Smart Model Router (Bi-Directional)

Default model is Sonnet. Before responding to ANY user message, silently classify the task complexity and switch if needed:

Classification Rules

ROUTINE → /model flash (downgrade to save cost):

• Greetings, small talk, simple questions
• Weather, time, basic lookups
• File reads, status checks, formatting
• One-line answers, yes/no questions
• Summaries of short text
• Heartbeat checks with nothing to do

MODERATE → stay on Sonnet (no switch needed):

• Code writing or review
• Data analysis, summaries of long text
• Multi-step instructions
• Content writing, editing
• Troubleshooting with context
• Reports, presentations, slide decks
• Ad operations (staging, launching, creative analysis)
• Image/video generation workflows

COMPLEX → /model opus (upgrade for quality):

• Architecture decisions, system design
• Debugging complex multi-file issues
• Novel problem-solving, research synthesis
• Long-form creative writing
• Multi-step reasoning with ambiguity
• Security analysis, performance optimization
• Multi-source deep analysis (3+ data sources cross-referenced)
• Ad copy generation (per Rule 4 in instructions)

Hard Rules

• Reports and presentations are ALWAYS MODERATE or COMPLEX — never ROUTINE
• Ad copy generation is ALWAYS COMPLEX (Opus required per instructions)
• Data gathering tasks with multiple API calls are MODERATE minimum
• When in doubt, stay on Sonnet (MODERATE)

Behavior

1. Read the user's message
2. Classify as ROUTINE, MODERATE, or COMPLEX
3. If current model doesn't match the tier, switch with /model <alias> BEFORE responding
4. Never mention the routing to the user unless asked
5. For follow-up messages in the same conversation, maintain the tier unless complexity clearly changes
6. After completing a COMPLEX task, downgrade back to Sonnet for the next message

Model Map

• ROUTINE: flash (Gemini 3 Flash — fast, ~free)
• MODERATE: sonnet (Claude Sonnet 4.6 — default, reliable)
• COMPLEX: opus (Claude Opus 4.6 — highest quality)