prompt-engineering

Use native JSON mode on OpenAI (response_format json_object) or tool use on Anthropic Claude with an input_schema defining your structure. Validate responses with Pydantic in Python or Zod in TypeScript instead of parsing raw text.

Use chain-of-thought for complex reasoning tasks like math, logic puzzles, and multi-hop question answering, where showing intermediate steps improves accuracy by 20-50%. It is unnecessary for simple factual recall or basic classification.

Use 2-5 diverse, high-quality examples; performance typically plateaus after 5-7 examples. Include edge cases, keep formatting consistent, and randomize example order to avoid position bias.

LangChain suits complex Python workflows like RAG pipelines, agents, and multi-step chains with LCEL composition. Vercel AI SDK is better for TypeScript/React apps needing streaming, type-safe structured outputs, and multi-provider support.

Use Anthropic prompt caching with cache_control on large repeated context blocks for up to 90% cost reduction on cached tokens. Also count tokens before execution, set temperature appropriately, and cache prompt prefixes across chained calls.

Sanitize user input by detecting patterns like 'ignore previous instructions', wrap user content in delimiters or XML tags, and instruct the model in the system prompt to ignore instructions embedded in user text. The included prompt-validator script automates these checks.