Lesson 6: Error Recovery & Edge Case Handling

Why Error Recovery Matters

Even the best prompts sometimes produce unexpected results. Mastering error recovery means you can handle edge cases gracefully and get reliable outputs even when things go wrong.

Common Error Types

Mental model: Think of error recovery as having a safety net — you plan for failure so you're never caught off guard.

Technique 1: Defensive Prompting

Anticipate errors and build safeguards into your prompts.

Example:

Write a product description for our new smartwatch.

Requirements:
- Exactly 3 paragraphs
- Include: features, benefits, call-to-action
- Do NOT include: competitor comparisons, pricing, fake statistics
- If you're unsure about a feature, say "up to" or omit it rather than guessing

Format: Markdown with H2 headers for each section.

Defensive techniques:

• Explicitly state what NOT to include
• Provide fallback instructions for uncertain cases
• Specify exact format requirements
• Include validation criteria

Technique 2: Output Validation Prompts

Ask the model to validate its own output before delivering it.

[Original task]

Before providing your final answer:
1. Check that all dates are plausible (not in the future for historical content)
2. Verify that all statistics have reasonable magnitudes
3. Ensure no competitor names are mentioned (unless specifically asked)
4. Confirm the response is within the specified length
5. If any check fails, correct the issue and explain what was wrong

Technique 3: Graceful Degradation

When the model can't complete the full task, get partial results with clear explanations.

Example:

Analyze this dataset and provide insights.

If you cannot complete the full analysis:
1. Explain what information is missing or unclear
2. Provide whatever insights you CAN derive
3. List exactly what additional information would be needed
4. Give a confidence level for the partial analysis

Never guess or make up data to fill gaps.

Technique 4: Retry Logic with Context

When the first attempt fails, provide context about what went wrong.

First attempt:

Convert this data to JSON format:
[Data here]

If output is malformed, retry with:

The previous JSON output had syntax errors. Here's what was wrong:
[Error details]

Please regenerate the JSON, ensuring:
- All strings are properly quoted
- No trailing commas
- Valid JSON syntax throughout

Here's the data again:
[Data here]

Technique 5: Boundary Testing

Explicitly test edge cases in your prompts.

Write a function that calculates the average of a list of numbers.

Handle these edge cases:
- Empty list: Return 0 or raise appropriate error
- Single element: Return that element
- Negative numbers: Include in average
- Zero values: Include in average
- Very large numbers: Handle without overflow

Show the function and test cases for each edge case.

Handling Specific Error Types

Fixing Hallucination

[Original prompt]

IMPORTANT: Only include information you are confident about.
- If you're unsure about a specific fact, say "I'm not certain, but..."
- Do not make up statistics, dates, or names
- If you cannot answer confidently, say so rather than guessing

Fixing Format Errors

Return your response as valid JSON with this exact structure:
{
  "key1": "string value",
  "key2": number,
  "key3": ["array", "of", "strings"]
}

Before outputting, verify:
- All braces and brackets are balanced
- All strings use double quotes
- No trailing commas
- Numbers don't have quotes around them

Fixing Incomplete Output

Write a comprehensive guide about [topic].

IMPORTANT: This must be complete. Do not stop mid-sentence or say "to be continued."
If the content is long, structure it with clear sections so I know it's complete.
End with a clear conclusion or summary.

Building Error Recovery Into Workflows

Workflow Template:

Step 1: Generate initial response
Step 2: Validate against criteria:
  - [ ] Format is correct
  - [ ] All facts are verifiable
  - [ ] Response is complete
  - [ ] No off-topic content
Step 3: If validation fails:
  - Identify specific issues
  - Retry with corrective instructions
  - Maximum 2 retries before asking for clarification
Step 4: Final delivery with confidence score

Common Mistakes to Avoid

• Mistake: Not specifying what "success" looks like — Fix: Always define validation criteria upfront.
• Mistake: Giving up after one failed attempt — Fix: Use retry logic with specific corrective feedback.
• Mistake: Ignoring edge cases — Fix: Explicitly test boundary conditions.
• Mistake: Assuming the model knows your constraints — Fix: State all requirements explicitly.

Professional Tips & Tricks

• Build a library of error recovery prompts for common issues.
• For critical workflows, always include a validation step.
• Track which errors occur most frequently — they become prevention items.
• Use "defensive prompting" by default for important content.

Key Takeaways

• Defensive prompting anticipates errors and builds safeguards.
• Output validation catches issues before delivery.
• Graceful degradation gets partial results when full completion isn't possible.
• Retry logic with context helps fix errors efficiently.
• Boundary testing ensures your prompts handle edge cases.

Next up: Building complete AI workflows with prompt chaining and orchestration.