Few-Shot Prompting Techniques

Core

Apply few-shot prompting to improve output consistency and quality · Difficulty 2/5

few-shotexamplespromptingconsistency

Prerequisites

Explicit Criteria over Vague Instructions

Few-Shot Prompting provides concrete examples of desired input-output behavior to guide the model. It is the most effective technique when instructions alone produce inconsistent output.

When to Use Few-Shot Examples

Output format must be precise and consistent (e.g., JSON, structured findings)

Instructions alone produce variable results

The task involves nuanced classification or ambiguous-case handling

You need to demonstrate reasoning patterns, not just outputs

Extraction tasks where hallucination is a risk

Best Practices

Target ambiguous cases: Focus examples on edge cases where the model struggles, not obvious cases

Show reasoning: Include WHY the example output is correct, not just WHAT it is

Use 2-4 targeted examples: Quality over quantity; cover different ambiguous scenarios

Diversify examples: Cover different scenarios (tool selection ambiguity, document structure variation, acceptable vs problematic patterns)

Generalization, Not Matching

Few-shot examples enable the model to generalize judgment to novel patterns. They teach the decision-making approach, not just a lookup table of pre-specified cases. This is why showing reasoning in examples is critical.

Reducing Hallucination in Extraction

For extraction tasks with varied document structures, few-shot examples showing correct handling of informal measurements, missing fields, and varied formats significantly reduce fabricated data.

Key Takeaways

✓Few-shot examples are more reliable than instructions for consistent formatting
✓Target examples at ambiguous cases, not obvious ones
✓Include reasoning in examples so the model generalizes judgment, not just pattern-matches
✓2-4 well-chosen examples outperform 10+ unfocused ones

Glossary Terms

Few-Shot Prompting

A prompting technique that provides concrete examples of desired input-output behavior to guide the model. More reliable than instructions alone for consistent formatting and nuanced decisions. Examples should target ambiguous edge cases, not just obvious ones.

Related Concepts

Concrete Input-Output Examples

Concrete examples eliminate ambiguity that prose descriptions create

Classification Consistency & False Positive Reduction

Use absolute criteria with concrete examples for each classification level

Test Yourself1 of 3

Your agent achieves 55% first-contact resolution, well below the 80% target. Logs show it escalates straightforward cases (standard damage replacements with photo evidence) while attempting to autonomously handle complex situations requiring policy exceptions. What's the most effective way to improve escalation calibration?