Skip to content

What is prompt engineering? A Q&A with one of our Prompt Engineers

What is prompt engineering? How we coach AI

At DeepDive, prompt engineering is fundamental to transforming raw web content into trustworthy intelligence reports. Our marketing team sat down with Finola, Prompt Engineer at DeepDive, to explore this critical discipline.

Hi Finola, so please tell us, what is prompt engineering?

At a very basic level, prompt engineering is making input tasks for Generative AI. We ask AI to collate, analyse or to write and we do that by prompting in natural language to ask AI to perform a specific output.

People might not realise it, but in fact we’re all prompt engineers.  Every time we talk to Siri or Alexa, that’s us prompting to get a response, although that’s a very basic example. I like to think of prompt engineering as writing incredibly precise instructions for a brilliant but literal-minded research assistant. Large Language Models (LLMs) are remarkably capable, but they need clear guidance to produce reliable outputs.

We can't afford ambiguity or inconsistent results when our AI extracts information from a court document or a company report, it needs to understand exactly what we're looking for, how to present it, and what evidential standards to apply. That's where prompt engineering becomes critical, it's the essential bridge between human and computer.

How did you get into prompt engineering?

I first started using LLMs to help me revise for my first degree at Durham University. This was actually the time when ChatGPT was first released, around 2022, and then I did my Masters in Computer Science at the University of Bristol. For my Masters project, we built a theoretical cancer risk prediction tool, and depending on what your risk factor was, we used LLMs to generate meal and exercise plans.

To be honest that was kindergarten stuff compared to how I prompt now. But what I did learn then was that that LLM's can be unreliable and I learnt the importance of prompt reiteration. The art of prompt engineering is an iterative process of refining your input message to the AI to get the best and most reliable outcome.

Did you ever think that you would be employed as a Prompt Engineer?

Not at all, growing up, my parents would always say, “Oh, you'll have a job which doesn't even exist today.” I mean, fair play to them, right?

What I’m doing right now is the sort of fusion skills that everybody needs to learn to stay relevant in the AI economy. Thinking with AI, using your judgement to decide when to step in and redirect the AI to improve efficiency of outputs. I like to think of the AI as my apprentices and my co-creators.

How do you structure effective prompts for DeepDive investigations?

Every effective prompt follows a similar architecture. First, we establish the context - telling the AI it's operating as an experienced compliance analyst or investigator. This sets the language, tone and standards of the output. We include examples and counter-examples. We show the AI what good extractions look like and explicitly demonstrate what to avoid, like inferring connections that aren't stated, or ignoring the descriptive language used in sources.

Next comes the specific task such as identifying business relationships, assessing source credibility or scanning for links to sanctioned entities. We're extremely explicit about what constitutes a valid extraction versus speculation.

Then we define the output format. Should statements include qualifying language like "allegedly" or "according to sources"? How should confidence levels be expressed? How do we get the LLM to give correctly formatted citations? How should values such as dates or currency be expressed?

Do you have to be an expert in other programming tools to be a good prompt engineer?

No, but there’s an important point to be made there actually. DeepDive isn’t just a wrapper around an LLM.  It’s an orchestration of several technologies including web search and entity resolution.  The value we derive from prompt engineering doesn’t come solely from the quality of the prompts themselves. It also depends heavily on the quality and breadth of the data we use alongside our prompts.

This is where a concept called Retrieval Augmented Generation, or RAG for short, comes in. Standard language models are trained on datasets up to a certain cutoff date and can’t access new information beyond that date.  So, if a model was trained before a certain date or simply lacks a  specific fact, it wouldn’t be in possession of all the information and could hallucinate or produce outdated info.

Our extensive web search and entity resolution allows us to use RAG to dynamically inject a rich,  targeted Body of Knowledge into the generation process, making our DeepDive reports more accurate and contextually relevant.

Finola, thanks for taking the time out from instructing your bot army. Before we finish, can you tell us about your AI superpowers that you use in your daily life?

Ha, I think my generation is now using AI chat tools more than Google! I’m hooked on NotebookLM right now. It’s an AI research assistant that you feed with context and sources and then it creates a conversational podcast based on all of the information you've submitted. It's a very, like, off the cuff kind of conversation. It's not very automated, as you might expect from a computer just reading off an article. It's properly articulated, they even make jokes, and it's a great way to learn and consume content instead of reading.

Thanks Finola, I’m getting onto NotebookLM this afternoon!  Will you and your bot army please come back soon to talk about further topics such as hallucinations, explainable AI and tokens?

Yes please, we would love to!

Want to read more? How prompt engineering helps DeepDive build a Body of Knowledge