analysis
- Architecture/design decisions
- Semantic/intent analysis
- Problem diagnosis and strategy
- Multi-component interaction
- Security/performance analysis
Use Haiku if dominant:
- Following defined rules/templates
- Repetitive mechanical tasks
- Command execution and collection
- Simple CRUD operations
- Format validation
Decision flowchart
Architecture/design? → YES → Sonnet
Multiple options? → YES → Sonnet
Security/performance? → YES → Sonnet
Defined rules only? → YES → Haiku
Detailed guide? → YES → Haiku
Large delegated? → YES → Sonnet
Simple commands? → YES → Haiku
Default: Sonnet (quality first)
Example
Task: Add validation logic
→ Analysis: Complex rules + security + error handling
→ Decision: Sonnet (3 criteria met)
Task: Add tags to files
→ Analysis: Template exists, repetitive
→ Decision: Haiku (rule-following)
For detailed criteria, see reference.md For more examples, see examples.md