PROMPT TEMPLATE: Few-Shot Classification with Chain-of-Thought

→ STRUCTURE (proven pattern for reliability)
  1. System prompt: Clear role + task specification
  2. Few-shot examples: 4-6 annotated examples (critical for accuracy)
  3. Reasoning instruction: "Think step by step"
  4. Output format: Structured JSON (category, confidence, reasoning)
  5. User input: Ticket to classify

→ COMPLETE TEMPLATE
  System: "You are a support ticket classifier. Assign each ticket to exactly one category. Prioritize accuracy over speed."
  
  Examples:
  EXAMPLE 1:
  Ticket: "My subscription won't renew"
  Category: Billing
  Confidence: 0.95
  Reasoning: Keywords (subscription, renew) directly map to billing issue
  
  EXAMPLE 2:
  Ticket: "App crashes when uploading photos over 50MB"
  Category: Technical
  Confidence: 0.92
  Reasoning: Crash + upload behavior = technical/performance issue, not usage question
  
  EXAMPLE 3:
  Ticket: "Can I export data if I cancel?"
  Category: Account_Management (not Billing alone)
  Confidence: 0.88
  Reasoning: Data export = account feature question (secondary billing context)
  
  [2-3 more examples following same structure]

→ USER INPUT FORMAT
  New ticket: [USER TICKET TEXT]
  Classify using chain-of-thought. Output JSON:
  {"category": "...", "confidence": 0.XX, "reasoning": "..."}

→ TUNING NOTES
  Model: Claude 3.5 Sonnet or equivalent (handles ambiguous tickets better)
  Trap: Zero-shot (no examples) drops accuracy 15-25% on edge cases
  Accuracy: Few-shot pattern achieves 89-94% on typical support queues
  Confidence threshold: Flag tickets below 0.75 confidence for human review
  Iterate: Collect misclassified tickets, add as new examples to improve model