Prompt engineering refers to methods for writing and organizing the instructions given to a language model for optimal outcomes. As the term implies, its primary focus is how to write effective prompts, and particularly system prompts. Anthropic describes it as having been the biggest component of AI engineering work in the early days of building with LLMs, when the majority of use cases outside everyday chat interactions required prompts optimised for one-shot classification or text generation, and as having been the focus of attention in applied AI for a few years before Context engineering came to prominence.
Usage
Within a context-engineering frame, prompt writing becomes one component of managing an entire context state that also includes tools, external data and message history, and Anthropic’s prompt-level guidance carries over into that larger practice: clear and direct language at the right altitude, distinct sections marked with XML tagging or Markdown headers, and the minimal set of information that still fully outlines the expected behaviour — where minimal does not necessarily mean short.
When It Applies
- Applies wherever the instructions given to a model are the main lever on its output. Anthropic identifies one-shot classification and text generation as the cases where this was the majority of the engineering work.
- Assumes the relevant state can be settled in advance, in writing. Anthropic argues this assumption breaks down for agents that operate over multiple turns of inference and longer time horizons, where information accumulates during the run and must be cyclically refined.
- Insufficient rather than wrong in the agentic setting: Anthropic presents context engineering as the natural progression of prompt engineering and as encompassing it, not as replacing it, and notes that the exact formatting of prompts is likely becoming less important as models become more capable.
- A well-established practice, described by Anthropic as the prior focus of applied AI work rather than as anything it proposes here.