100 Prompt Engineering resources for developers

Instruct the model to 'think step-by-step' before providing a final answer to improve performance on logic and math tasks.

Provide 3-5 high-quality examples of input-output pairs within the prompt to define specific formatting and stylistic requirements.

Use tags like <context>, <instruction>, and <examples> to help models like Claude 3.5 Sonnet distinguish between metadata and task data.

A multi-step process where the model generates a response, identifies its own potential errors, and verifies facts before final output.

Prompt the model to generate a high-level outline first, then expand each section to reduce latency and improve structural coherence.

Explicitly list forbidden words, topics, or formats (e.g., 'Do not use markdown') to prevent unwanted output behavior.

Define a specific professional role (e.g., 'You are a Senior Site Reliability Engineer') to prime the model for specific technical jargon and tone.

Generate multiple outputs for the same prompt and use a majority vote or LLM-as-a-judge to select the most accurate result.

Using RAG or sliding windows to pass only the most relevant 20-30% of document context to avoid the 'lost in the middle' phenomenon.

Force the model to return valid JSON by providing a strict schema, often implemented via libraries like Instructor.

A matrix testing tool that allows you to run prompts against multiple models and check outputs against predefined assertions.

A platform for tracing, debugging, and evaluating LLM applications with integrated dataset management for regression testing.

An enterprise-grade stack for logging LLM calls, managing prompt versions, and running automated evaluation workflows.

A Python framework for unit testing LLM outputs based on metrics like faithfulness, relevance, and hallucination scores.

An open-source observability proxy that tracks costs, latency, and token usage for OpenAI and Anthropic requests.

A middleware for logging and versioning prompts, allowing developers to roll back to previous prompt iterations without code changes.

An open-source library for detecting vulnerabilities, hallucinations, and biases in LLM-based applications.

Distributed tracing for LLMs to understand the latency impact of each step in a complex multi-agent chain.

A suite of tools for visualizing and inspecting the execution flow of LLM chains and pipelines.

Open-source observability for RAG and LLM applications, focusing on embedding visualization and retrieval evaluation.

The industry standard for getting structured data (JSON) from LLMs using Pydantic models for Python or Zod for TS.

A unified interface for generating type-safe JSON objects across OpenAI, Anthropic, and Google Gemini models.

A library for neural text generation that uses regex and context-free grammars to guarantee 100% valid JSON or code.

A programming language that combines logic programming with LLM generation to constrain output at the token level.

A template language for controlling LLMs, allowing you to interleave generation, control flow, and fixed strings.

A lightweight library that uses LLMs to power standard Python functions, handling the prompt engineering under the hood.

Microsoft's library for replacing complex prompt engineering with schema-based type definitions to guide model output.

A library specifically designed to fill in the values of a JSON schema while the structure is pre-defined, saving tokens.

A fast backend and frontend for structured generation, optimized for high-throughput serving and complex control flow.

A control plane to manage multiple LLMs with built-in retries, fallbacks, and load balancing across providers.