Format Examples — Controlling Output Structure with Few-Shot

Core 6 min +25 XP

💡

THE ANALOGY

A style guide with annotated examples. Telling writers 'write in AP style' produces variability. Showing them three exemplary articles with annotations produces consistent house style. Few-shot format examples do the same for Claude's output.

⚠️ EXAM TRAP — The Wrong Answer People Choose

Using few-shot only for content guidance and ignoring its power for format control. For complex nested JSON, showing examples is often more reliable than describing the schema.

KEY POINTS

1 Format examples show exact output structure — more reliable than format descriptions for complex structures.

2 All examples must use identical format — variation in examples creates variation in output.

3 Include null/missing cases explicitly — prevents Claude from inventing values.

4 Combine few-shot format examples with schema validation for production reliability.

5 Input/output pair examples teach both content judgment AND format simultaneously.

Format Examples vs Format Instructions

# ❌ Instructions only — produces variable nesting
instructions = """
Extract parties from contracts. Return JSON with:
- parties: array of party objects with name, role, and address
- Each party may have a 'represented_by' attorney field
"""
# Result: Claude nests 'represented_by' differently each time

# ✅ Examples show exact structure — produces consistent output
few_shot = """
Extract contract parties. Follow this EXACT structure:

EXAMPLE 1 (two parties, one with attorney):
Contract: "Between Acme Corp ('Buyer') and TechCo Inc ('Seller')
represented by counsel Smith & Jones LLP, effective January 2024."
Output:
{
  "parties": [
    {
      "name": "Acme Corp",
      "role": "buyer",
      "represented_by": null
    },
    {
      "name": "TechCo Inc",
      "role": "seller",
      "represented_by": "Smith & Jones LLP"
    }
  ],
  "party_count": 2
}

EXAMPLE 2 (three parties, none with attorneys):
Contract: "Agreement between GlobalCorp ('Licensor'), RegionCo ('Licensee'),
and PayCo ('Payment Agent')."
Output:
{
  "parties": [
    {"name": "GlobalCorp", "role": "licensor", "represented_by": null},
    {"name": "RegionCo", "role": "licensee", "represented_by": null},
    {"name": "PayCo", "role": "payment_agent", "represented_by": null}
  ],
  "party_count": 3
}

Now extract parties from: {contract_text}
"""

Example 1 explicitly teaches null for missing represented_by — preventing Claude from omitting the field entirely or inventing an attorney.

Format Consistency Across Examples

# ❌ Inconsistent — creates inconsistent output
examples = [
    # Example 1: date as string
    {"effective_date": "2024-01-15"},
    # Example 2: date as object
    {"effective_date": {"year": 2024, "month": 1, "day": 15}},
]
# Claude will sometimes use one format, sometimes the other

# ✅ Consistent — output is predictable
examples = [
    {"effective_date": "2024-01-15"},   # ISO string
    {"effective_date": "2024-03-22"},   # ISO string
    {"effective_date": None},           # null when missing
]

Key Takeaways

Show structure rather than describe it for complex nesting
null/missing cases must be in examples — prevents fabrication
Identical format across all examples — no variation
Two examples minimum to show pattern generalization
Combine with schema validation for production safety