LLM Prompt Patterns

Prompt patterns are reusable techniques for steering large language models toward useful, reliable output. They range from simple instructions and few-shot examples to multi-step reasoning scaffolds, role framing, self-consistency checks, and formal output specifications that a model must follow exactly.

Why it matters now

The craft is changing quickly. Early prompt engineering focused on phrasing: finding the right words to coax better answers. Practitioners are now paying more attention to structure. They define schemas for responses, use constrained decoding so outputs are guaranteed to parse, and design prompts that fit cleanly into software pipelines.

Reasoning-oriented techniques such as chain-of-thought are also evolving. Researchers are experimenting with variations that add iteration, branching, or aggregation across multiple attempts. As newer models handle more reasoning internally, some older tricks are being reassessed. Meanwhile, reliability and machine-readable output matter more than ever for production systems.

What you'll find here

This hub tracks developments including:

• Reasoning frameworks: new and refined approaches to step-by-step reasoning, and debate over when they still help.
• Structured outputs: schemas, JSON modes, grammar-based and constrained decoding, and the tooling around them.
• Practical guidance: evolving best practices for writing prompts that hold up across models and versions.
• Evaluation and testing: how models, including open-weight systems, are assessed for behavior, robustness, and safety.
• Research and tooling: papers, libraries, and techniques that change how developers interact with LLMs.

Whether you build with language models or want to understand how they're being directed, this is where to follow the shift from clever wording to engineered interaction.