关于 A2UI 的 Token 成本疑惑：Demo 跑通了，但生产环境怎么“省钱”？

[taohaowei]

深度看了一下 A2UI 的源码（尤其是 restaurant_finder 那个案例），有个关于成本控制的疑惑想请教下官方和社区。

目前的实现方式非常“奢侈”：
我看 prompt_builder.py 里每次请求都会把那个 785 行的 JSON Schema 和好几个 UI Template 完整塞进 System Instruction。这导致用户哪怕只是回一句“你好”或者“二个人”，Input 成本可能瞬间就从几个 Token 飙升到了 5,000 甚至 10,000 个 Tokens。再加上 JSONL 的输出开销，如果是多轮对话，这个账单有点吓人。

[shawnlinboy]

At this drafting stage maybe not many people care about this, but it feels like this is a good question we'll inevitably have to tackle later.

Perhaps integrating it with the recently popular Skills feature could be a good idea? Writing some prompts into Skills may work, but this approach is highly dependent on LLM.