Context Engineering FAQ

Context engineering is the systematic design of the information an LLM receives during inference: task, instructions, retrieved evidence, memory, tools, examples, constraints, and output requirements.

Prompt engineering focuses on wording instructions. Context engineering includes the whole information payload: what is retrieved, selected, compressed, ordered, remembered, cited, and evaluated.

Include answer-critical task details, priority instructions, authoritative evidence, necessary examples, durable memory, and the output contract. Leave raw history and low-signal source dumps outside the active window.

Yes. RAG is still useful for large, changing, permissioned, or source-sensitive corpora. Long context helps reason over a larger working set, but it does not replace selection, provenance, or access control.

Context rot is degraded or inconsistent model performance as input length grows. It can come from distractors, stale history, buried evidence, repeated instructions, conflicting sources, or sheer length.

It can. Safe compression preserves answer-critical facts, qualifiers, and provenance. It should be tested against original context on cases where dates, exceptions, and conflicts matter.

Evaluate retrieval, assembly, and generation separately. Check whether the right evidence was found, whether it was placed and labeled well, and whether the answer was faithful to that evidence.

No. Curated Context is model-neutral. It complements claudecontext.com, which is Claude-specific, but the guidance here is about context curation across LLM systems.