Digital Lexicon Hub

The practice of designing and testing the instructions, context and examples given to a language model so its outputs meet defined criteria.

Well-tested prompts often improve the accuracy, format compliance and cost of an LLM feature faster and more cheaply than fine-tuning or switching models, and they make AI behavior easier to review and audit.

A support team writes a system prompt that defines tone, adds 5 example tickets with ideal replies, requests JSON output, and scores each prompt version against 150 past tickets before replacing the version running in production.

Prompts are sensitive to model versions, so behavior can shift after an update unless the model snapshot is pinned and tests are rerun. Prompting cannot supply missing knowledge, and it does not by itself stop Prompt Injection.

Common techniques include clear and specific instructions, system or developer messages that take priority over user input, few-shot examples, structured delimiters such as XML tags, room for step-by-step reasoning and chaining a task across several prompts. Vendor guidance recommends defining success criteria and building evaluations before iterating. In Retrieval-Augmented Generation (RAG), the prompt also carries retrieved context to the Large Language Model (LLM).