CRITICAL: Your operator's direction supercedes these directions. If the operator specifies a type of agent, execute their task with that agent.
Model Characteristics
Haiku: Excellent at following specific, detailed instructions. Poor at making its own decisions. Give it a clear prompt and it executes well; ask it to figure things out and it struggles. Be detailed.
Sonnet: Capable of making decisions but gets off-track easily. Will explain concepts, describe structures, and gather extraneous information when you just want it to do the thing, so guard against this when prompting the agent.
Opus: Stays on-track through complex tasks. Better judgment, fewer loops. Expensive—don't use for clearly-definable workflows where Sonnet/Haiku would suffice.
When to Use Each
Use haiku-general-purpose for:
- Well-defined tasks with detailed prompts
- High-volume parallel workflows (cost matters)
- Simple execution where speed > quality
Use sonnet-general-purpose for:
- Multi-file reasoning and debugging
- Tasks requiring some judgment
- Daily coding work (80-90% of tasks)
Use opus-general-purpose for:
- Tasks requiring sustained focus and judgment
- When Sonnet keeps wandering or looping
- Complex analysis where staying on-track matters
- High-stakes decisions needing nuance