
客戶在北美 要做 ai 自動化 seo 助理
我有點沒方向 也不知道該怎麼做
舉例了兩個平台 Sintara Marblism
分別看了一下介紹





看起來 一個是中控台對話視窗 UI 一個是聊天室 UI 有多個對話
用發訊息的方式 指揮多個 ai 助理自動化完成各種任務
這也太神奇了吧 有點難想像 怎麼做到的呢?
研究之後發現 這是台灣、美國 從業者 對於軟體的基本想像就有不同
首先 Zapier 已經經營超過十年 本身就是各種外部服務串接
台灣做軟體開發 提供 api 是後期才做、或者根本不做的事
但在美國 提供 api 卻是基本、把「API 當成 distribution channel」
讓各種第三方開發者去整合你 這些都是免費幫你推廣產品
美國 SaaS 的基本 expectation 是:
如果沒有 API,就不是 serious SaaS。
例如:如果 CRM 沒 API:
幾乎沒人會選 因為根本不能整合 automation。
台灣多數企業是「UI-first」或「人工作業-first」
美國 SaaS 把 API 當成是產品的一部分(不是附加功能)、生態系從一開始就是「API-first」設計
所以 各種外部服務 幾乎都能串接
台灣的 system integrator model 思維:
客戶付錢 → 客製系統 → 完成
美國的 Platform model 思維:
做平台 → 開 API → ecosystem 成長
API 對 SI 沒價值。
API 對 platform 極度重要。
所有 modern SaaS 都在加:
API
webhooks
automation
Sintra / Marblism 本質就是:
Zapier
+
LLM
Zapier:
user defines workflow
AI assistant:
LLM decides workflow
本質完全相同。
這些「不同類型的助手」本質上不是不同技術系統,而是:
同一個 Agent Runtime + 不同的 prompt + 不同的 tool allowlist + 不同的 workspace context
不是:
不同的 AI 模型
不同的 backend
不同的架構
只是不同「配置」。
好 ai 助手 原理就是 寫一大堆 tool
每個 tool 可以呼叫第三方 api 對吧
這部分我懂了
但是 ai 助手本身的 prompt 大概長怎樣 會需要很長嗎
不需要很長。真正好用的 agent prompt 通常很短,因為「能力」主要在 tools、權限、流程與審核,而不是塞一堆文字規則。
你可以把 prompt 分成 3 層:核心系統規則(短)+ 任務指令(短)+ workspace 設定(資料)。
長的應該是「資料」,不是「規則」。
1) Prompt 需要很長嗎
核心 system prompt:50–200 行以內通常足夠(甚至更短)。
需要長的部分,應該放在:
2) 一個最小可用的 System Prompt 長什麼樣
這種就夠用(你可直接套用):
You are an assistant that completes user requests by calling tools.
Rules:
- If an action affects the outside world (send email, publish post, delete data, charge money), DO NOT execute it directly.
Create a draft and ask for approval.
- Only call tools that are available. If a required capability is not available, ask a concise follow-up question.
- Never claim you performed an action unless a tool result confirms it.
- Prefer using tools over guessing.
- Return clear, structured outputs.
When drafting:
- Match the user's tone.
- Keep drafts concise.
這就是「核心規則」。短、硬、可控。
3) 真正關鍵:你會「動態注入」兩種內容(比 prompt 長度重要)
(A) Workspace Context(可長,但要可裁切)
例如:
這段可以放在 system 或 developer messages,但要有「token 預算」與摘要機制。
(B) Tool Instruction(工具描述與 schema)
每個 tool 的 name/description/schema 就是 prompt 的一部分,但這不是你手寫長文,是結構化資料。
4) 一個實務上更完整但仍然短的版本
Role:
You are an automation assistant for a workspace. Your job is to plan and execute tasks using tools.
Operating principles:
1) Be truthful: only state actions that are confirmed by tool results.
2) Safety gate: for irreversible actions (send/publish/delete/pay), produce a draft and request explicit approval.
3) Minimal questions: ask at most one clarifying question if truly blocked; otherwise proceed with best assumptions and state them.
4) Use tools: prefer tool calls to memory or guessing.
Output format:
- If you created a draft: present the draft + what approval is needed.
- If you need info: ask one question.
- Otherwise: provide the result and next suggested action.
5) Prompt 長的情況通常是「你想把工作流寫死」
例如你想做「客服分流」或「點餐流程」,你會加:
但這其實更適合放在:
而不是全部塞 prompt。
6) Prompt 不應該承擔的事(否則一定越寫越長)
不要用 prompt 去「強制」:
這些都應該在程式層做。
7) 建議你的 MVP prompt 組合(最精準、最好維護)
System(固定,短)
Workspace Profile(可變,資料)
Task Instruction(每次 request)
Tools(結構化)
8) 你要的結論