Pick the activity that belongs to context engineering rather than prompt engineering

The four options look similar because they all change what the model sees. The difference is *where* the change happens. Options 0, 2, and 3 all edit the literal text of the prompt itself. Rewording a system prompt to use bullets, adding more few-shot examples, and changing the assistant's role description are all prompt-engineering moves. They tune the instruction string. Option 1 is different. It does not change a single word of the instruction. It changes what *else* gets put into the context window on this turn: which chunks the retriever pulled, which memory facts the memory store returned, which tool outputs from the prior step are still relevant. That is context engineering. In 2026 practice the distinction matters because most production LLM bugs come from the surrounding context, not the instruction. A perfect system prompt and the wrong three retrieved chunks still gives you a wrong answer. A mediocre system prompt and the right three chunks still gives you a usable one.

**Context engineering** is the system-design discipline of composing the full input, retrieval, memory, tool outputs, history, schema, that arrives at the model on a given turn. **Prompt engineering** is the narrower craft of writing the instruction string itself. Options 0, 2, and 3 are all prompt-engineering moves. They edit characters inside the system or user prompt: formatting, examples, role description. The instruction changes; the surrounding context stays the same. Option 1 is the context-engineering move. The instruction string can stay byte-identical from turn to turn. What changes is the set of retrieved chunks, the memory facts pulled from the user's profile, the tool outputs from the prior reasoning step. The decision of which chunks to include, in what order, with what budget, is exactly the context-composition problem. A useful test is the question 'would this change still apply if the instruction were locked?' Reformatting bullets, adding examples, and changing the role description all touch the instruction itself. Picking the top three retrieved chunks instead of the top fifteen is independent of instruction wording, that is the giveaway. In a 2026 production app the prompt is typically a small fraction of the final input; the bulk is composed at runtime from databases, vector stores, tool calls, and summaries.