Retry Patterns — Three-Component Retry for Effective Error Recovery

Core 6 min +25 XP

💡

THE ANALOGY

A proofreader marking a draft. Not 'this is wrong' — but specific marks on each error with the correction. The author addresses each mark precisely. Claude's retry needs the same specificity: field name, what it got, what's expected.

⚠️ EXAM TRAP — The Wrong Answer People Choose

Retrying with the original prompt unchanged — produces identical wrong output. The retry must include all three: original document, previous failed output, AND specific error descriptions.

KEY POINTS

1 Three required retry components: original document + failed output + specific errors.

2 Without the failed output: Claude doesn't know what to change.

3 Without specific errors: Claude doesn't know what's wrong.

4 Error format: field name + what was extracted + what's expected.

5 Never retry when source document lacks the information — produces fabrication instead.

The Three-Component Retry

def build_retry_prompt(document: str, failed_output: dict, errors: list) -> str:
    """All three components are required — missing any one degrades retry effectiveness."""
    error_text = "\n".join([
        f"• {e['field']}: {e['error']}"
        + (f"\n  Got: {e['got']}" if 'got' in e else "")
        + (f"\n  Expected: {e['expected']}" if 'expected' in e else "")
        for e in errors
    ])

    return f"""Your previous extraction failed validation. Fix the specific errors listed.

ORIGINAL DOCUMENT:
{document}

YOUR PREVIOUS (INCORRECT) EXTRACTION:
{json.dumps(failed_output, indent=2)}

ERRORS TO FIX:
{error_text}

Fix only these errors. Do not change fields that were correct.
If a value is not in the document, return null — do not invent values."""

Specific Error Messages

# ❌ Vague — Claude can't fix this
{"field": "invoice_date", "error": "Invalid format"}

# ✅ Specific — Claude knows exactly what to change
{
    "field":    "invoice_date",
    "error":    "Must be ISO format YYYY-MM-DD",
    "got":      "January 15, 2024",
    "expected": "2024-01-15"
},
{
    "field":    "total_amount",
    "error":    "Must be a decimal number in dollars, not a string with currency symbol",
    "got":      "$4,999.00",
    "expected": 4999.00
}

When Not to Retry

# Source-missing errors produce fabrication on retry — never retry these
NOT_RETRYABLE = [
    "not found in source document",
    "not in the document",
    "document does not contain",
    "absent from"
]

def should_retry(error: dict) -> bool:
    return not any(p in error.get("error", "").lower() for p in NOT_RETRYABLE)

Key Takeaways

All three components: document + failed output + specific errors
Field + got + expected in every error message
Never retry missing-data errors — fabrication
Cap at 2-3 retries — escalate to human review after
“Fix only listed errors” — targeted correction, not re-do