← All courses
Intermediate · 7 lessons
Context engineering: beyond the prompt
Why prompts stopped being enough: the context window as a budget, retrieval and memory, compression, repository maps for coding agents, and context-quality evals in production.
- 01Why prompts stopped being enoughThe anatomy of agent context and the context rot phenomenon: why optimizing a single string no longer works and what context engineering is.40 min
- 02The context window as a budgetTokens, input pricing, KV-cache, and lost-in-the-middle: learn to allocate the window across budget lines and protect a stable prefix.50 min
- 03Retrieval, memory, and stateThree external sources of context: hybrid search and contextual retrieval, memory notes beyond the window, and structured task state.50 min
- 04Compression and summarizationWhen and how to compress history: schema-based compaction, anchor facts, tool-result cleanup, and what must never be compressed.45 min
- 05Context for coding agentsA repository map instead of the whole repository, CLAUDE.md patterns, task context from diffs and tests, sub-agents as context isolation.50 min
- 06Multi-turn context in productionConversation history as a liability: cache-friendly request layout, observability of context composition, overflow, and poisoned context.50 min
- 07Measuring context qualityRetrieval metrics, faithfulness, and end-to-end evals: how to prove a context change improved the agent instead of breaking it.45 min