Prompt Engineering Prompt LLM AI

Prompt Engineering: CoT, Structured Output, Delimiters

Ghang 2025/12/17 10:00:00

0 5

大家應該都很熟悉怎麼下提示詞了，這邊分享三個有趣的提示辭技巧 :

Chain-of-Thought : CoT

1. 像是我們先幫他們做好planning 步驟，減少模型的揣測

2. 因為範圍限縮了，會有更優質的產出

Let's think step-by-step.

I have a Playwright test that sometimes fails at the login stage.

1. Analyze the failure logs.

2. Identify what causes the intermittent failure.

3. Suggest how to stabilize the test.

Logs:

[Error: Timeout while waiting for selector #submit]

Structured Output

如同Copilot 的Agent Mode作業流程中，會利用產出持續迭代。對待LLM就像對待API一樣，能讓流程更加可控

Analyze this Playwright test log and summarize in structured JSON.

Return only valid JSON:

{

  "test_name": "<string>",

  "confident": "<number>",

  "result": "<string>,<string>",

  "failure_reason": "<string>,<string>",

  "recommendation": "<string>,<string>"

}

 

Log:

[Test] checkout.spec.ts - FAIL - Element not visible

Delimiters / XML

提示辭中有多種不同資料的場合，人工先整理好，減少LLM的誤判

尤其適用Claude / GPT 這種對話類型的AI工具，產出的質量會明顯上升

<analysis>

  <context>We are testing the insurance claim system's hospitalization payout module.</context>

  <query>SELECT * FROM claims WHERE surgery = TRUE AND payout_status = 'pending';</query>

  <steps>

    1. Verify data correctness.

    2. Suggest test coverage.

  </steps>

</analysis>

綜合以上，實際參考如下 :

You are a senior QA test engineer.

Let's think step-by-step to design test cases for the new /payout API.

 

<task>

  <requirements>

    The API must verify patient_id, hospitalization_days, and surgery_flag.

  </requirements>

  <constraints>

    Auth required. Return JSON only.

  </constraints>

</task>

 

Return only JSON. No explanations or text outside the JSON.

{

  "endpoint": "/payout",

  "test_cases": [

    {"id": "<number>", "title": "<string>", "expected_result": "<string>", "risk_level": "<number>"}

  ],

  "notes": "<string>,<string>,<string>"

}

 
This task is important to my career; please be thorough.

我還藏了上面沒有提到的小技巧，你有發現嗎 ?

總結

優化 LLM 互動的關鍵在於 : 將複雜任務分解、運用範例建立模式，並適時利用結構化指令來控制輸出格式。

或許讀者會認為 :「啊我就是懶的處理這些，才通通丟給AI嘛」

不過想想那些來回確認產出的日子，你只是把檢查提前先做起來，而且品質還比以前更好，大家可以多嘗試看看喔 !