Context engineering
The discipline of designing the inputs (prompts, retrieved documents, tool schemas, memory state) that a language model sees at inference time. Replaces the narrower "prompt engineering" framing. In AI-native publishing, context engineering is the day-job skill: what the model produces is mostly a function of what you give it to work with.
In depth
The "prompt engineering" framing treats the input as a short instruction you phrase well. That framing made sense when models had small context windows and limited tool access. It doesn't match how production agents actually work. A modern agent receives a system prompt with rules, a user message, retrieved documents selected by a RAG layer, tool schemas describing what it can call, the running conversation, and sometimes explicit memory state. "Prompt" is too narrow a word for that surface. "Context" is the accurate one.
Context engineering is the practical skill of curating that surface. It's less about clever wording and more about retrieval, schema design, state management, and what gets cut when you're close to the window limit. The empirical finding from operator work is that most "the model got it wrong" failures are actually context failures upstream: the right document wasn't retrieved, the schema didn't constrain the tool call, the system prompt didn't tell the model what good output looks like. Fix the context, the output fixes itself.
The day-to-day craft is closer to information architecture than to copywriting. Operators spend more time on what to retrieve, what to summarise, what to drop, and how to format what stays, than on phrasing the instruction. The leverage is in the curation. A good context engineer can get a mediocre model to ship excellent work, and a bad one will get the strongest model on the market to produce drift.
Examples
- An editorial drafter that fails on tone. The fix isn't "write a better prompt." It's adding three gold-standard past articles into the context so the model can pattern-match on house voice.
- A research agent that hallucinates sources. Fix the tool schema so every citation must include a retrievable URL, add a fetch-and-verify step before the draft, and the hallucination rate drops to near-zero without touching the prompt itself.
- A multi-step agent running close to its context window. The intervention isn't a stronger model, it's a summarisation step that compresses earlier turns into a structured state object, freeing tokens for the live work without losing the thread.
Usage notes
Context engineering isn't a replacement for prompt engineering. Prompts are still part of the context. The point of the broader term is to stop pretending that the prompt is the only lever. When something goes wrong, the first question is "what did the model actually see?", not "how should I reword the instruction?"
Also known as
context engineering
These aliases are what the site's build-time auto-linker matches against to cross-reference this term across the FAQ and machine-readable endpoints.