Save Money

**Judgement:** A well-designed, practical routing skill for Anthropic Claude that focuses on cost savings without sacrificing answer quality. It is strong when used exactly as intended (Haiku + Sonnet with sessions_spawn available), but is less plug-and-play if your stack uses different model names or providers. **What it does well** - Enforces a clear, concrete decision rule: simple, short, single-step tasks stay on Haiku; anything analytical, multi-step, long, or structured *must* escalate to Sonnet. - Provides detailed escalation triggers and multilingual examples, which likely improves correct routing across many real-world prompts. - Includes explicit prohibitions on doing complex work on Haiku (no long code, no tables, no reports, no multi-step reasoning), which should reduce low-quality answers on the cheap model. **Key risks / limitations** - **Provider/model specificity:** The logic and example model IDs are tailored to Claude (Haiku / claude-sonnet-4-20250514). In environments without those exact models, you’ll need to adapt the configuration; otherwise, sessions_spawn calls may fail. - **Heuristic classification:** Complexity detection is rule-based, not dynamic. Some edge cases will still be misrouted (e.g., subtle but complex questions that look short, or verbose but simple requests). - **Potential over-escalation:** The “when in doubt, escalate” rule favors safety and quality over maximal cost savings; costs might be higher than absolutely necessary in borderline cases. - **Instruction interaction:** If combined with other skills that also try to control routing or enforce long-form behavior on the base model, there’s potential for conflicting instructions. **Best-fit scenarios** - A Claude-based assistant where you explicitly want Haiku as the default and Sonnet only for clearly complex work. - High-traffic org assistants (support, ops, engineering Q&A) where total token cost matters but you can’t compromise on complex answers. - Developers comfortable with tweaking model names who want a ready-made, opinionated routing policy to adapt for OpenAI/Gemini equivalents.