Look, I spend my days helping organizations – from established enterprises to ambitious startups – navigate the real world of AI. Not the hype, not the theoretical, but the actual implementation. And right now, the biggest bottleneck isn’t compute power, or data availability (though those are challenges, especially here in the MENA region). It’s prompt engineering. Specifically, business leaders understanding what it is, why it matters, and how to leverage it.
Beyond “Asking Nicely”: The Core of Prompt Engineering
Most people think prompt engineering is just about phrasing your question to ChatGPT in a clever way. It’s… not. It’s a discipline. It’s about understanding how Large Language Models (LLMs) – the engines behind these tools – think. They don’t “think” like humans. They predict the most probable next token (word or part of a word) based on the massive dataset they were trained on. Your prompt isn’t a request; it’s the beginning of a text sequence. You’re essentially steering that prediction.
We saw this acutely with a client, a large bank in Lebanon, trying to automate customer service responses. They initially tried feeding the LLM raw customer complaints. The results were… disastrous. Generic, unhelpful, and sometimes even offensive. Why? Because the LLM was predicting based on the entirety of its training data, not just the context of banking or customer service. The fix wasn’t a better LLM; it was a meticulously crafted prompt that included:
- Role Definition: “You are a highly empathetic and knowledgeable customer service representative for [Bank Name].”
- Task Instruction: “Respond to the following customer complaint in a professional and helpful manner.”
- Contextual Information: “The customer is a premium account holder. Refer to our internal knowledge base for specific product details.”
- Output Format: “Respond in a concise paragraph, followed by a list of relevant FAQs.”
Suddenly, the responses were relevant, helpful, and on-brand. That’s prompt engineering. It’s about controlling the prediction, not just asking a question.
The Rise of RAG and Why It Changes Everything
For a long time, prompt engineering was about maximizing what was already in the LLM. Now, we’re entering the era of Retrieval-Augmented Generation (RAG). RAG is a game-changer, especially for businesses dealing with proprietary data. Instead of relying solely on the LLM’s pre-trained knowledge, RAG allows you to feed it relevant information at the time of the prompt.
The technique traces back to the original Retrieval-Augmented Generation paper, the 2020 paper that introduced retrieval-augmented generation as models which combine pre-trained parametric and non-parametric memory for language generation.
Think of it like this: the LLM is a brilliant generalist. RAG gives it access to your company’s specific expertise. We built a RAG system for a logistics company in Dubai. They had decades of operational data – shipping routes, customs regulations, pricing agreements – locked in spreadsheets and databases. Without RAG, the LLM couldn’t answer questions like “What’s the fastest route to ship goods from Jebel Ali to Beirut, considering current customs delays?” With RAG, it could. The prompt doesn’t need to contain all that information; it instructs the LLM to retrieve it from a designated knowledge base.
This is critical in the MENA region. Data privacy regulations are evolving, and many organizations are hesitant to share sensitive information with third-party LLM providers. RAG allows you to keep your data secure while still leveraging the power of AI.
Prompt Libraries: Your Competitive Advantage
Don’t treat prompt engineering as a one-off task. It’s an iterative process. And the best prompts aren’t discovered; they’re built. That’s where prompt libraries come in. A prompt library is a centralized repository of tested and optimized prompts for specific business use cases.
At Webspot, our applied-AI consultancy in Lebanon, we’ve developed internal prompt libraries for common tasks like content creation, data analysis, and code generation. But more importantly, we help our clients build their own libraries, tailored to their unique needs and data. This isn’t just about saving time; it’s about creating a competitive advantage. If you can consistently generate higher-quality outputs with AI than your competitors, you’re going to win.
I often tell leaders: “Your prompt library is your new intellectual property. Protect it, refine it, and treat it as a core asset.”
The Human-in-the-Loop Imperative
LLMs are powerful, but they’re not perfect. They hallucinate (make things up), exhibit biases, and can be easily misled. Blindly trusting AI-generated outputs is a recipe for disaster. That’s why a “human-in-the-loop” approach is essential. This means having a human review and validate the LLM’s outputs before they’re used for critical business decisions.
This is particularly important in contexts like Lebanon, where trust and reputation are paramount. A single inaccurate or insensitive AI-generated response could severely damage a company’s brand. The human reviewer isn’t just checking for errors; they’re ensuring that the output aligns with the company’s values and ethical standards.
Think of the human as an editor, not a replacement. They refine, contextualize, and ensure quality. This isn’t about slowing things down; it’s about building trust and mitigating risk.
Beyond Text: Multimodal Prompts and the Future
Prompt engineering isn’t limited to text. LLMs are increasingly becoming multimodal, meaning they can process and generate different types of data – images, audio, video, code. This opens up a whole new world of possibilities. Imagine a prompt that says: “Analyze this customer support transcript and the associated screen recording. Identify the key pain points and suggest improvements to the user interface.”
We’re starting to see this in areas like fraud detection, where LLMs can analyze both textual data (transaction descriptions) and visual data (images of checks or credit cards). The ability to combine different modalities will be a key differentiator in the coming years.
I discuss these emerging trends in detail in my book, Applied AI for Future Ready Organizations. It’s a practical guide for leaders who want to move beyond the hype and start building real AI solutions.
The 2026 Correction: Supervision, Not Prompting, Is the Scarce Skill Now
This article first ran in March 2026, when spelling out a role, a task, contextual information and an output format was the difference between a usable model response and a generic one. Five months on, that gap has mostly closed. Current-generation models default to much stronger instruction-following: they infer role and format from a short, plain request, ask a clarifying question when the ask is ambiguous, and no longer need the four-part scaffolding the Lebanese bank example above required just to get a usable customer-service reply. The craft described in this piece has not become useless — it still sharpens a one-off request, and a documented prompt library is still worth building — but it is no longer the scarce, differentiating skill it was in March. It is table stakes now, the way a clean spreadsheet formula was table stakes a decade after spreadsheets shipped.
What the models have not absorbed is judgment about autonomy. Anthropic's own guidance on building agents is explicit that the deployments worth trusting are the ones that integrate meaningful human oversight alongside clear success criteria — oversight is a design requirement, not an afterthought bolted on later. OpenAI's documentation for its Agents SDK draws the same line at the control level: guardrails run automatically, but for anything sensitive, human review pauses the run so a person or policy can approve or reject a sensitive action. Two competing vendors, describing the same leadership decision in different words: what may an autonomous system do without a person watching, and what must stop and wait for one.
That decision does not write itself into a prompt. It gets drawn organization by organization — a spending limit here, a customer segment that may never be contacted automatically there, a message class that always needs sign-off before it goes out — and then it has to be checked, not just declared, because a policy nobody verifies quietly stops being true. I now spend more client time on that boundary and its checks than on prompt wording. Prompt engineering earned its place in this article; agent supervision is what earns a leader's attention today.
Actionable Takeaways: Start Today
- Invest in Training: Don’t expect your team to become prompt engineering experts overnight. Provide them with training and resources.
- Start Small: Identify a specific business problem that can be solved with AI. Don’t try to boil the ocean.
- Build a Prompt Library: Document your best prompts and share them across your organization.
- Implement Human-in-the-Loop: Always have a human review and validate AI-generated outputs.
- Explore RAG: If you have proprietary data, investigate how RAG can unlock its value.
- Define the Supervision Boundary: Write down, in specific terms, what your agents may do without a person watching — and build the check that catches it when they cross that line anyway.
The future of AI isn’t about building smarter algorithms; it’s about learning how to communicate with them effectively. Prompt engineering is the key. And it’s a skill that every business leader needs to master. You can find more of my thoughts on AI strategy at jonahtebaa.com.
