A content-generation app may need to draft an article, use retrieved facts, and decline a request unrelated to editing. A prompt is the instruction and supporting input given to a language model for a response. Its value depends not just on wording the task, but on supplying relevant information for the job. section 6.1.2 section 7.3.1 section 8.2.3
For a task more complex than a simple zero-shot request, structure the prompt around:
- An instruction that states what the model should do. section 6.1.2
- Context that helps steer the response, such as the conversation so far, a domain role, or retrieved documents. section 6.1.2 section 7.3.1
- Input variables or examples where they make a successful structure reusable across topics and content types. section 6.1.2
In retrieval-augmented generation, retrieved chunks become prompt context. Design that context around user needs and decide whether responses should rely mainly on the retrieved database or also draw on the model’s general knowledge. Multiple chunks make prompts long, so input-token cost and the model’s handling of information in the middle of a long prompt matter. section 7.3.1
Keep a content editor on task
When users ask a content-editing app for weather forecasts or inappropriate content, a system prompt can direct the model to handle only edits and revisions and politely decline unrelated requests. Pair that boundary with off-topic examples and desired refusals in instruction fine-tuning data; the prompt is one safeguard, not the whole safeguard strategy. section 8.2.3
