model-routinglisted

# Model Routing Strategy ## Purpose Define a consistent model routing pattern across all workflows. The same model should not be used for every task — route by task complexity to control cost and latency. ## Routing Tiers | Tier | Task examples | Model requirement | Cost target | |------|--------------|-------------------|-------------| | **Classification** | Prospect eligibility, reply intent, approval status, error classification | Minimum capable — fast, cheap | Lowest available | | **Validation** | Output quality check, contract compliance, format verification | Minimum capable — deterministic checks preferred | Zero (use regex/schema first) | | **Generation** | Draft email content, content creation, metrics summary | Capable — quality matters | Highest budget per call | | **Routing** | Command dispatch, argument parsing, config resolution | Zero — must be deterministic | No model needed | ## Implementation Pattern ``` 1. Classify the task tier (classification / validation / generation / routing) 2. If routing or validation-checkable via regex: no LLM call 3. If classification: use the cheapest model in LLM_PROVIDER_ORDER 4. If generation: use the most capable model in LLM_PROVIDER_ORDER 5. Never default to generation model for classification tasks ``` ## Configuration - `LLM_PROVIDER_ORDER` — ordered list of fallbacks, cheapest to most capable per tier - Instance-specific model names are NOT defined here — they are instance config - This skill defines the pattern