Building Robust Prompts
Production prompts must handle the unexpected. Users will provide malformed input, ask edge-case questions, and trigger scenarios you didn't anticipate.
💡 Production Reality: The prompts that work perfectly in testing often fail in production because they don't handle edge cases.
Common Failure Modes
Missing Input
User provides empty or null values
Malformed Input
Input doesn't match expected format
Out of Scope
Request outside your domain
Ambiguous Request
Unclear what the user wants
Defensive Prompt Patterns
Pattern 1: Input Validation Instructions
Before processing the request:
1. Check if the input is a valid product URL
2. If the URL is malformed, return: {"error": "invalid_url", "message": "Please provide a valid product URL"}
3. If the URL is from an unsupported site, return: {"error": "unsupported_site", "message": "We only support Amazon, eBay, and Walmart"}
4. Only proceed with analysis if validation passes
Pattern 2: Fallback Responses
Classify the customer inquiry into one of these categories:
- billing
- technical_support
- product_question
- complaint
- other
If you cannot confidently classify the inquiry (confidence < 70%),
return "needs_human_review" and explain why in the reason field.
Return format:
{
"category": "string",
"confidence": 0.0-1.0,
"reason": "string (only if needs_human_review)"
}
Pattern 3: Explicit Edge Case Handling
Extract the price from the product description.
Special cases to handle:
- If multiple prices exist, use the current/sale price (not original)
- If price range (e.g., "$10-$20"), use the lower bound
- If price is "Free" or "$0", return 0
- If no price found, return null
- If price is in foreign currency, return as-is with currency code
Return: {"price": number|null, "currency": "USD"|"EUR"|etc, "notes": "string if any special handling applied"}
Graceful Degradation
When the AI can't fully complete a task, it should provide partial results rather than failing completely.
Analyze this financial document and extract:
1. Revenue figures
2. Expense breakdown
3. Profit margins
4. Year-over-year growth
For each field:
- If clearly stated: extract the value
- If can be calculated: show the calculation
- If partially available: extract what's possible and note what's missing
- If unavailable: mark as "not_found" with reason
Never fail completely - always return as much as you can extract.
Error Response Formats
When you encounter an error, return a structured error response:
{
"success": false,
"error": {
"code": "ERROR_CODE",
"message": "Human-readable message",
"details": "Additional context",
"suggestion": "What the user can do to fix it"
}
}
Error codes:
- INVALID_INPUT: Input doesn't match expected format
- MISSING_REQUIRED: Required field is missing
- OUT_OF_SCOPE: Request outside supported functionality
- AMBIGUOUS_REQUEST: Cannot determine user intent
- PROCESSING_ERROR: Error during processing
🔑 Key Takeaway: Anticipate failures before they happen. Build prompts that validate inputs, handle edge cases explicitly, and degrade gracefully when full completion isn't possible.