Claude Code 如何建立模型路由器，為模型選擇評分
Your model choice is a bet. Build a router that keeps score
作者為 Claude Code 製作了一個小型模組，分析用戶的第一個提示詞，詢問 Haiku 任務需要哪個模型及 effort，再在聊天旁顯示評分。每項評分附有短海報和 WHY 方框；文章亦説明評分準則、模組組成、切換模型的快取成本，以及在 Claude Desktop app 建立自訂模組的步驟。成本數字是作者估算，實際數字會因人而異。
Most of the time I pick the same Claude model I used last session, out of habit. So I built a small Claude Code mod that scores that habit.
It reads your first prompt, asks Haiku which model and effort the task needs, and shows a verdict beside the chat: on target, one step off, overshot or undershot. Each verdict comes with a short poster and a WHY box. The WHY box is the part I learn from.
The full post covers the rubric I use, how the mod is put together, what a model switch costs in cache terms, and a recipe you can paste into the Claude Desktop app to build your own. The cost numbers are my own estimates, so yours will differ.
Read the full post, with the recipe and the rubric: https://digitaldias.com/blog/2026-11-05-model-choice-is-a-bet/
來源：DEV Community · dev.to