Context Engineering

Context engineering selects, organizes, and maintains information for model calls: instructions, conversation state, retrieved evidence, tool descriptions, and results. Anthropic's engineering article focuses on choosing useful information within limited context.

Related concepts

Prompt engineering expresses the task. RAG retrieves external evidence. Agent memory retains state across steps or sessions. Context engineering decides what the current call should receive and how that information changes.

More text is not enough

Lost in the Middle found position-sensitive performance in tested long-context models. A larger window does not automatically make evidence useful.

Engineering recommendations

For document review, maintain the objective, confirmed facts with sources, unresolved questions, and material needed next. Carry this state forward and retrieve originals as needed.

When removing repeated tool output, preserve its source and retrieval path. When summarizing history, retain constraints, pending work, and decisions. Keep assumptions distinct from established facts. Label external documents as data.

Evaluate multi-turn tasks for lost constraints, stale evidence, traceable citations, completion rate, token use, and latency. Test before and after compaction rather than judging only summary fluency.

See context windows, RAG, agent memory, and prompt injection.