Prompt engineering is the practice of designing and refining the instructions, examples, and context sent to a language model. A prompt may be a single command or a structured set of directions. Clear, task-specific inputs help the model produce more useful and accurate responses without changing its training. section 1.4.1 section 2.0
Build a prompt around the job it needs to do:
- State the task clearly. A chatbot can use role-based messages to establish consistent behavior across a conversation. section 1.1.2
- Add relevant context or a few examples when the model needs to follow a pattern. Recent conversation turns can help maintain continuity; examples enable in-context learning without fine-tuning. section 1.4.1
- Make repeated prompts reusable with a fixed instruction and variable fields, then test and refine the outputs. section 1.4.1 section 2.0
Research summarization
A research engine needs different prompts for different stages: turn a broad question into targeted web searches, extract key facts and arguments from retrieved pages, and combine the summaries into a report with citations. Separating these tasks makes each prompt’s purpose clear. section 4.5 section 4.8
Prompts also define an agent’s scope, available tools, and routing criteria. A router should account for ambiguous requests; a supervisor needs descriptions of specialist capabilities. When factual information is missing, clearer wording alone is not enough: retrieval can supply external context, and a grounded prompt should instruct the model to rely on it. section 12.4 section 1.4.2
