少量範例學習提示詞 — Few-Shot Learning 完全指南

# 任務：分類客戶反饋

Example 1: "產品太好用了！" → POSITIVE
Example 2: "非常棒的服務！" → POSITIVE
Example 3: "我喜歡它！" → POSITIVE

問題：全是短、正面的反饋。無法處理複雜或負面反饋。

# 任務：分類客戶反饋

Example 1: "產品太好用了！" → POSITIVE
Example 2: "界面很直覺，但報表功能需要改進" → MIXED
Example 3: "與競品相比，定價太貴，而且缺少重要功能" → NEGATIVE
Example 4: "客服回應慢，但產品本身不錯" → MIXED
Example 5: "無法與現有系統整合，這是我們的交易破局點" → NEGATIVE

覆蓋：短/長、正面/負面/混合、技術/業務問題

❌ 模糊：
Input: "The market is saturated"
Output: LOW

為什麼輸出是 LOW？可能是市場機會、競爭程度、增長潛力都不清楚。

✓ 清晰：
Input: "市場中已有 20+ 競品，年增長率只有 2%"
Output: LOW_OPPORTUNITY

現在邏輯清晰：飽和市場 + 低增長 = 低機會

# 部分 1：任務說明（清晰、簡潔）
Your task is to [do X with Y].

# 部分 2：評估標準（可選，但推薦）
Evaluation criteria:
- [Criterion 1]
- [Criterion 2]
- [Criterion 3]

# 部分 3：輸出格式
Output format: [JSON / Markdown / etc.]

# 部分 4：範例（3-5 個）
Example 1:
Input: [input]
Output: [output]
Reasoning: [why this output] (可選但有幫助)

Example 2:
Input: [input]
Output: [output]

... [更多範例]

# 部分 5：邊界情況或特殊規則（可選）
Special cases:
- If X, then Y
- If Z, then W

# 部分 6：用戶的實際問題
Now process this:
[actual input to process]

# 不推薦的排序
Example 1: [複雜邊界情況]
Example 2: [簡單情況]
Example 3: [複雜情況]
結果：Claude 被複雜性淹沒

# 推薦的排序
Example 1: [簡單、清晰的情況] ← 建立基礎
Example 2: [中等複雜] ← 增加難度
Example 3: [邊界情況] ← 教會 Claude 異常情況
Example 4: [複雜情況] ← 展示完整能力
結果：Claude 逐步理解，最後掌握複雜情況

Task: Classify customer feedback sentiment with intensity.

Output format:
{
  "sentiment": "POSITIVE | NEUTRAL | NEGATIVE",
  "intensity": 1-5 (5 being strongest),
  "primary_topic": "product | service | pricing | other"
}

Example 1:
Input: "Great product!"
Output: {"sentiment": "POSITIVE", "intensity": 3, "primary_topic": "product"}

Example 2:
Input: "The product works, but customer service never replies to emails. This is frustrating."
Output: {"sentiment": "NEGATIVE", "intensity": 4, "primary_topic": "service"}

Example 3:
Input: "比起 A 公司，你們的定價貴了 30%，但功能更完整。有點值得。"
Output: {"sentiment": "POSITIVE", "intensity": 2, "primary_topic": "pricing"}
Reasoning: 承認缺點但認可價值=溫和正面

Example 4:
Input: "Meh, it's okay I guess."
Output: {"sentiment": "NEUTRAL", "intensity": 1, "primary_topic": "other"}

Example 5:
Input: "Failed to integrate with our CRM. Support couldn't help. Switched to competitor."
Output: {"sentiment": "NEGATIVE", "intensity": 5, "primary_topic": "product"}
Reasoning: 導致流失=最強烈的負面

Task: Review code snippet and rate on APPROVED | NEEDS_CHANGES | MAJOR_REWORK

Example 1:
Input:
```python
def calculate_total(items):
    return sum(item['price'] * item['qty'] for item in items)
```
Output: APPROVED
Reasoning: 清潔、簡潔、正確處理空列表

Example 2:
Input:
```python
total = 0
for item in items:
    total = total + item['price'] * item['qty']
return total
```
Output: NEEDS_CHANGES
Reasoning: 功能正確但應使用內建 sum() 和列表推導式

Example 3:
Input:
```python
total = 0
for i in range(len(items)):
    total += items[i]['price'] * items[i]['qty']
if total < 0:
    total = 0
return total
```
Output: MAJOR_REWORK
Reasoning: 索引訪問低效，負數邏輯可疑（為什麼？），應確認業務規則

Task: Screen job candidates, output STRONG_YES | YES | MAYBE | NO

Context: Hiring for Senior Backend Engineer role, 5+ years required

Example 1:
Input: "10 years backend at Meta, Kubernetes expert, led 3 migrations"
Output: STRONG_YES

Example 2:
Input: "4 years as full-stack dev, knows Python and some ops"
Output: MAYBE
Reasoning: 經驗略短，但有相關背景，可能有潛力

Example 3:
Input: "2 years as junior frontend dev, no backend experience"
Output: NO
Reasoning: 經驗太少，方向錯誤

Example 4:
Input: "6 years backend, strong fundamentals, but last job was in PHP/WordPress"
Output: YES
Reasoning: 經驗充足，基礎扎實，技術棧可學習

Task: Classify news articles → TECH | BUSINESS | SCIENCE | POLITICS | OTHER

Example 1:
Input: "Google releases new quantum processor with 1000 qubits"
Output: TECH

Example 2:
Input: "Apple-EU antitrust settlement reaches $500M agreement"
Output: BUSINESS
Reasoning: 雖然涉及 Apple，但重點是商業協議，不是技術

Example 3:
Input: "Researchers discover new species in Amazon rainforest"
Output: SCIENCE

Example 4:
Input: "新總統上臺，承諾改革稅制"
Output: POLITICS

Example 5:
Input: "AI company raises Series B funding"
Output: BUSINESS
Reasoning: AI 是主題，但融資是商業事件，不是技術突破

Task: Determine proper HTTP status code

Example 1:
Input: User requests resource they don't have permission to view
Output: 403 Forbidden

Example 2:
Input: Server can't connect to database
Output: 503 Service Unavailable

Example 3:
Input: User submits invalid JSON
Output: 400 Bad Request

Example 4:
Input: Endpoint doesn't exist at requested URL
Output: 404 Not Found

Example 5:
Input: Server receives duplicate request within 1 second (idempotency)
Output: 409 Conflict
Reasoning: 重複請求應明確告訴客戶端重試可能有問題

Task: Rewrite in style of technical documentation (clear, direct, no fluff)

Style guidelines:
- Active voice
- Short sentences
- No marketing language
- Define jargon on first use

Example 1:
Original: "The innovative solution leverages cutting-edge technology..."
Rewrite: "This solution uses PostgreSQL and Redis to improve performance."

Example 2:
Original: "Users will be delighted to discover the intuitive interface"
Rewrite: "The interface requires minimal training. Most users complete tasks on first attempt."

Example 3:
Original: "Our world-class team of experts has developed..."
Rewrite: "Engineers built this over 18 months, incorporating feedback from 50+ beta users."

Task: Validate customer data, output VALID | INVALID_FORMAT | INVALID_LOGIC | SUSPICIOUS

Example 1:
Input: {"email": "john@example.com", "age": 30, "signup_date": "2026-04-11"}
Output: VALID

Example 2:
Input: {"email": "not-an-email", "age": 30}
Output: INVALID_FORMAT
Reasoning: 電郵格式錯誤

Example 3:
Input: {"email": "john@example.com", "age": 150, "signup_date": "2026-04-11"}
Output: INVALID_LOGIC
Reasoning: 年齡不合理（超過 120）

Example 4:
Input: {"email": "john@example.com", "age": 25, "signup_date": "2025-01-01", "last_purchase": "2026-04-10"}
Output: SUSPICIOUS
Reasoning: 15 個月內才購買一次=可能不是活躍用戶

用戶問題：「這個複雜產品應該定什麼價格？」

動態選擇相似範例：
1. 找出過去類似「複雜、高價」的定價案例
2. 與通用範例一起提供
3. Claude 看到相似背景會更準確

偽代碼：
examples = [
    generic_example_1,
    generic_example_2,
    find_most_similar(user_input)  # ← 動態選擇
]

驗證流程：
1. 隱藏每個範例的輸出
2. 用 Claude 預測輸出
3. 比較實際 vs 預測
4. 如果準確度 < 90%，範例需要改進