top of page
Writer's pictureMorgan Hunter

GPT Prompt Engineering: 101

GPTs cannot read your mind. They are also not magical. When interacting with artificial intelligence start by imagining how you would ask your fellow human to answer the question or complete a task. Now that you're in the right headspace, here are some specifics to help you on your way. But first...what is prompt engineering?


Prompt engineering involves designing the initial input or "prompt" given to a model in a way that elicits the most accurate and relevant output. This blog explores strategies and tips for effective GPT prompt engineering.

 

Let's Go!

GPT models generate responses based on patterns and associations learned during training from a diverse dataset. The model's performance is contingent on how well the prompt is structured. Follow these examples to improve your GPT responses.

 

1) Be Clear and Specific

A clear and specific prompt will yield better results. Be explicit about what you need. For example, instead of asking, "How do I make a cake?" specify the type of cake and any dietary restrictions or preferences, such as "How do I make a gluten-free vanilla cake with ingredients commonly found at home?"


Prompt: Write a 500-word article on the benefits of intermittent fasting for athletes, focusing on muscle recovery and performance enhancement.


Explanation: This prompt is specific about the topic (intermittent fasting for athletes), the length of the article (500 words), and the aspects to focus on (muscle recovery and performance enhancement).

 

2) Break It Down

For complex requests, breaking the prompt into smaller, manageable parts can be more effective. Instead of a broad prompt that may lead to a generalized response, ask specific questions that guide the model through the thought process you want to explore.


Prompt: Explain the process of securing a home loan in three parts: eligibility criteria, necessary documentation, and the typical timeline from application to approval.


Explanation: The prompt breaks down the complex process into manageable sections, guiding the model to structure its response in a clear, step-by-step manner.

 

3) Use Instructions

Instructive prompts tell the model how to structure its response, enhancing the likelihood of receiving the output in the desired format. For example, if you need a summary, you might begin your prompt with "Summarize the following text," ensuring that the model knows to condense rather than discuss or analyze the text.


Prompt: Summarize the following text in bullet points, highlighting the main argument, supporting evidence, and conclusion.


Explanation: The prompt specifies not only that the text should be summarized but also dictates the format (bullet points) and the elements to include (main argument, evidence, conclusion).


4) Provide context

Providing context helps the model generate more accurate and relevant responses. If the task involves a specialized topic, including a brief background or defining complex terms within the prompt can significantly improve the output.


Prompt: Discuss the impact of the 2008 financial crisis on European banking regulations, considering the changes in policies introduced by the European Central Bank post-crisis.


Explanation: This prompt provides contextual background (2008 financial crisis) and specifies the focus on resultant European banking regulations, aiding the model in tailoring the response to the specific after-effects of the crisis.

 

5) Use Feedback Loops

Use the model's responses to refine your prompts. This iterative process can help you discover the optimal ways to communicate with the model for different types of tasks. Each interaction is an opportunity to learn which prompt strategies are most effective.


Initial Prompt: Describe the steps involved in baking bread.

Response: (Lists generic steps)


Refined Prompt: Can you provide more detail on the proofing process mentioned in your response, including time and temperature guidelines?


Explanation: Using the response to the initial prompt, the refined prompt asks for specific details about a part of the process that was too general, guiding the model to provide the needed depth.

 

 

6) Leverage Examples

When appropriate, provide examples within the prompt to illustrate exactly what kind of response you are looking for. This can be particularly useful for creative tasks or when the output needs to follow a specific format.


Prompt: Write a complaint letter to a mobile phone company about a recent issue with service outages. Model it after this example: start with a greeting, state the issue, mention how it affected you, demand a resolution, and close with your contact information.


Explanation: The prompt uses an example structure for the letter, ensuring that the model produces a response that meets specific expectations for format and content.

 

7) Consider the Model’s Limitations

Understanding what GPT does not do well is just as important as knowing its strengths. The model may struggle with highly technical content, recent events, or personal advice. Tailoring your prompts to avoid these pitfalls can prevent irrelevant or incorrect responses.


Prompt: Provide a historical overview of the Roman Empire's expansion but avoid discussing specific battles as I am focusing on political and diplomatic strategies.


Explanation: Acknowledging that detailed accounts of battles might not be the model's strength, this prompt directs the focus to political and diplomatic strategies, areas where the model can provide reliable information based on historical data.

 

8) Experiment and Learn

Lastly, the key to mastering prompt engineering is experimentation. The more you interact with GPT models, the better you will understand how your prompt choices influence the responses. Don’t be afraid to try different approaches and learn from unsuccessful prompts.


Initial Prompt: Generate ideas for a science fiction story set in outer space.

Response: (Basic ideas)


Refined Prompt: Based on your previous response, can you develop a detailed plot for the idea involving a rogue AI on a space station?


Explanation: After reviewing the initial basic ideas, the refined prompt asks for a detailed plot expansion on a specific concept, demonstrating how iterative prompting refines and improves outcomes.



GPT Prompt Template

Now that you know the basics you can use this template to create a GPT prompt:


Objective: [Clearly state what you need]

E.g., "Create a detailed plan," "Summarize the key points," etc.


Subject: [Specify the subject or topic clearly]

E.g., "vegan chocolate chip cookies," "market analysis for the tech industry in Q2 2021," etc.


Details:

1. [List any specific details or parameters]

E.g., "Use coconut oil instead of butter," "Focus on changes due to emerging technologies," etc.

2. [Add more specific elements or constraints as needed.]

3. [Include any specific formatting or structural preferences]

E.g., "List ingredients first, followed by steps," "Use bullet points for key insights," etc.


Additional Instructions: [Any other instructions or preferences].

E.g., "Provide sources if possible," "Keep the tone professional," etc.

 

Completed Example


Objective: Write a blog post.

Subject: The impact of remote work on urban economies.

Details:

1. Discuss how remote work has changed city landscapes and local businesses.

2. Mention specific cities as case studies: New York, London, and Tokyo.

3. Conclude with predictions for the next five years.

Additional Instructions: Use a conversational tone and include relevant statistics.


Conclusion

Effective prompt engineering is crucial for harnessing the full potential of GPT models. By being clear, specific, and thoughtful about the information you provide and the questions you ask, you can improve the quality and applicability of the model's responses. Whether you're using GPT for content generation, data analysis, or customer service, these strategies will help you communicate more effectively with one of today's most powerful AI tools.

19 views0 comments

Recent Posts

See All

Comments


bottom of page