The Spice is in the Prompt

The Medium article Let me tell you what I (really, really) want — or how to use an LLM to get information out of structured data sets by Eliot Salant offers an understanding on how to get meaningful answers from large language models (LLMs). Rather than relying on vague or overly polite prompts, we have the power of clarity, specificity, and iteration. If we want good information, we need to ask good questions—and that means knowing what we really want. In the article we are walked through a journey of refining prompts, showing how even small tweaks can dramatically improve the relevance and usefulness of AI-generated responses.

One of the most compelling insights mentioned in the article, is a technique where queries are built step by step, layering context and constraints to guide the model toward more accurate outputs. The importance of giving the LLM permission to say “I don’t know,” it's emphasised. This helps reduce hallucinations and encourages more grounded responses. This approach mirrors human conversations, where clarification and feedback loops are essential to understanding. By treating the LLM as a collaborative partner rather than a magic oracle, Salant demonstrates how users can extract reliable information.

The article is a call to action for anyone working with AI: be intentional, be iterative, and don’t be afraid to experiment. It’s not just about getting answers, it’s about asking better questions.