The Art of AI Context Engineering

Your AI does not have a prompt problem. It has a context problem.Context engineering is what comes after prompt engineering: deciding what an LLM sees, in what order, with what memory, at what cost, and with what defences, so that RAG pipelines, AI agents, and coding agents such as Claude Code hold up in production and not only in a demo. This Second Edition was rewritten for September 2026: 310 pages, 16 chapters, working Python code for every technique, every source dated.Million-token context windows did not fix hallucination, forgotten instructions, or agents that report "done" on work they never finished. They made the problem larger, because there is now more context to get wrong. The systems that work are the ones whose context was engineered. This book teaches how, from first principles to a running system.What you will build and understand- How a model actually reads its context, in six principles you can apply the same day: a prompt sets a scene, rules lose force with distance from the action, a long transcript trains the model on its own behaviour, and reliability comes from an independent check, never from the model's own report.- Retrieval-Augmented Generation that is measured, not assumed: hybrid search, reranking, contextual retrieval, and a complete pipeline you run against a gold set before spending anything on model calls.- Memory, compaction, tool-result clearing, and prompt caching, with the documented prices and the order in which to apply them.- AI agents and the Model Context Protocol (MCP): tools as the contract, loop detection, the false "done", sub-agents, instruction files (CLAUDE.md, AGENTS.md), and when a second agent is worth three to ten times the cost.- Securing the context: indirect prompt injection, the lethal trifecta, the 2025 and 2026 incidents, and the layered defences that held.- Evaluation and records: gold sets, LLM-as-judge and its biases, regression runs on every prompt change, and traces an auditor can read.Why this book and not the othersEvery technique from chapter 4 onward ships as a small Python project that runs offline in a test mode with no API key, so you can read it, change it, and break it before you trust it. Twelve code projects, 144 tests, 18 figures. Every factual claim about a model, tool, paper, or price carries its date, because most of them will have moved by the time you read this, and the book tells you where to check.Who it is forDevelopers, architects, and technical leads building on Claude, GPT, Gemini, or open models who need AI systems that are reliable, auditable, and affordable to run.New in the Second EditionFive new chapters: how a model reads its context, coding agents, multi-agent context, securing the context, and evaluation. Every surviving chapter rebuilt around published evidence from 2023 to 2026.Stop tuning prompts. Start engineering the context. Read the free sample, then get the book.