Few-Shot Template Builder

What this prompt does

Provides two prompts: a meta-prompt that generates high-quality few-shot examples (including edge cases) for any task, and the production prompt template that uses those examples to condition the model at inference time.

When to use it

• Classification, extraction, or structured generation tasks where zero-shot accuracy is insufficient
• When you have a format requirement the model consistently violates without examples
• When you need to demonstrate handling of edge cases (ambiguous inputs, boundary conditions)

When NOT to use it

• Tasks where examples would consume most of the context window (use fine-tuning instead)
• Open-ended creative generation where examples constrain output unnecessarily

The Prompt

Meta-prompt: Generate few-shot examples

Production prompt: Use generated examples at inference time

{{TASK_INSTRUCTION}}

Examples:

Input: {{EXAMPLE_INPUT_1}}
Output: {{EXAMPLE_OUTPUT_1}}

Input: {{EXAMPLE_INPUT_2}}
Output: {{EXAMPLE_OUTPUT_2}}

Input: {{EXAMPLE_INPUT_3}}
Output: {{EXAMPLE_OUTPUT_3}}

Input: {{ACTUAL_INPUT}}
Output:

Why it's structured this way

Meta-prompt for example generation — Writing good few-shot examples is time-consuming and error-prone. Generating them with an LLM is faster and produces more diverse coverage than hand-writing them.

"Include at least one edge case" — Edge case examples have disproportionate value. They teach the model how to handle the inputs that break zero-shot prompts.

"Exactly examples" — Without a count constraint, the meta-prompt generates a variable number of examples. Consistency matters for downstream prompt assembly.

Trailing "Output:" with no value — The open-ended Output: at the end of the production prompt is the in-context learning trigger. The model completes the pattern established by the examples. Adding a newline after Output: without a value prevents the model from generating a preamble before the answer.

Variations