ChatGPT API Prompting Guide & Best Practices

A Developer's Guide to Prompt Engineering

This is not a programming or ChatGPT API integration tutorial, these are some key ideas and points to keep in mind while using ChatGPT API.

Large Language Models provide unprecedented capabilities to your software products. As developers, it is crucial to learn this tech and integrate it within our projects. Platforms like OpenAI provides API to interact with these models and we can leverage that to enhance our software.

Recently, I completed a fascinating course, "ChatGPT Prompt Engineering for Developers." I learned some incredible techniques to get the most out of OpenAI's ChatGPT API, and I'd love to share those insights with you in this post.

Guidelines

When interacting with ChatGPT, clarity is crucial. Here are some useful strategies:

Separate instructions and input texts: This allows you to test your instructions with various inputs. For example, if you're asking ChatGPT to summarize an article, make the instruction "Summarize the following article:" separate from the article text itself.
Use delimiters: They help specify the boundaries between the instruction and the input, reducing the likelihood of prompt injections. A colon or a new-line can serve as simple but effective delimiters.
Request structured output: If you need to parse ChatGPT's response programmatically, ask it to structure its output in a specific way. For instance, you might request, "List the key points from the following text as bullet points:".
Check input conditions: To reduce the chances of the model producing irrelevant or inaccurate responses (a phenomenon known as "hallucination"), specify any conditions the input must meet. For example, "If the text contains a date, provide the day of the week for that date."
Few-Shots Prompting: If possible, providing a few examples of the desired input-output pattern can guide the model to produce similar results.

Iterative Prompt Development

Crafting the perfect prompt with ChatGPT is indeed an iterative process, and it requires a keen understanding of the objective and a willingness to experiment and learn.

Let's break it down into actionable steps:

Define the Objective: The first step in prompt crafting is to clearly understand the desired output. Ask yourself, what do you want ChatGPT to generate? For example, if your goal is to extract key points from a text, your objective would be a list of the main ideas from the input text.
Write an Initial Prompt: Based on your objective, write an initial prompt. A prompt is a command or question given to ChatGPT to direct its response. For the objective mentioned above, an initial prompt could be, "List the main ideas in the following text:".
Test the Prompt: Now, it's time to test the prompt. Run it through the model and see what kind of output you get. Is it in line with your objective? If not, it's time to iterate.
Analyze the Output: Analyze the output from the model. What aspects of the response align with your objective, and what parts are off target? This will give you clues about how to adjust your prompt.
Refine the Prompt: Based on your analysis, adjust your prompt. For instance, if the model isn't quite capturing the main ideas as you intended, you might refine the prompt to be more specific, such as, "Summarize the following text into bullet points:".
Repeat the Process: Continue to test, analyze, and refine your prompt until it consistently generates the desired output. Remember that this is an iterative process. It might take several rounds of refinement to get it just right.
Generalize the Prompt: Once you have a prompt that works well for a specific case, try to generalize it for other similar cases. This means testing the prompt with a variety of input texts to ensure it works across a broad range of scenarios.

Summarization

Summarization is a powerful use case of ChatGPT API, but it's essential to know how to use it effectively. Here are some guidelines:

Define the purpose: If the summary is going to be used in a specific way, make sure to specify this in the instruction.
Stay focused: Ask the model to focus on specific parts of the input if needed. For instance, if you're only interested in the financial aspects of a business report, you might ask, "Summarize the financial information in the following report:".
Extract instead of summarizing: In some cases, it might be more useful to extract key information instead of summarizing. For instance, you could prompt, "List the names of all the people mentioned in the following text:".

Inferring

ChatGPT can also infer insights from a text, performing tasks such as sentiment analysis, categorization, classification, and labeling. Here are some ways you can leverage this functionality:

Multitask: You can ask the model to perform more than one task in the same instruction, and to generate the output in a specific format. For example, "Analyze the sentiment of the following review and classify it as positive, negative, or neutral:".
Test on multiple examples: It is possible that a prompt that works well on one set of inputs does not perform as well on others. To ensure that your instruction works well across a broad range of inputs, test it with a variety of examples.
Include your own labels/tags: If you have a predefined set of categories, you can include them in the instruction, and ask the model to select the most relevant one(s) for the given input. For instance, "Classify the following text into one of these categories: Technology, Environment, Politics, or Culture:".

Transforming

ChatGPT is also capable of performing a variety of text transformations, including language translations & format transformations.

Language Identification & Translation: You can ask the model to identify the language of an input text, or translate it into another language. For instance, "Translate the following Spanish text into English:".
Tone Transformation: You can transform an input text into a different tone, such as formal, casual, or conversational. For example, "Rewrite the following formal text in a casual tone:".
Format Transformation: The model can transform text from one format to another, such as from JSON to HTML or CSV to JSON. For instance, "Convert the following JSON data into HTML table format:".
Proofreading: You can ask the model to proofread a text, making corrections for grammar, punctuation, and spelling. For example, "Proofread and correct any errors in the following text:".

Expanding

ChatGPT can also expand a short input text into a longer, more detailed piece of writing, such as a blog post, an essay, or an email reply. Here are some things to consider:

Provide context: If the output is being used in a specific context, make sure to provide that context in the prompt. For instance, "Write a reply to the following email, expressing appreciation for the sender's suggestions and agreeing to implement them:".
Disclose AI involvement: If the output is being communicated to a user, it's recommended to disclose that it's AI-generated to maintain transparency.
Adjust the temperature: The temperature parameter controls the model's creativity. A lower temperature (closer to 0) makes the model's output more deterministic, while a higher temperature (closer to 1) allows for more creative responses.

Developing a Chatbot

You can use ChatGPT to create a chatbot with specific behaviors:

Set the role: Use the 'system' role to frame the conversation. This message tells the model how it should behave in response to the 'user' messages. For example, "You are a helpful assistant that always provides detailed responses:".
Provide user context: Include the user's name, details, and any other relevant context in the initial user message. For example, "The user is a beginner programmer asking for help with Python syntax errors:".
Remember the model's limitations: The ChatGPT model cannot remember earlier interactions, so you have to provide previous messages each time you make a request. For example, if the user asked a question in a previous message, include that message in the next request if it's relevant to the ongoing conversation.
Control the temperature: For output that is intended to be presented to the user, you can use a higher temperature for less predictable responses. For output intended to be parsed programmatically, use a lower temperature for more reliable results.

Understanding these prompt engineering concepts can significantly improve your interactions with the ChatGPT API, making your applications more effective and user-friendly.

The purpose of this blog is to give you a glimpse of the course "ChatGPT prompt engineering for developers". The course was packed with useful insights and techniques for anyone working with the ChatGPT API.

Here is the link to the course: https://www.deeplearning.ai/short-courses/chatgpt-prompt-engineering-for-developers/

I hope you find this valuable, if did awesome 👌 share it with your folks if it's relevant to them. If you have any suggestions/comments please feel free.

Happy Coding!