Prompt Engineering: Part 1 – Introduction to Effective Communication with LLMs

If you’re a power user of large language models (LLMs) or a developer who’s been curious about integrating LLMs into your apps, you’ve probably heard about prompt engineering. At its core, prompt engineering is about asking the right questions clearly and effectively. Think of it as guiding the model towards giving you useful responses.

In this segment, we will talk about what you need to know in just a few minutes to effectively interact with any LLM model. We will use several real-world examples to illustrate these best practices. In Part 2, we will focus on how developers can build these best practices in their applications to facilitate obtaining the best answers from agentic systems.

Why Prompt Engineering Matters

A great prompt helps LLMs like ChatGPT, Claude, and Gemini understand exactly what you are looking for. On the flip side, a vague prompt can lead to ambiguous answers or completely irrelevant outputs. Effective prompt engineering reduces the guesswork, saves development time, and ultimately gets you more accurate results. Although prompts can be multimodal, meaning that they could include text, images, videos, and/or audio inputs, we will focus on text-only prompts in this document. (Note: Newer LLM releases such as ChatGPT 5 and Claude 4 handle vague prompts better than earlier models, but clear and specific prompting still significantly improves accuracy and reduces hallucinations.)

How Prompt Engineering Works

An effective prompt generally has four components: Persona, Context, Task, and Format. The Persona component usually comes first, followed by either the Task or Context component (the order of these two sections doesn’t matter), ending with the desired Format expected for the response. Here is a general description of how these four components should be structured:

• Persona: Define clearly the role you want the AI to assume (e.g., product manager, sales executive).
• Context: Provide necessary background information (e.g., current project details, specific business opportunity).
• Task: Clearly state what you want the AI to accomplish (e.g., generate meeting agenda, document methods, or list use cases).
• Format: Specify the desired output format (e.g., answer in three bullet points, generate a structured table, and/or provide inline comments).

Example (Illustrating Major Components of a Text Prompt):

You are a specialist facilitator for cross-functional product teams in industrial software. I’m a product manager at Digital Industries. My team (software architect, 4 developers, QA, marketing/strategy, and myself) is having a full-day planning meeting. We want to get everyone on the same page for a new AI-powered topology optimization module for our 3D CAD system. Create a specific agenda for this meeting. Allocate a time slot to each individual or group depending on their function. Take breaks every two hours, and include a shared brainstorming session and a concluding alignment session. Output should be a markdown table with the following columns: Time, Duration, Presenter, Topic, Goal.

Key to identifying component:      Persona                          Context                          Task                   Format

Be clear and precise/specific in your prompts, which may require longer prompts for more clarity and context. If the description of the Context becomes too long, it is fine to include the Task both before and after it. It is also important to include a verb or command as part of the Task, and stating how you want the response to be formatted.

Other Key Prompt Engineering Techniques

The following techniques will help you further improve communication with LLMs.

1. Role Prompting

Specify clearly who the model should pretend to be. It helps the model set the right tone, language, and specificity.

Example (Technical Documentation Generation):

You are a senior software engineer writing documentation for an API. Clearly explain how to authenticate requests and handle errors.

2. Chain-of-Thought Reasoning

Ask the model to reason step-by-step. This is especially effective in tasks that involve logical reasoning or calculations.

Example (Simulation Software):

Perform thermal analysis on the system using the specified boundary conditions. Then, double the temperature of the base plate, repeat the thermal analysis and compare the two simulations. Think through the process step-by-step:

• Define the assumption and boundary conditions
• Describe the simulation methods and results of the first simulation
• Predict how doubling base temperature might affect the system
• Perform the modified simulation to observe change
• Compare both results and explain different across heat stress and gradients
• Conclude with the physical reasoning of observed change

Clearly explain your reasoning at each step and summarize key takeaways.

This kind of detailed prompt encourages the model to lay out the logic clearly, reducing mistakes.

3. Few-Shot Prompting

Give your LLM a few examples of what you’re expecting. This works well when you need the model to understand the format or style you are after. In the following example, the first two User/Agent sentences are written by you to show the LLM how you want its answer to look. Then, you pose the question and wait to hear the response from the LLM in the desired form.

Example (Customer Support AI):

User: How do I reset my password?
Agent: To reset your password, go to the login page and click on “Forget Password”. Follow the instructions on the login page (which also might be sent to your registered email address) and create a new password.

User: How do I update my billing information?
Agent: Firstly, go to your profile. Open the Account and select the “Billing” tab. In the billing tab, you should see the Billing Information. Go ahead and update your payment details. Don’t forget to click “Save” to apply these changes.

User: How can I delete my account?
Agent: (Wait for the answer from the LLM)

4. Multi-stage or Chain Prompting

Multi-stage or chain prompting refers to breaking down a complex task into smaller logical steps and then asking the LLM for solutions. It is highly effective for complex tasks as LLMs have limited context window and can benefit from guided stage-based reasoning.

Example (Topology Optimization AI):

Prompt I – You are given a rectangular beam fixed on both sides, identify the boundary condition and possible loading scenarios for topology optimization (wait for LLM answer)

Prompt II (Chain) – Based on the scenario 1 you suggested, outline the CAD design objectives and constraints for an AI-driven topology optimization. (wait for LLM answer)

Prompt III (Final) – Suggest any two design features that satisfy the design objective under those constraints. (wait for LLM answer)

There are many other Prompt Engineering techniques, which you can learn by studying the documents in the Resources section below.

An Example of Ineffective vs. Effective Prompt

Here is an illustrative example tailored specifically to Digital Industries, highlighting an ineffective prompt and its improved, effective counterpart.

Ineffective Prompt to ChatGPT 4o:

Tell me about automation in digital industry.

Response from ChatGPT 4o:

Automation in digital industries involves using technology to perform tasks without human intervention. It can increase efficiency and reduce errors.

Effective Prompt to ChatGPT 4o:

List three specific use cases of Robotic Process Automation (RPA) within manufacturing in Digital Industries and briefly explain their benefits.

Response from ChatGPT 4o:

1. Order Processing Automation: Reduces manual data entry errors, accelerates order fulfillment, and improves customer satisfaction.
2. Quality Inspection Automation: Enhances product quality through consistent, accurate defect detection, lowering rejection rates.
3. Predictive Maintenance Scheduling: Minimizes downtime by proactively identifying equipment failures, thereby reducing operational costs.

As you can see, the improved prompt explicitly defined the context (Robotic Process Automation), domain (manufacturing within Digital Industries), and desired format/length of the response (three specific use cases). Focusing on manufacturing ties directly to practical applications within Digital Industries, resulting in concise and relevant answers. Furthermore, requesting a numbered list ensures responses are organized and actionable. Note that I wanted a short response to place in this document. In your case, you could get a more detailed response if you provide the role and ask the LLM to frame its response around structured thinking, explain its reasoning step-by-step, and/or use few-shot prompting as explained above.

Other pitfalls to avoid as you create your own prompt:

• Ambiguous Instructions: Be explicit. Instead of saying, “Explain this,” say, “Explain this as if teaching a new programmer.”
• Overloading Requests: Keep prompts concise. Don’t ask multiple unrelated questions in one prompt.
• Ignoring Structure: Structure prompts clearly, breaking down complex tasks into smaller sub-prompts if needed.
• Outdated Information: When tasks require up-to-date or domain-specific data, combine prompt engineering with Retrieval Augmented Generation (RAG) or fine-tuned embeddings to ground the model in reliable sources.

What Should Not Be Done with Prompt Engineering

Prompt engineering is powerful, but certain boundaries must be respected, especially in the software business context. In other words, to avoid ethical and legal risks, interactions with LLMs need to abide by certain guardrails that go beyond technical optimization for functionality. A few examples are provided below. (Note that the list below is not exhaustive and, because compliance requirements can change rapidly, you should check the organization’s latest AI usage policy or data governance rules before using LLMs for any sensitive tasks.)

• Sharing Proprietary or Sensitive Company Information: LLMs, especially third-party hosted models, are not designed for secure data handling. Sharing proprietary data risks unintended leaks or exposure. Internal IP or sensitive employee data must remain confidential and secure.
• Requesting Competitor Data or Non-Public Information: LLMs are trained on publicly available data; attempting to extract non-public or private competitor data is unethical and typically impractical. Such attempts may also inadvertently produce misleading or fabricated responses.
• Generating or Validating Source Code Without Expert Review: Although LLMs can generate code, they cannot ensure legal compliance with industry standards. Risks include potential infringement of third-party IP, and potential “Copyleft” infringement if the copied code was governed by a “Copyleft” open-source license. To avoid such issues, developers should always review and validate AI-generated code.
• Relying on Prompt Engineering for Precise, Real-Time Data Queries: Without special techniques such as Retrieval Augmented Generation (RAG), LLMs do not have real-time data access or database querying capabilities.
• Attempting Automated Security or Access Control Operations: LLMs are inherently probabilistic and not suitable for mission-critical security decisions or authorization processes. Automated access via prompts can unintentionally bypass crucial security checks or validation.
• Expecting Legal or Regulatory Advice: Models provide generalized information and do not account for nuanced legal contexts. Relying on prompt-generated advice for compliance or legal decisions poses significant risks.
• Embedding LLM Outputs Directly into Critical Customer-facing Applications Without Validation: Prompt-generated outputs can contain errors, biases, or inappropriate content. Direct integration without review can damage trust, brand reputation, and/or cause operational issues.

In Part 2, we will provide more examples and dive deeper into topics tailored specifically for software developers of AI-enabled applications.

Resources for Going Further

To dive deeper into prompt engineering, here are some concise and reliable resources:

• OpenAI Cookbook – Prompt Engineering Examples: Practical examples covering everything from basic prompting to JSON formatting (see here).
• DeepLearning.AI – ChatGPT Prompt Engineering for Developers: Interactive course great for developers who want hands-on practice (check it out).
• Anthropic Claude Prompting Interactive Tutorial: Clear tips and best practices useful regardless of the LLM you’re using (visit tutorial).
• A Framework for Prompt Engineering and AI Interaction Design: Examines how prompt engineering can be approached responsibly to mitigate risks (read paper).

Final Thoughts

Traditional approaches to improve performance of GenAI models, which include retraining and fine-tuning, require substantial technical expertise and computational resources. Prompt engineering offers an additional pragmatic and efficient method of influencing AI behavior and improving its output, without any changes to the underlying model architecture. By mastering a few key strategies as discussed here, you will quickly improve the effectiveness of interactions with LLM-powered applications, saving time and reducing frustrations along the way. Give these techniques a try in your next project; you will notice the difference immediately. (Note: This document was prepared with the help of AI.)