Large language models (LLMs) are revolutionizing how we interact with AI. However, crafting effective prompts and managing the complex interactions between LLMs, tools, and data can be a significant hurdle. Enter PDL, the Prompt Declaration Language, a new approach designed to streamline and simplify LLM programming. Imagine building sophisticated LLM applications without wrestling with intricate code. PDL offers precisely that: a simple, declarative language based on the familiar YAML format. This means you can define prompts, chain LLM calls, integrate external tools, and manage data flow, all within a clean, human-readable structure. PDL's magic lies in its data-oriented design. Each block in a PDL program contributes data to a context that serves as the input for subsequent LLM calls. This intuitive approach simplifies complex prompting patterns like chatbots and agents, where maintaining conversation history is crucial. Forget about tedious manual plumbing – PDL handles it seamlessly. Beyond its ease of use, PDL promotes robustness. Built-in type checking and role definitions help prevent errors and ensure consistent interactions. Whether you're using a local LLM or one hosted by a provider like IBM watsonx or Replicate, PDL offers consistent control. But PDL is more than just prompts. It integrates smoothly with external tools and libraries using code blocks, allowing you to incorporate Python code directly into your PDL programs. This flexibility opens doors to powerful applications like retrieval-augmented generation (RAG), where relevant context is retrieved from external sources before querying the LLM. The provided example showcases how PDL, combined with a few lines of Python, can efficiently perform RAG on a code generation task. PDL's declarative approach also extends to agent-based systems. The example in the research paper demonstrates a ReAct agent implemented in PDL, highlighting the language’s ability to handle dynamic, LLM-directed tool calls. This capability allows developers to create more goal-driven and adaptable AI applications. Surprisingly, PDL can even generate PDL! This meta-programming feature empowers LLMs to create their own execution plans, opening up exciting possibilities for automated problem-solving. The research used this feature to analyze the GSMHard dataset of math problems, revealing inconsistencies in the ground truth data, demonstrating the practical value of PDL’s human-readable format. PDL aims to tackle the complexity barrier that often hinders LLM development. By offering a simple yet powerful declarative approach, it empowers developers to focus on the creative aspects of LLM programming, leaving the tedious details to PDL. With its intuitive syntax, robust features, and focus on user control, PDL has the potential to be a game-changer in the world of LLM application development.