Each week we find a new topic for our readers to learn about in our AI Education column.
We once believed we could learn by osmosis.
That by reading enough words, by sitting in lectures and breakout sessions listening to presenters drone on, by watching filmstrips and videos, by listening to podcasts and audiobooks, we would just absorb the information of the world and become a more intelligent person. Like a guru, we could be passive and just take it all in and emerge a better, more valuable person on the other side.
It doesn’t really work that way, of course, knowledge and skill are learned mostly in practice. It’s not what you know, it’s what you do. While there is a stage of learning where books and lectures are useful, to use and manipulate knowledge takes, well, work. We have to do it.
And it’s no different when using generative AI. For all its vaunted intelligence, AI is pretty dumb when you try to tell it what you want it to do, unless you have some practice and have learned how to craft your instructions to an AI model to get precisely the kind of results you’re looking for. With a good deal of experience under our belt, we often get what we want from AI on our first try.
As it turns out, the craft of honing the optimal artificial intelligence instructions has its own discipline, prompt engineering, and that’s what we’re going to discuss this week in AI Education. Personally, we’re still a work in progress on this front.
What Is a Prompt?
Most of us, at this point, have fired questions at an AI chatbot willy-nilly, whether using voice via a smart speaker or mobile device, or with text queries with Claude or ChatGPT or some other broadly available large language model. Far fewer of us have paid attention to how we’re asking those questions and the strategies we’re using to hone in and get the response we want.
To discuss prompt engineering, we first need to understand the prompt. At its simplest, a prompt is information supplied to an AI system that tells or helps it determine what to do.
Amazon Web Services defines a prompt as natural-language text asking generative AI to perform a task. Prompts can range from a question such as “What is duration?” to elaborate instructions containing background material, examples, data, formatting requirements and restrictions. SAP similarly describes a prompt as an input or command that instructs an AI system to perform a particular task or generate a particular response.
The everyday conception of a prompt is something typed into ChatGPT, Claude, Gemini or another AI interface. But prompts inside commercial AI applications can be much more complicated. They can combine instructions written by developers with material entered by users and contextual information supplied by software. The prompt the model actually receives may bear little resemblance to the short request entered by the human user.
What Is Prompt Engineering?
Prompt engineering is the systematic practice of designing, testing and improving those inputs so that an AI system produces more useful, accurate, consistent and appropriate outputs. McKinsey describes it succinctly as designing AI inputs to produce optimal outputs. AWS calls it the process of guiding generative AI toward desired results by selecting appropriate formats, phrases, words and symbols. Stanford describes the discipline as designing and refining inputs to guide AI-model behavior. Oracle emphasizes its iterative character: users experiment with wording, formatting, contextual information and other variables until outputs better meet their objectives.
That final point is important. Prompt engineering is not about discovering a magical phrase that “unlocks” an LLM. It resembles experimental design more than the casting of spells. A prompt engineer establishes an objective, designs an instruction, observes the result, identifies deficiencies, modifies the prompt and tests again. For production applications, organizations may evaluate different prompts against test datasets and track qualities such as accuracy, consistency, latency, cost and adherence to required formats.
Prompt engineering also increasingly overlaps with context engineering. IBM’s 2026 prompt-engineering guide describes the progression from simply writing better prompts toward designing the broader context available to a model—including user intent, conversation history, retrieved information and structured data. Retrieval-augmented generation (RAG), for example, can insert relevant proprietary information into the context supplied to an LLM.
Use Cases
The range of possible applications is enormous because prompting applies wherever generative AI does. AWS highlights classification, subject-matter applications, analytical reasoning and creative work, while IBM identifies applications including chatbots, software development, cybersecurity and specialized knowledge work.
Financial applications are particularly easy to imagine. An analyst can prompt an AI system to compare quarterly earnings transcripts and identify changes in management language. A compliance employee can ask it to compare a communication against internal policy. An investment researcher can instruct it to extract specified metrics from filings. An advisor can generate a first draft of a client meeting summary. Operations employees can classify documents or correspondence. Developers can generate and review code.
Prompting becomes even more useful when organizations standardize repeatable tasks. Instead of allowing 500 employees to invent 500 prompts for summarizing an investment document, a firm can create a tested template specifying which information should be extracted, which sources may be used and how the answer should be structured. OpenAI’s financial-services resources, for example, describe enterprise prompts for data analysis and modeling, research and synthesis, policy and regulatory interpretation, contract and covenant analysis, data extraction, and Excel, business-intelligence and ERP workflows.
A Prompt Engineering Toolbox
The simplest technique is zero-shot prompting: ask the model to perform a task without giving it examples.
“Classify the following client email as a service request, complaint, trade instruction or other.”
When the desired result is difficult to describe, few-shot prompting can provide examples demonstrating what the user wants. IBM identifies zero-shot and few-shot prompting among core techniques.
Another technique is role or persona prompting, in which the model receives a perspective or function: “You are assisting a financial advisor preparing for a client review.” The role does not magically give the model professional credentials or new knowledge. Its usefulness lies in supplying context about audience, priorities, vocabulary and expected behavior.
Constraint prompting establishes boundaries: use only specified documents; stay below 500 words; produce JSON; don’t calculate values when required inputs are missing; separate facts from assumptions.
Structured-output prompting defines the desired answer architecture. Instead of requesting “an analysis,” the user can request columns for company, revenue, operating income, year-over-year change, source and source date.
Prompt chaining decomposes complicated work into stages. Stanford includes prompt chaining among its prompting techniques. A research workflow might first extract facts, then classify them, then compare them with previous periods, and finally draft a narrative based on the validated results.
Techniques commonly discussed under the prompt-engineering umbrella also include generated-knowledge prompting and various reasoning-oriented approaches. The larger principle is decomposition: complex tasks often become more manageable when they are divided into clearly specified operations rather than packed into one sprawling instruction.
Best Practices
- Start with the desired outcome. Users frequently begin composing a prompt before deciding exactly what they want. A better process defines the deliverable first.
- Then provide relevant context. Identify the audience, purpose, available information and important constraints. SAP’s guidance recommends defining context and the desired outcome and providing detailed guidance about length, format and style.
- Tell the model what information it should use. This becomes particularly important in finance. “Answer from the attached prospectus and cite the relevant passages” creates a fundamentally different task from “Tell me about this fund.”
- Specify the output. Tables, schemas, headings and required fields reduce ambiguity and make outputs easier to evaluate automatically.
- Provide examples when describing the result is harder than showing it.
- Break complicated workflows into stages. Extraction, calculation, interpretation and communication do not necessarily belong in the same prompt.
- Test prompts against difficult cases rather than judging them from one impressive response. Because LLM outputs can vary, production prompts should be evaluated systematically. AWS explicitly notes the stochastic nature of model responses.
- Finally, verify the result. Prompt engineering improves the probability of obtaining a useful output; it does not make an LLM authoritative.
Become a Better Prompt Engineer
The most productive training method is practice tied to actual work.
Take one recurring task—summarizing research, preparing for client meetings, comparing funds, reviewing earnings reports or drafting internal communications—and create a reusable prompt. Define the objective, context, approved sources, constraints and output format. Run it against several examples. Record where it fails. Revise it.
Then compare results across models. Stanford’s training materials emphasize iterative refinement and comparing prompt behavior across different AI tools.
Build a personal or firm-approved prompt library, but resist treating prompts as immutable recipes. Models change, applications change and tasks change. A good prompt is a tested starting point, not scripture.
Most importantly, develop the habit of interrogating outputs. Ask: What information did the model use? What did it assume? What is missing? Can the claims be verified? Is the format suitable for review? Would the answer change if the task were phrased differently?
That mindset turns prompting from conversational trial and error into professional practice.
