Directly ask the model to perform a task, e.g.: ``` Summarize the following article in three sentences. [Input article] ``` 2. **Role-Based Prompts**
Assign the model a specific identity or role to guide tone and knowledge. ``` You are a career coach. Give me advice on how to negotiate a salary raise. ``` 3. **Chain-of-Thought Prompts**
Encourage step-by-step reasoning or breakdown of complex tasks. ``` What is the square root of 144? Explain the steps before giving the answer. ``` 4. **Zero-Shot Prompts**
Provide only the instruction and input — no examples. ``` Translate the following sentence into Portuguese: "How are you?" ``` 5. **Few-Shot Prompts**
Include examples of input/output pairs to guide the model. ``` Translate the following: English: Good morning → Portuguese: Bom dia English: Thank you → Portuguese: Obrigado English: I’m hungry → Portuguese: ``` > Note: a prompt can be of more than one type simultaneously, e.g. Prompt 1 above is also a zero-shot prompt.  ## 5.7 Prompt Comparison | Pattern | Use Case | Example Role/Task | | ---------------- | -------------------------------------- | --------------------------------------------- | | Instructional | Clear single-step tasks | “Summarize this paragraph.” | | Role-Based | Tone/personality alignment | “You are a financial advisor…” | | Chain-of-Thought | Complex reasoning, multi-step problems | “Explain each step before giving the answer.” | | Zero-Shot | Fast and generic tasks | “Translate to Spanish…” | | Few-Shot | Custom structure or format | “English → French examples…” | ## Prompt Pattern Comparison Example **Task**: Recommend three books for someone interested in artificial intelligence. The goal is to compare how different prompt formulations influence the quality, depth, and tone of the AI’s responses. ### 1️⃣ Instructional Pattern **Structure:** Clearly instruct the AI to perform a task or deliver information in a specific way. **Example Prompt:** ``` List and briefly describe three essential books for beginners who want to learn about artificial intelligence. ``` ### 2️⃣ Role-Based Pattern **Structure:** Ask the AI to _take on a persona_ or _expert role_ before performing the task. **Example Prompt:** ``` You are a university professor teaching an AI Literacy course. Recommend three foundational books that would help your students understand artificial intelligence. ``` ### 3️⃣ Chain-of-Thought Pattern **Structure:** Instruct the AI to _show its reasoning_ step-by-step before providing the final answer. **Example Prompt:** ``` Think step by step. First explain what qualities make a good introductory book on artificial intelligence, then recommend three that meet those criteria. ``` ### 4️⃣ Zero-Shot Pattern **Structure:** Provide the task _with no examples or additional guidance._ **Example Prompt:** ``` Recommend three books for someone interested in artificial intelligence. ``` ### 5️⃣ Few-Shot Pattern **Structure:** Provide _examples_ before giving the task, so the AI can mimic the style or format. **Example Prompt:** ``` Example 1: Topic: Data Science Books: 1. "Data Science for Business: What You Need to Know about Data Mining and Data-Analytic Thinking" by Provost & Fawcett, O'Reilly Media 2. "Python for Data Analysis: Data Wrangling with pandas, NumPy, and Jupyter" by Wes McKinney, O'Reilly Media 3. "Exploratory Data Analysis: Descriptive Statistics, Visualization, and Dashboard Design" by Leandro de Castro, CRC Press Now follow the same format for this topic: Topic: Artificial Intelligence ``` ## 5.8 Why Iteration Matters Prompting is often **not a one-shot process**. Even good prompts may return incomplete, vague, or misleading results. The key is to **test, analyze, and refine**. Iteration helps you: - Clarify vague requests - Add or remove unnecessary detail - Adjust tone or output format - Explore different framing approaches ## 5.9 The Iterative Prompting Process 1. **Draft the initial prompt** _Write a basic version of your request._ 2. **Review the output** _Is it accurate? Detailed enough? In the right tone or format?_ 3. **Refine the prompt** _Add missing instructions, clarify goals, or adjust roles._ 4. **Test again** _Compare outputs, note improvements, and repeat if needed._  ## Exercise 2: Prompt Iteration and Debugging Examples **Poor Prompt:** ``` Explain the stock market. ``` **Improved Prompt (Iteration 1):** ``` Explain the stock market in simple terms for a high school student, using 3 key points. ``` **Improved Prompt (Iteration 2):** ``` Explain the stock market in simple terms for a high school student. Use 3 bullet points and include one analogy related to daily life. ``` ## Exercise 3: Prompt Experiment Classwork **Goal:** Learn how prompt structure changes the quality, depth, and usefulness of AI-generated insights. Use the file **FinancialSample.xlsx** (provided below). It contains sales data by product, country, segment, and month — perfect for analytical prompts. **Activity Instructions:** You will interact with the same dataset using **four prompt patterns**. Paste each prompt into any LLM of your preference and record how the response changes. [Spreadsheet download](../Data/FinancialSample.xlsx) Available at: [https://learn.microsoft.com/en-us/power-bi/create-reports/sample-financial-download] **1. Instructional Prompt** ``` Analyze the dataset FinancialSample.xlsx and list the three countries with the highest total sales. ``` **2. Role-Based Prompt** ``` You are a financial analyst preparing a short executive summary. Using the dataset FinancialSample.xlsx, identify the three countries with the highest total sales, and explain what may be driving these results (e.g., product mix, segment, or profit margin). Write your answer in bullet points suitable for a management report. ``` **3. Chain-of-Thought Prompt** ``` Think step-by-step. 1. Load the dataset FinancialSample.xlsx. 2. Determine total sales per country. 3. Rank the countries in descending order of total sales. 4. Discuss one possible business implication from the pattern you observe. Provide your reasoning before the final answer. ``` **4. Few-Shot Prompt** ``` Example: When analyzing regional sales data, the analyst should: - Summarize the top-performing region - Identify key product contributors - Highlight one actionable insight Now apply the same structure to the dataset FinancialSample.xlsx. ``` ### Reflection and Comparison Create a short table comparing the outputs: | Prompt Type | Clarity of Output | Depth of Insight | Alignment with Task | Notable Difference | |--------------|------------------|------------------|---------------------|--------------------| | Instructional | | | | | | Role-Based | | | | | | Chain-of-Thought | | | | | | Few-Shot | | | | | ### Discussion Questions 1. Which prompt produced the most accurate or useful result? Why? 2. How did adding *role* or *examples* change the AI’s tone or structure? 3. How might bias or ambiguity in your wording affect the output? ## 5.10 Reverse Prompting: A Prompt Analysis Technique Beyond writing prompts, there are methods to **analyze and understand** how prompts work. These techniques help you learn, debug, and improve your prompting skills. Reverse prompting means starting from an **output** and asking: ``` What kind of prompt would generate this result? ``` It is useful for: - Understanding how specific outputs were created - Reverse-engineering effective prompt patterns - Teaching and learning prompt design by working backwards > Reverse prompting is not a prompt type itself — it is an **analytical method** for understanding model behavior and improving your own prompt engineering skills. To reverse-prompting you can use a technique called **meta-prompt**, which asks the LLM to reverse prompt an output. Here is a general-purpose structure you can adapt to your context and use: ``` Here is an AI-generated output: [PASTE OUTPUT HERE] Your task is to reverse prompt this output. That means: propose one or more possible prompts that could have generated it. Explain your reasoning and show how different phrasings might affect the response. ``` ### Reverse Prompting Example **Output (from the AI):** ``` 1. Set a consistent study schedule. 2. Use active recall and spaced repetition. 3. Take short breaks to avoid fatigue. ``` **Reversed Prompt:** ``` What are three key steps a student should follow when preparing for an exam? ``` This shows how reverse prompting can help: - **Understand AI outputs in their domain** (if you see a list of steps, you can practice inferring what kind of question led to it). - **Learn effective prompt patterns** (e.g., phrasing requests around “three key steps” often yields structured, actionable advice). - **Apply to your projects** by practicing both sides: writing prompts and analyzing outputs to refine your own strategy. ## Exercise 4: Reverse Prompting Below is an AI-generated output. Your task is to **work backwards** and write the possible prompt(s) that might have generated it, and then **use a reverse prompt method** to do the same and compare with your proposed prompts. **AI Output:** ``` 1. Identify a clear market pain point that your AI solution can address. 2. Build a prototype to demonstrate feasibility. 3. Prepare a pitch deck that communicates your value proposition and business model. ``` **Your Task:** - Propose at least **two different prompts** that could have led to this output. - Compare how each proposed prompt would shape the AI’s response. - **Use a meta-prompt to reverse prompt this output** and compare with your proposed prompts. ``` Here is an AI-generated output: 1. Identify a clear market pain point that your AI solution can address. 2. Build a prototype (MVP) to demonstrate feasibility. 3. Prepare a pitch deck that communicates your value proposition and business model. Reverse prompt this output: propose at least two different prompts that might have produced it. For each proposed prompt, explain why it would likely generate this output and how changing its phrasing might alter the answer. ``` ## 5.11 Tips for Troubleshooting and Improving Prompts - If output is **too generic**: Add specific instructions or examples. - If output **hallucinates**: Ask the model to “only use provided information” or to look for verifiable sources. You can also ask the model to cite the sources. - If output **misses important parts**: Reframe the prompt with clearer structure. - If output is **inconsistent**: Use role and format specifications. **Hallucination** is when a model (confidently) outputs incorrect or invented information. > Note: Once you understand how the LLM generates its output, hallucinations become a natural outcome and our role as moderators becomes evident. It is also important to acknowledge that the GenAI models can create statistics to provide data and results, and these statistics are not always directly available in the sources. ## 5.12 Assessing Prompt Quality Five criteria to judge the effectiveness of a prompt-output pair: | Criterion | Description | | ----------------------- | ------------------------------------------------------- | | **1. Relevance** | Is the response aligned with the prompt's intent? | | **2. Completeness** | Does the answer cover all requested aspects? | | **3. Clarity** | Is the response easy to understand and well-structured? | | **4. Factual Accuracy** | Are the claims and data points correct and verifiable? | | **5. Format** | Does the output follow the requested format or style? | **Techniques to Reduce or Identify Hallucinations** - Ask the model to **cite its sources** or say "Only use the provided text/file." - Clarify that the model should say **“I don't know”** if unsure. - Specify that the answer should be based **only on the uploaded file** or context. - Ask the model to **explain data and generalizations** presented. ## Reflection 1. How does communicating with an AI model differ from communicating with another person?
2. Which prompt pattern felt most intuitive to you, and which required the most experimentation?
3. How might prompt engineering change the way you research, write, or solve problems?
Reflect on how **intentional interaction** with AI changes your role from passive user to _active collaborator_. Each prompt you craft is a design decision — shaping not just what the AI produces, but how you think about inquiry, evidence, and creativity. > Remember that the goal of prompt engineering is not perfection, but **partnership**: using language to co-create meaning and insight. ## 📘 Further Reading - Prompt Engineering Guide, https://www.promptingguide.ai/ - OpenAI Platform, https://platform.openai.com/docs/guides/prompt-engineering - White, J. et al. (2023). _A Prompt Pattern Catalog to Enhance AI-Assisted Creativity and Problem-Solving._, arXiv: 2302.11382 https://arxiv.org/abs/2302.11382. - Zhou, Y. et al. (2023). _Large Language Models Are Human-Level Prompt Engineers._, arXiv:2211.01910 https://arxiv.org/abs/2211.01910 - Mollick, E. & Mollick, L. R. (2024). _Co-Intelligence: Living and Working with AI._, Portfolio. - Russell, S. & Norvig, P. (2020). _Artificial Intelligence: A Modern Approach_ — Chapter 24, “Natural Language Systems and Interaction.”, Pearson. - Dendritic Institute (2025). _AI Literacy Series – Module 5: Fundamentals of Prompt Engineering._ (Slides & video lecture)