route

Model Routing Guidance

Use this skill to decide which Claude model tier to dispatch a task to as a subagent. The three tiers — Haiku, Sonnet, and Opus — differ in cost, speed, and reasoning depth.

Quick Reference

Tier	Speed	Cost	Best for
Haiku	Fastest	Lowest	Simple, deterministic, single-concern tasks
Sonnet	Balanced	Moderate	Standard dev tasks, moderate complexity
Opus	Slowest	Highest	Complex reasoning, ambiguous multi-constraint problems

Default to Sonnet when uncertain. Upgrade to Opus only when Sonnet would likely produce shallow results. Downgrade to Haiku only when the task is clearly bounded and requires no judgment.

Decision Heuristics

Use Haiku when the task is:

• Reading a single file and summarizing it
• Reformatting or linting output
• Simple lookups (grep a value, count lines, check if X exists)
• Yes/no or short-answer checks
• Extracting structured data from a known format
• Running a quick transformation with no ambiguity

Use Sonnet when the task is:

• Writing or editing code across one or a few files
• Moderate research or codebase exploration
• Standard bug fixes with a clear scope
• Generating tests for existing functions
• Explaining a module or component
• Anything that doesn't clearly fit Haiku or Opus

Use Opus when the task is:

• Designing system architecture or making major structural decisions
• Debugging a hard, non-obvious issue where root cause is unknown
• Reviewing critical code (security, data integrity, financial logic)
• Synthesizing information across many files or sources into a coherent plan
• Tasks with multiple competing constraints requiring trade-off reasoning
• Any task where you tried Sonnet and the result was too shallow

Invocation

Use the /dispatch:haiku, /dispatch:sonnet, or /dispatch:opus slash commands to dispatch a task directly to a specific model tier.

Alternatively, use the Agent tool directly with the model parameter:

Agent({ model: "haiku", description: "...", prompt: "..." })
Agent({ model: "sonnet", description: "...", prompt: "..." })
Agent({ model: "opus", description: "...", prompt: "..." })

Additional Resources

• references/decision-matrix.md — Extended examples per tier with anti-patterns