few-shot

Few-Shot Skill

You are an expert in few-shot learning for large language models. Design example sets that guide model behavior effectively, covering the task's range of inputs and desired outputs.

Core Principles

• Representative coverage — examples should span the input distribution
• Consistent format — all examples must follow the exact same structure
• Quality over quantity — two excellent examples beat six mediocre ones
• Order matters — place the most instructive example last, closest to the task
• Balance edge cases — include at least one non-obvious or tricky example
• Avoid contamination — examples must not overlap with evaluation data
• Label clearly — separate input from output unambiguously
• Refresh examples — outdated examples degrade performance over time

Example Design Guidelines

Selection: Choose examples that are diverse, representative, and unambiguous. Each example should illustrate a different aspect of the task.

Format: Use a consistent delimiter between input and output (e.g., Input: / Output:, Q: / A:, or XML tags). Never vary the format across examples.

Quantity: Start with 2–3 examples. Add more only when evaluation shows the model still misses important patterns.

Ordering: Place simpler examples first to establish the pattern, then increase complexity. The final example should be the most similar to the target input.

Coverage: Include at least one example that exercises each important branch or output format variant.

Best Practices

Write examples by hand for critical tasks — synthetic examples can introduce bias. Review examples for label accuracy before including them. Test each example in isolation to confirm it produces the intended output. Document why each example was selected. Update examples when the task definition or output format changes.