Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
by Jason Wei et al.
Shows that prompting a large model with a few worked examples containing intermediate reasoning steps substantially improves arithmetic, commonsense, and symbolic reasoning, and that the effect emerges only at sufficient model scale. The origin of chain-of-thought prompting.
More resources on Prompt Engineering
Prompt Engineering Guide
Reference covering prompting techniques with worked examples: zero-shot and few-shot, chain-of-thought, self-consistency, ReAct, and tree of thoughts, plus model-specific notes and adversarial prompting risks. Readers finish able to choose techniques deliberately.
Prompt Engineering for LLMs: The Art and Science of Building Large Language Model-Based Applications
Two engineers who built GitHub Copilot explain why prompts work in terms of how models complete text, then build up few-shot patterns, chain-of-thought, retrieval-assembled context and evaluation for production applications rather than one-off chat sessions.
Effective Context Engineering for AI Agents
Anthropic's applied AI team reframes prompting as one part of curating the whole context window. Covers context rot, calibrating system prompt specificity, minimal non-overlapping tool sets, just-in-time retrieval, and compaction, note-taking and sub-agents for long-horizon tasks. This is the resource that answers the topic's hardest question - where prompting stops and retrieval, memory or architecture must take over. It gives concrete mechanisms (compaction, structured note-taking, sub-agent isolation) rather than the vague 'context is the new prompt' takes that flooded blogs afterwards.
GPT-5.2 Prompting Guide (OpenAI Cookbook)
OpenAI's own guide to steering its current flagship reasoning model: controlling verbosity, preventing scope drift, reasoning-effort settings, long-context handling, tool-call parallelism, schema-driven extraction from documents, and a migration table from earlier models. Reasoning models broke a lot of received prompting wisdom, and this is the clearest first-party account of what changed.
Anthropic's Interactive Prompt Engineering Tutorial
Nine notebook chapters with graded exercises: prompt structure, being direct, roles, separating data from instructions, output formatting, step-by-step reasoning, examples, hallucination control, plus chaining and tool use. You write and debug real prompts, not read about them. The single best free way to actually practise rather than read - every chapter ends in an exercise that is graded against expected output, which is the only mechanism on this list that catches the gap between understanding a technique and being able to apply it. Its example code targets an older Claude model, but the failure modes it drills are model-independent.
Prompting Best Practices (Claude Platform Docs)
Anthropic's continuously updated prompting reference, organized as per-model behavioural guidance, then techniques that apply to all current models - clarity, examples, XML structuring, extended thinking, tool use, agentic loops - then migration notes for older prompts.