using-generic-agents_skill

Overview

This skill helps you choose which generic agent variant to run for a given task by mapping task characteristics to three agent profiles: Haiku, Sonnet, and Opus. It distills trade-offs between cost, focus, and autonomy so you pick an agent that matches task complexity and budget. The goal is faster, cheaper runs for routine jobs and more capable agents for high-stakes or long-horizon work.

How this skill works

The skill inspects task prompts, success criteria, and constraints (cost, parallelism, tolerance for drift) and recommends one of the three agent profiles. It highlights why an agent was chosen and lists conditions that would change the recommendation. If an operator specifies an agent explicitly, that direction is honored immediately.

When to use it

• Selecting an agent before running automated workflows or tests
• Planning resource allocation across many parallel jobs
• Choosing an agent for debugging, code review, or local development tasks
• Deciding when to escalate to a more capable model for critical decisions
• Standardizing agent selection across a team or CI pipeline

Best practices

• Provide clear, specific prompts for Haiku to get reliable, low-cost execution
• Use Sonnet for everyday development tasks that need judgment but not extreme focus
• Reserve Opus for complex, high-risk, or multi-step analysis where staying on-track matters
• Start with the recommended agent and only escalate if results loop or wander
• Document success criteria and constraints so the skill can make consistent recommendations

Example use cases

• Run Haiku for bulk data formatting jobs and simple API integrations where speed and cost matter
• Use Sonnet for multi-file debugging, feature implementation, and codebase reasoning
• Choose Opus for architecture reviews, research synthesis, or decisions requiring sustained context
• Automate agent selection in CI: Sonnet for test triage, Opus for complex failure diagnosis
• Scale large batches by assigning Haiku to high-volume, well-specified tasks and Sonnet for edge cases