The Real Value of Prompt Engineering: How to Leverage LLMs
I interview industry experts on prompt engineering and provide you with their perspectives in pursuit of a signal within the noise.
In November 2022, a research lab and startup named OpenAI launched a language-model-based product called ChatGPT. Spreading like wildfire, ChatGPT reached 100 million users just 2 months after launch, and in 2024 received around 10 million queries every day, each with a unique and specific intent. Navigating enormous information landscapes has never been easier, and for many, ChatGPT or similar LLM-based products have become staples of daily life.
With the increased popularity of the LLM platform, industries are beginning to look towards leveraging its capabilities to their advantage. Prompt Engineering is the practice of designing inputs for AI tools that will produce optimal outputs. However, in an industry full of innovation, it can be hard to see past potential into reality, and thus the task becomes to separate signal from noise. In other words, I want to know how to leverage the rising LLM platform to our advantage.
In this post, I’ll take you through the realities of prompt engineering, and interview some friends that I would consider experts in the field to provide you with actionable insights.
Landscapes of meaning…
To understand the rise of prompt engineering, it’s important to thoroughly understand how these models work, at a high level at least. The most important thing to remember about modern AI models is this:
A transformer model finds the next best token in a sequence by carefully navigating a high-dimensional space of data embedded with meaning
I find it helpful to imagine this as walking through a field of random objects, which are sorted by their relevance to one another. Suppose you’re looking for a cheeseburger, and you notice there is a napkin near you. Looking around the napkin, you notice a ketchup bottle a little further away. Walking towards the ketchup, you notice a hot dog, and soon a spatula as well. Eventually, you find your cheeseburger, surrounded by objects with some sort of relation to it. The more related the object, the closer it is to the cheeseburger.
When you give a model a prompt, it walks through this field, although its “field” is actually a space with thousands of dimensions, making it effectively incomprehensible to us humans. However, the same concept effectively applies, for each token in the prompt to a transformer, it will peruse its field of similarly ordered objects.
For example, when you give a sequence of words to a language model, it traverses its semantic space and attempts to make a prediction as to what the next most probable word in the sequence might be. That’s where the different responses from language models come from, and thinking about them this way allows you to optimize your prompts to ensure effectiveness. Think of prompting as essentially guiding the model through this semantic space, and coercing it in different directions.
So where do these vast semantic spaces come from? Did you know that when you tell a language model “let’s take a deep breath”, it performs better on certain tasks? This is because these models are trained on massive corpuses of text, and the relationship of different pieces of data to one another. So, during training a positive correlation between this particular phrase and positive outcomes is drawn, and thus when you use this phrase in a prompt, it guides the model in a certain direction.
The semantic spaces I mentioned before are built during training on these massive corpuses of human-generated data, and so we find these quirks. Even though a language model has no lungs, or ability to take deep breaths in any way, it can still perform better simply because we are guiding it towards a certain area of its training data. This, in my opinion, is one of the fundamental aspects of prompt engineering as a whole.
Breaking down prompt engineering…
So now that we’re equipped with some prerequisite knowledge, we can dive into its applications. Since transformers have become more popular, learning how to work with them has become the obvious course of action. With such an influential technology, knowing how to best leverage these models will be a highly valuable skill.
However, the nature of this work is where things get a bit murky, and with new technologies–especially those with marketing teams–hype is prevalent. That’s why I’ve turned to a couple of industry experts with comprehensive background working with and studying AI, to gain a more nuanced perspective.
Introducing our experts…
Evin Tunador is a content creator, consultant, and independent researcher in the field of Artificial Intelligence. He has written premium courses on prompt engineering with Learn Prompting and runs an excellent AI-research-focused YouTube channel. Evin also has a Substack and article on self-teaching artificial intelligence.
Mike Bybee is an expert software developer and researcher in AI, known as 'stereoplegic' across AI Discord channels. He offers a broad and practical perspective on prompt engineering and the industry as a whole. Mike can also be found on Hugging Face and has curated some excellent collections of research papers.
So, what even is prompt engineering?
Simply searching the term “prompt engineering” reveals that it has gained widespread reach, with most references hovering around the following definition:
The practice of designing inputs for AI tools that will produce optimal outputs
However, given the nature of buzzwords and the implications they bring with them, I asked our experts what they thought when they heard the term “prompt engineering” itself.
“To be honest, it depends on the source. If it’s someone serious like Greg Kamradt or Sam Witteveen—individuals who regularly demonstrate that they are more than just prompt engineers—then I take them seriously. I think at best it's an incredibly useful skill… At worst, it's a way to lazily seek clout you haven't earned.”
- Mike Bybee“I don't know who invented this term, but I think it was likely someone who also had an interest in putting that skill on their resume and marketing themselves to prospective employers who don't know anything about AI. The word 'engineering' gives a false sense of science that is just not there… I'm not sure what word would be better, but awhile ago Sander Schulhoff and I were throwing around words like 'design', 'curation', or 'experimentation'. Something that compares it to a learned art form or skill would be better.”
- Evin Tunador
Both of our experts expressed criticism towards the potentially hyperbolic nature of the term. While prompt engineering does have legitimate meaning, for many individuals in the industry with more insight, it seems to leave a bad taste in the mouth.
What we can distill, however, is that prompt engineering looks more like an art form, with careful and iterative improvement and experimentation being key. Experts in the field are interested in how to get the best out of models, and are researching and developing different strategies for interaction. While there’s certainly skepticism around the term, and charlatans will likely take advantage of its flashy tone, there is real value here.
What does it look like?
While it’s simple to say that prompt engineering, done well, increases the efficiency of language models for specific tasks… what does that look like in practice? On top of that, where does this practice actually create value, and how does this change across time? These questions are critical for understanding what to actually do with language models, and where to expend your effort.
With that, I decided to ask Evin and Mike what prompt engineering looks like today, as well as in the future.
What’s the core value…
To start, I wanted to know what principles lay at the heart of prompt engineering, what do the best prompters get right?
“First, as these models get better we're actually seeing far less of a need for complicated, unintuitive prompts, and I expect that trend to continue for most use-cases… I see the future of prompt engineering becoming more accessible as we stop using weird tricks found in research papers and start better utilizing basic communication skills… such as concise writing, clearly defining the problem at hand, and explicitly stating expectations.”
- Evin Tunador
Evin brings up an excellent point. AI companies aren’t looking to build products that only work under an expert guided scalpel, but they’d like to appeal to the general population. While these highly research-based and carefully-crafted tools have historically been most useful to qualified experts, these days almost anyone can have a productive interaction with language models.
This doesn’t mean that there isn’t still research required in this field, but we can expect communication with language models to feel much more human as time goes on. Mike shared a similar sentiment, but was sure to emphasize that models still are fundamentally mechanistic.
“You know how I feel about anthropomorphizing AI models–they're still just software. The fact that certain prompts or phrasings elicit better results from LLMs–and diffusion models, etc.–doesn't change this principle; in fact, I'd say it emphasizes it.”
- Mike Bybee
He goes on to describe the nature of software and how even while natural language is more complex than other software mediums, the same basic rules apply.
“Software output is only as good as its inputs enable it to be… I think the good prompt engineers understand this, even if only implicitly, and thus their skill is in optimizing the input to maximize the quality of the outputs they receive. Natural language is far more nuanced, but essentially no different in this sense than any other input dialect, such as JSON, XML, etc.”
- Mike Bybee
So software is software, even if it’s a gigantic collection of billions of weights and biases, it still follows the same rules of input and output. It’s important to recognize that it’s not human, although AI companies will continue to make it feel more human to work well with a general audience. Language models specifically will continue to feel more human, and understand more of the complexities of human language, but will still be machines.
Evin describes the nature of interacting with these machines, and how they might benefit our work and society as a whole.
“I predict the majority of the value will come as individual people get better at using these machines like highly competent personal assistants or even coworkers. As a bonus, those increased communication skills might even have spillover effects into human-to-human interaction.”
- Evin Tunador
It’s nice to conceive of a future with better human-to-human communication, where people are much more capable of expressing themselves. It sounds like learning these communication skills will be crucial for leveraging transformer based models in the future.
Prompting, in practice…
However, it’s unlikely that companies will be willing to pay someone simply for their good communication. So what practical skills can prospective prompters pursue in order to gain an edge with coming industry trends? Mike and Evin had some excellent insights into this question.
“You're still going to be expected to create pipelines of some sort, and probably some sort of ETL [aka Extract, Transform, and Load] data work. You'll have to know Python, JavaScript, etc., just in a different sense than a front end or back end dev. I'd say it still comes down to various frameworks/libs: Langchain, LlamaIndex, DSPy, etc. i.e. tangible things you can list on a resume.”
- Mike Bybee
So from a practical standpoint, unfortunately to really find a job as a “prompt engineer”, you’ll still need to be in the wider field of software engineering. However, it’s not as simple as just knowing how to code. When asked what kinds of skills and experience a prompt engineer might look into, Evin emphasizes the importance of creative degrees and writing skills.
“An english or creative writing degree, or some other field like law that teaches you how to write, would be the best thing in the long term. Time and time again I've seen competent writers get far better outputs from these models without even ever hearing of the term "prompt engineering" before they begin. Second to that would be a low-level comprehension of the math behind these models, which is a lot to ask but it does give you a huge leg up in terms of intuitively understanding their limitations and unusual quirks.”
- Evin Tunador
So it would seem that prompt engineering is like any other subfield of software engineering, and requires a specific combination of skills, both technical and otherwise. There are no easy roads to true success, and the same applies with this up-and-coming professional field. While the industry will certainly favor those who are capable of working well with language models, this likely doesn’t mean simply talking to ChatGPT all day.
Transformer models, as well as other state of the art architectures, are largely considered to be in their early stages. Perhaps infancy is selling them short, however there is certainly a massive road of development ahead for researchers. As we see these models advance, working with them and integrating them into existing applications and tools will become the next wave of engineering. The real value of this technology will be found in this practice of integration.
Navigating the changing landscape…
In June 2007, Apple released the iPhone. It wasn’t the first smartphone, it wasn’t the first phone with a camera or even internet capabilities. In many ways, the iPhone wasn’t special, but today the iPhone alone holds a 30% global market share of smartphone sales, and a 60% market share in the United States. So what set it apart?
With its multitouch touchscreen, sleek design, intuitive UI, and eventual App Store, the iPhone got one thing right that the rest of the competition didn’t–it could be popular. Apple didn’t need to have any grand innovative ideas in order to beat out their competition, all they had to do was make a device which was accessible and appealing to the average person. By making accessibility a priority, the iPhone lead the charge, making smartphones the massively popular platform of today.
In November 2022, I believe OpenAI did the same thing with ChatGPT. Without taking the parallels too far, it’s clear that we’re on the verge of a new exploding platform. Companies across the US are in a race to build AI related products, and even my mom is telling me about how she uses ChatGPT and Perplexity at her day job. There’s no question that AI has exploded in popularity as of late, but the exact implications remain to be seen.
However, we can assume that in the same way that the App Store catalyzed a billion dollar market, the rise in the popularity of Large Language Models and related technologies will as well. Prompt engineering, while not without its hype and charlatans, is an inevitable high-value skill of tomorrow.
The future likes to arrive unexpectedly, and so we can’t say anything with certainty. However, that won’t stop me from trying. Current hype cycles will fade, but there is real substance to these technologies, and they will be put to use across the world. It is my hope that we can learn to understand them, and maybe they can help us learn to understand ourselves.
Author’s Note
Thank you so much for reading! This was a blast to research and write and I’m excited to continue producing this kind of content!
Huge shoutout to Evin and Mike for being willing to help me with this, I’d recommend that you check them out on social media and across the internet.
Here are some links to help you find Evin:
LinkTree: https://linktr.ee/tunadorable
YouTube: https://www.youtube.com/@Tunadorable
Substack:
Here are some links to help you find Mike:
X (Twitter): https://x.com/SciArtMagic
HuggingFace: https://huggingface.co/stereoplegic