prompt-engineering-patterns

Select 1-10 examples using semantic similarity retrieval with embedding models, ensuring consistent input-output formatting across all examples. Balance example count against your context window budget, typically allocating around 38% of tokens to examples.

Define a Pydantic schema describing the expected fields and use structured output mode or JSON mode in your prompt. Add error handling with a fallback prompt that retries with a simpler instruction if JSON parsing or schema validation fails.

Chain-of-thought prompting elicits step-by-step reasoning by adding phrases like "Let's think step by step" or providing reasoning examples. Use it for math, logic, and multi-step tasks, but skip it for simple lookups or latency-sensitive applications.

Run both prompts against the same test suite and measure accuracy, latency, and token usage. Use a statistical significance test such as a t-test with p < 0.05 to confirm the winning variant's improvement is not due to chance.

Inconsistency usually comes from ambiguous instructions, missing output format constraints, or high sampling temperature. Add explicit format requirements, include examples, and measure consistency by running identical inputs multiple times.

Avoid few-shot prompting when examples do not match the target task, when token limits are tight, or when the model already performs well zero-shot. Poorly chosen examples can cause the model to overfit to example patterns.