prompt-engineering-patterns

Define a Pydantic schema describing the expected fields, then instruct the model to respond with JSON matching that schema. LangChain's with_structured_output method automates this binding, and you should add fallback handling for JSON decode or validation errors.

Use chain-of-thought prompting by adding 'Let's think step by step' or providing few-shot examples with explicit reasoning traces. For higher reliability, apply self-consistency by sampling multiple reasoning paths and taking the majority answer.

Use semantic similarity selection with embedding models to retrieve examples closest to each query, or diversity sampling via clustering to cover edge cases. Balance example count against your context window token budget.

Skip chain-of-thought for simple factual lookups, direct retrieval, creative writing, and latency-sensitive real-time applications. CoT adds tokens and response time, so reserve it for math, logic, multi-step planning, and debugging tasks.

Run both prompts against the same test suite, measuring accuracy, latency, and token usage per variant. Use a statistical significance test such as a t-test on the metric distributions before declaring a winner.

Inconsistency usually comes from ambiguous instructions, missing output format constraints, or examples that do not match the target task. Add explicit format requirements, edge-case examples, and a verification step, then measure consistency across repeated runs.