A Pragmatic Guide to GenAI
It seems as if GenAI is everywhere these days…
… and not necessarily in a bad way… but it’s become so very prevalent… every Google search comes with an AI overview, and Copilot has been sewn into every Microsoft product. It feels inescapable in the news too — Meta has been spending beaucoup bucks to attract AI talent and OpenAI’s bumpy rollout of ChatGPT 5 means that the “king of AI” has yet to be crowned.
With all this buzz, investment, and disruption, it’s quite clear that
GenAI is integrating into our lives (whether you want it or not)
and reshaping them for the better (hopefully, at least partly).
LLMs like ChatGPT, Gemini, Claude, DeepSeek, and others can aid us in basic tasks like drafting email, and summarizing documents. But applications of AI have propelled multiple breakthroughs across many scientific fields. AI has given us ways toforecast adverse weather more accurately andidentify tumors in patients more reliably. It’s even changed how researchers approachdrug discovery and development andcrop protection. It’s impacting many different industries, and seems to integrate into new ones every day.
It’s quite clear, things are much different today than what they were like only a few months ago…
… and (un)fortunately things will continue to change as time goes on.
With all of these changes, it’s easy to get wrapped into the excitement (i.e., fear) — we have to realize that we are living through history right now. These systems are amazing breakthroughs, yet their applications and impacts are still broadly unknown and underexplored.
Even in the face of all of this uncertainty, many industries have been going “all-in” on AI, but they are coming with mixed results. Some companies are rethinking their AI strategy as they face challenges amid their rollout (e.g., Taco Bell and the 18,000 water cups). In addition, end-users are frequently finding hallucinations which undermine GenAI’s usefulness. And many are raising concerns over AI’s environmental, ethical, and societal impacts. Even economists are getting nervous as fears are growing that we are riding an AI bubble.
We are in (yet again) unprecedented times 😪
So it’s quite apparent that the current state of GenAI isn’t perfect, but it has the potential to do so much good. It can empower visionaries to create prototypes themselves with far less effort, and it can offer humans a helping hand as we need to process more and more data. I truly believe GenAI can make one significantly more productive and can help them get ahead in life… if it is used wisely.
My hope is to pass some wisdom along with this blog post, and hopefully you, the reader, can make pragmatic choices when using GenAI.
Cheers,
Nate
P.S., out of personal pride, none of this was written by GenAI (aside from spellcheck).
What exactly is GenAI?
🤔
What exactly is GenAI? 🤔
At a very high level, Generative AI (GenAI) is a type of artificial intelligence which can create “new” and “original” content based on an input prompt. This form of AI is able to learn patterns from large amounts of training data and use these patterns to generate new outputs based on a user’s prompt.
To get more detailed, GenAI uses a particular technology at its core: Generative Pre-Trained Transformers (GPTs), often simply called “transformer models”.
Transformer models are built using vast amounts of training data — basically as much as companies can get their hands on. Companies like will spend billions of dollars in compute cost to tokenize data (i.e., making text interpretable by a machine) and use it to train their transformer model (the GPT). This GPT model will then be fine-tuned towards a particular modality — like chat, or image generation. If you’re interested in more of the nitty gritty of how GPTs are training and fine-tuned, I’d encourage these articles from Sapien, from Analytics Vidhya, and from Airbyte.
But the idea is that these models have a broad and approximate knowledge about the tokenized data that it’s been trained on. The hope is the proximity of these tokens in hyper dimensional space correlates with true knowledge.
This technology enables us to ask plain text questions to an LLM which has a generally good understanding of how language works. You’re also able to provide additional context to the models to improve its predictive power in a particular space. The combination of language, context, and human feedback enable users to work so much faster than they were able to previously — we now have a new tool. And as David Thomas and Andrew Hunt said in the Pragmatic Programmer:
“Like the craftsman, expect to add to your toolbox regularly. Always be on the lookout for better ways of doing things.”
GenAI is powerful tool and it is ushering in a new era of working. However, we mustn’t be careless in how we use and apply this tool.
We must fully understand its benefits and risks before we dive in.
Benefits of GenAI
(1) Automation of repetitive tasks to enable deeper thinking
In my humble opinion, GenAI’s greatest asset is automation. I personally use Google’s Gemini to construct personalized drafts of emails, to perform initial searches of research topics, convert handwritten notes into documents, and to curate the latest weekly news for biotech — it can effortlessly traverse large amounts of data quickly to help me come to conclusions faster.
So now, rather than spending time copying notes, drafting emails, and searching the web; I am able to take on more strategic thinking.
(2) The Great “Rubber Duck” for brainstorming
In software engineering, developers have often used a yellow rubber duck as a sounding board when debugging. The process of explaining it aloud to the duck helps the programmer get to the bottom of the error quickly — I think GenAI can be useful for this as well.
When one is brainstorming or debugging, they are often a “subject matter expert” of that topic. Having a “mirror” with approximate knowledge of everything can be useful assuming you can call B.S. on the model.
(3) AI Applications have the potential to do large amounts of good
There have been numerous applications of AI, well before the notoriety of ChatGPT. We have been using AI and GenAI for many medical applications such as drug discovery and development, medical diagnostics, and personalized treatment plans. It also has been used in environmental projects to identify next generation pesticides, optimize energy consumption, and to model impacts of climate change. Societally, it’s enhancing accessibility for disabled persons through real-time captioning, live translations, visual assistance, and advanced prosthetics. Achievements in these fields are measured in years, not weeks so the jury is still out on the quantifiable impacts of AI in these fields, but initial efforts have been promising.
(4) Economic disruption
From my experience, GenAI has enabled my business to hit the ground running much faster than I originally anticipated. Through back-and-forth conversations with GenAI, I was able to identify the initial steps I needed to take to start my business, and verify these steps through conversations with experts. It helped me do initial research on potential clients and draft personalized emails — I’d have to double check them and rephrase the drafts, but it would give reasonable starting points.
I also find myself learning new tools and technologies faster. I am able to make prototypes of new software ideas so I can learn new concepts, and then recreate them myself to solidify the concept.
If you know what you’re doing, GenAI can enable substantial productivity growth and help you get an idea to the market faster than ever. You’re able to do more with fewer resources.
“With great power comes great responsibility”
One could argue that Voltaire or Winston Churchill was the real originator of this quote, but regardless, Spiderman’s words ring true today.
Risks of GenAI
1) Blind trust, misinformation, and AI hallucinations
In my opinion, the greatest risk associated with GenAI is not from the AI itself, but from users who are not critical of its output.
Since its inception, AI responses have been riddled with hallucinations making them hard to trust. This ought to be known by now; yet, there are numerous cases of users blindly trusting LLM responses and facing serious (and sometimes deadly) effects.
In today’s day-and-age, we must be critical of everything we see on the internet, especially if we suspect it may come from GenAI.
(2) Bias and Discrimination
GenAI models are trained on vast datasets that represent much of the internet. However, if these datasets contain any biases, the models can amplify them resulting in discrimination against minorities. And being frank, the internet can be an awfully sexist, racist, and mean place.
(3) Intellectual property, copyright infringement, and deepfakes
GenAI models are built using as much data as they can get their hands on. This data grab has caused legal arguments about which data are freely available for training, and which data are not. And at the same time, there is a growing concern about how “content” created by GenAI may be too similar to training examples (i.e., real art) resulting in copyright infringement and plagiarism. Also, bad actors are actively using AI to spread misinformation and generate “deepfakes,” used for exploiting and extorting the masses, individuals, and minors.
(4) Opacity and the AI “black box”
To most, AI is seen as this “black box” — it’s difficult to understand how GenAI arrives at its final outputs. Although we can understand how it works in theory, this opacity to the public raises concerns about its accountability and explainability.
(5) Economic disruption
While GenAI is changing the shape of the economy, there are ways it’s changing things for the worse. Just as machinery did during the Industrial Revolution, AI in today’s age will make many jobs obsolete — most of them being lower-level positions. This will force many workers to adapt by honing their skills with AI, or by changing careers… Disruptive indeed.
O.K.
So if after reading all of that, you still are thinking… “actually, GenAI might be the right choice for my problem… I’m ready to try solving it with GenAI.”, keep reading – I want to share with you a set of principles that I use to guide how Harms Informatics implements GenAI.
-
Pragmaticism: the act of dealing with things in a sensibly and realistic way.
A route based on practically, rather than theoretical possibilities.This word came to me from a book near and dear to my heart — The Pragmatic Programer. This is a book written to describe the art of software engineering by defining a set of approaches to writing code, independent of language or framework. I’ve been following these approaches for a while now, and they’ve guided me (and many others) through the field of software engineering.
With this context, I came across an article from the Wall Street Journal describing how AI is being used in China to “[build] practical, low-cost tools that boost China’s efficiency and can be marketed easily.”
Reading this article and having a predisposition for pragmatism, I can’t help but believe that this is an appropriate approach for AI as well.
Harms Informatics’
Pragmatic Principles for GenAI
as of 2025/09/26
GenAI is an amazing tool, but it’s prone to mistakes. I’d essentially treat GenAI as an eager intern who is chronically online and really fast at reading and writing. It will attempt to reason through complex requests but it will often fall flat unless you provide additional context. You may also need to iterate to refine its answers — this may require a bit of handholding during ideation.
GenAI responses are only as good as the details you provide. Clear, direct prompts with specific context yield higher quality responses. Think of your prompt as you saying, “what all would I tell a co-worker if they were taking on this task? What information might they need?”
GenAI should only be used by subject matter experts in their subject field. In my humble opinion, you need to have the background knowledge to be able to validate (or call B.S. on) an AI’s response. When the stakes are low, experts can effectively use GenAI for subjects outside of their expertise; however, the responses must be validated by a human before any action is taken.
Take time to reflect on what you need GenAI for. What exactly is the task and how can AI help? What are the consequences if you full-heartedly trust the AI’s response? Should you be focusing on precision to avoid false positives? Or recall to avoid false negatives? Do you truly need AI to do this, or would this add complexity? Depending on your task, these answers change.
With these principles in mind, we’re now ready to effectively work with GenAI. So go ahead and start asking questions to your favorite LLM. But remember …
Context is key!
🔑
Context is key! 🔑
If you’ve worked with LLMs, you may have noticed this yourself: shorter prompts tend to give wilder answers, while the answers tend to focus (i.e., improve) as you add additional details.
This observation has been studied thoroughly, resulting in the field of Prompt Engineering – the act of designing inputs, or prompts, for GenAI models. You can think of prompts as the “instruction manual”, “recipe”, or “contract” for the AI to follow. Through trial and error, I’ve come up with a template which has worked well for me, and hopefully it can be a reasonable starting point for you!
Nate’s go-to prompt template
# Task executive overview
> Describe what you'd like done at a high level
> e.g., Explain quantum chemistry to me like I'm a 5 year old
> e.g., Given an input SMILES string, provide a letter grade (A, B, C, etc) corresponding to it's diffusivity in water. Explain why you provided that grade.
# Tone and persona context
> Describe how you'd like the AI to respond and think.
> e.g., you are an elementary school teacher. use a simple, engaging, and storytelling tone for educational purposes
> e.g., you are a college professor. use a formal, academic, and instructional tone for technical purposes
# Task context and background information
> Include references to files, websites, definitions which may be important.
> Anything which may be relevant to understanding the task.
# Step-by-step task description
> Provide detailed descriptions of each step for the AI to follow when solving the task.
> Also known as "chain-of-thought" prompting.
## Steps
1. ...
1.1. ...
1.2. ...
2. ...
3. ...
## Rules & Exceptions
- Always prioritize ...
- In step 2, never ...
- ...
# Output formatting
> Describe the deliverable. What specific "thing" are you expecting?
> e.g., a PDF of the description written in Times New Roman, 12pt font, single spaced, 1-inch margins on all sides. Include figures with subtitle descriptions, if applicable)
> e.g., a JSON response with the following schema: {"GRADE": <str>, "RATIONALE": <str>}
# Examples
> If you can, provide an example or two of how the AI should approach the task and what kind of outputs it should expect.
> By providing one or more high-quality examples of the task and the expected output, you give the model a concrete pattern to follow.With this kind of prompt, you should get higher quality responses, but remember, your response is still an unverified prediction from a model. So as David Thomas & Andrew Hunt said in The Pragmatic Programmer:
“[T]hink critically about what you read and hear. You need to ensure that the knowledge in your portfolio is accurate and unswayed by either vendor or media hype.”
So, where do we go from here?
My hope is that this blog provided you with some basic knowledge and guiding principles for using GenAI. We have a great and powerful tool and I think people ought to be using it. But we must use it in an ethical, practical, and pragmatic way.
We must be critical when deciding when to use this tool, and when to not. We can let these tools perform complex tasks, but we must trust our own reasoning first. We can be visionaries when applying these tools to novel problems, but we must do the due diligence and validate their responses.
In simplicity, GenAI is a tool for the people of today’s industrial revolution, but we must use it responsibly.
