Smart Routing Decision Framework

When This Activates

Use this framework BEFORE starting complex work to determine the best model:

• Should this use Opus (architecture), Sonnet (default), or Haiku (exploration)?

• Local Ollama is ONLY for non-critical tasks (commit messages, Enchanted API)

IMPORTANT: Local LLM is NOT for Critical Work

DO NOT route development tasks to local Ollama. Local is only appropriate for:

• Commit message generation

• PR description drafts

• Enchanted/mobile API access

• Personal tinkering

Quick Decision Matrix

Task Type Route To Model

"Where is X?" CLAUDE Haiku (fast exploration)

"What does Y do?" CLAUDE Haiku (simple lookup)

"How does Z work?" CLAUDE Sonnet (reasoning needed)

Code review CLAUDE Sonnet (quality matters)

Commit message Local (non-critical)

Multi-file refactor CLAUDE Sonnet

Debug intermittent bug CLAUDE Sonnet

Architecture decision CLAUDE Opus

Write new feature CLAUDE Sonnet

Fix across codebase CLAUDE Sonnet

Model Selection Guide

Use Opus when:

• Greenfield architecture decisions

• Multiple valid approaches with non-obvious trade-offs

• High cost to reverse if wrong choice

• Cross-cutting concerns (security, scalability, performance)

Use Sonnet when (DEFAULT):

• ALL implementation work

• Bug fixes and debugging

• Feature additions

• Refactoring

• Code review

• Most planning tasks (75%)

Use Haiku when:

• File/code exploration

• Simple lookups ("where is X?")

• Documentation updates

• Simple formatting

Use Local ONLY when:

• Commit messages (mlx commit )

• PR descriptions (mlx pr )

• Enchanted app queries

• Personal experimentation

Tool-Need Detection

Any task requiring tools should use Claude (never local):

• edit , modify , change , update , fix (file operations) → Sonnet

• create file , write file , new file → Sonnet

• commit , push , pull request → Sonnet (local for msg only)

• run test , execute , build , deploy → Sonnet

• install , npm , pip , yarn → Sonnet

Simple Question Patterns

For quick lookups, use Haiku:

• "what is/does/are..." → Haiku

• "where is/are/do..." → Haiku

• "find", "search", "look for", "locate" → Haiku

For reasoning questions, use Sonnet:

• "how does/do/to..." → Sonnet

• "why does/is/do..." → Sonnet

• "explain", "describe" (complex) → Sonnet

Workflow

• Default to Sonnet for all development work

• Use Haiku only for simple exploration/lookups

• Use Opus only for complex architecture decisions

• Use Local only for commit messages and non-critical tasks

MCP Tools (All Use Claude)

Memory tools (use with appropriate Claude model):

• memory_query

• Instant hybrid search (Haiku OK)

• memory_functions

• Function lookup (Haiku OK)

• smart_read

• File reading (Sonnet for analysis)

• smart_edit

• File editing (Sonnet always)

Local tools (non-critical ONLY):

• local_ask

• ONLY for commit messages, personal queries

• local_review

• NOT for critical code review

Quality Over Cost

• Critical work: Always use Claude (Sonnet default)

• Architecture: Use Opus - worth the cost for high-stakes decisions

• Exploration: Haiku is cheap and good enough

• Trivial: Local is fine for commit messages only