Smart Model Switching

**Quick judgment**: A well-designed, practical routing policy for Claude-only stacks that will significantly reduce costs for mixed workloads while preserving access to stronger models for demanding tasks. It’s simple, opinionated, and easy to operationalize. **What it does well** - Starts everything on Haiku and escalates to Sonnet/Opus only when complexity or risk justify the extra cost. - Provides concrete triggers (code length, paragraphs, task types, “>30 seconds of human thinking”) that are straightforward for both humans and agents to follow. - Documentation (decision tree, quick reference) is unusually clear, making integration and maintenance easier. **Key risks / limitations** - **Claude-only**: Hard-wired to Haiku/Sonnet/Opus, so it’s not directly reusable across other providers or model sets without adaptation. - **Heuristic misclassification**: Edge cases (e.g., short but high-stakes tasks, or deceptively simple prompts needing deep expertise) may stay on Haiku or Sonnet when Opus-level reasoning would be safer. - **No built-in feedback loop**: The guideline “if Sonnet struggles, go to Opus” assumes some external signal or human judgment; the skill itself doesn’t define how struggle is detected. **Recommended scenarios** - Claude-centric agent systems with diverse workloads where API cost is a real concern. - Tools that do a lot of routine Q&A and status checks, with occasional heavier coding or architectural work. - Environments where developers want a simple, documented policy for when to pay for stronger models, rather than ad-hoc or model-hardcoded decisions.