How to use AI tools effectively - Prompt Engineering basics

Over the past year, my journey through the landscape of AI tools has led to a remarkable observation: no two AI systems are the same, and each one thrives under a specific set of instructions. It's similar to unlocking those special moves in classic arcade games — each AI has its own combination that unleashes its full potential. But beneath these varied interactions lies a shared foundation, which I’ll be unpacking in today’s post on Basics of Prompt Engineering. This piece marks the beginning of a series dedicated to the art of ‘How to use AI tools effectively’. We’ll be using OpenAI’s ChatGPT as our case study, exploring the ins and outs through practical examples. And although our focus today is on ChatGPT, the techniques we’ll discuss are applicable across the board — whether you're tinkering with Claude, Gemini, or Copilot.

When engaging with AI language models, precision is paramount. Picture yourself at a bakery; if you request a pastry but don't specify which kind, you're at the mercy of the baker's choice, which may not satisfy your actual craving. This scenario mirrors the interaction with AI. Unlike a human who would typically probe with questions like 'What type of pastry would you like?' or 'How many would you prefer?', AI lacks the capacity for such intuitive reasoning. These models don't possess the ability to ask follow-up questions to clarify vagueness; instead, they rely on algorithmic logic to make educated guesses. Thus, to extract the most accurate and relevant outcomes from AI, it's crucial to eliminate ambiguity by offering detailed instructions and context right from the start. This approach minimizes the AI's need to assume and maximizes your chances of receiving the precise information or results you seek. In our quest for precision, I’ll use a generic and vague example where I request to write a paragraph about city of Paris, without much context or information.

The AI responded with a generic but accurate description, fulfilling the basic requirement of the prompt. Despite the AI delivering exactly what was requested – a brief on Paris – the simplicity of the ask leads to a correct yet unremarkable result. To avoid this pitfall and achieve a richer, more distinctive output, I often apply this basic formula:

CONTEXT + INSTRUCTIONS

See the enhanced prompt for Paris below, where additional context has been woven in. Notice the difference it makes in the depth and detail of the AI's response.

Let’s address the elephant in the room—why on Earth do I greet and thank AI algorithms as if they're sentient? It might sound like I’m practicing my manners on a glorified calculator, but believe it or not, there’s solid science behind it. Research from Google DeepMind (link to research paper: https://arxiv.org/abs/2309.03409) has shared some interesting findings: it turns out that AIs can actually get better at their jobs when we treat them with a bit of human decency. Who knew? By saying ‘please’ and ‘thank you’, we’re not just being old-fashioned. We’re training these digital beings to handle tasks with a touch more empathy and precision. It’s a bit like teaching a very literal child how to pick up on social cues. So, next time you see me thanking Siri for setting an alarm, know that I’m doing my bit to shape our AI overlords!

As you can see, adding detailed context and instructions can significantly enhance the AI’s responses. For copywriting, my tip is to draft your initial text and then engage with your AI tool for refinement. Although AI can’t perfectly mimic individual writing styles, it is adept at polishing text based on the prompts you provide. I made a note of an excellent example from Bani Kaur, who utilises ChatGPT as an editor to analyse her writing for any logical reasoning gaps, building analogies to find a creative way to express a common idea, and to give her "one word for many words" e.g. instead of using "smile very wide" use "beam". We can build upon Bani’s prompt engineering technique to add these specific set of instructions but also examples. So, the new enhanced formula will go like this:

Context + Instructions + EXAMPLE(S)

Have a look at the example below and see the different result you can get from honing your prompt with additional details we have learned about. Notice, you are no longer getting a ready text but a very precise feedback.

Now moving away from writing, one of my favourite applications of AI tools is their prowess in sifting through and analysing large datasets, a task that can be quite daunting manually. By employing AI, we can swiftly transform raw data into actionable insights. Imagine the AI highlighting trends in sales over time, identifying our best-selling products, and even predicting future demand based on historical patterns. This capability is not just about understanding what happened in the past—it’s about leveraging that knowledge to make informed decisions. Whether it’s optimizing stock levels, tailoring marketing strategies, or setting sales targets, AI helps demystify complex data, enabling businesses to focus on strategic decisions rather than getting bogged down in numbers. Thus, with the right prompts, AI can serve as a crucial tool in not only navigating but also mastering the commercial landscape. Take, for instance, the example of hypothetical AI generated 'Bakery Sales' figures, which you can download here. While using this file, I’d like to introduce new type of of our prompt engineering formula:

Context + Instructions + INPUT DATA

The rollout of ChatGPT’s new feature at the end of 2023 called CustomGPT, was a game-changer, quickly complemented by a dedicated ‘store’ offering a variety of optimized custom AI chatbots. In my own workflow, I regularly leverage these CustomGPTs for various projects, especially for data analysis tasks. To give you a very specific data analysis prompt I often craft, see example below. This prompt get the analysis job done without hitting the 'token limit' – a constraint we're all too familiar with.

Below I attached a brief snippet of the outcome I got from a CustomGPT called ‘Diagram & Data: Research, Analyze, Visualize’, developed by Max & Kirill Dubovitsky.

I encourage you to experiment with the sample 'Bakery Sales' file provided,

Bakery_Sales_Data.xlsx

Download XLSX • 13KB

or feel free to use your own data—just remember to avoid any sensitive or private information! Apply one of the formulas mentioned in today’s post and observe the insights that emerge. Share your findings and any 'eureka' moments in the comments below, or let me know if you’re still exploring the nuances of prompt engineering. For those keen to delve further, there are numerous online resources with premade prompts ready for use. For example, try prompts that instruct your AI to 'Act as if…' it was a personal trainer, translator, SEO expert, and more. Check out this collection of prompts for more ideas.

This brings us to the end of the first part of our series on using AI tools effectively. In the upcoming posts, we'll explore fun and creative aspects of AI, focusing on text-to-image and text-to-audio prompting, and delve deeper into advanced prompt engineering techniques. Let’s continue to demystify the art of communicating with AI together.

Until next time,

Dorian