ChatGPT and other Generative AI tools can create content for you. To get the best results, you need to use good prompts. Here you’ll find out how to write good prompts.
What is a prompt? What is prompting? What is prompt design?
In a nutshell, a prompt is what you type into ChatGPT (or most other Generative AI tools). It’s what you request the AI-chatbot to generate or do for you.
Giving Generative AI tools instructions is called prompting. Other terms commonly used are prompt design or prompt engineering. They indicate that there is more to a prompt than just a quick sentence.
Prompt Engineering – a skill of the future?
For some, prompt engineering is the skill of the future. The main argument for this perspective: prompt engineers know how to consistently generate the output they are looking for. There was even a job ad circulation for a position as a prompt engineer at Anthropic offering a salary in the range of 250k to 375k USD. Anthropic is another company that develops and provides a Large Language Model (its chatbot is called Claude btw).
As this is the most famous and one of few examples, it’s not unreasonable to argue, that publicity was one of the motivations of the mentioned job ad.
For others, prompting, prompt design or prompt engineering, however you call it, will become obsolete. The reasoning is, that Large Language Models and the chatbots that build on top of them develop so fast and will become so mature, that no special skills will be required. Everybody will be able to get the required results.
Buy or build good prompts?
As you surf the internet, chances are high, that you’ve come across articles or posts on blogs or social media about the perfect prompt for x or lists of prompts for different use cases. In addition, marketplaces have opened where you can purchase single prompts (quite often for a handful of dollars, in contrast to the high salary mentioned in the previously mentioned job ad).
The prompts you’ll find on these lists and marketplaces can be a good starting point to test for the use case you have in mind. They might also serve as a shortcut or source of inspiration. It’s very likely though, that they won’t help you achieve your goal the way you want. Alternatively, with a specialized tool like eggheads for chat-based employee training the prompts you need are already baked into the software, so that you don’t need to take care about the right prompt.
Compare it to researching for an image in a picture database. You start with a specific picture in your mind. Even though there are thousands of pictures in the database, scrolling through it for hours won’t surface a result that matches what you have in mind.
From our point of view, there is no such thing as the perfect prompt or only one prompt that delivers the desired result. It’s more helpful to understand what makes a good prompt.
Basics of a good prompt
ChatGPT and other AI-powered chatbots made it as simple as possible to access powerful AI technology – through chat. Chat is one of the easiest ways to interact with a computer. However, a useful prompt requires more than just a loose string of words. To get the best results, it’s essential to give it the right instructions. These are our tips for writing good prompts.
1. Start with the end in mind
What do I need? For whom? Why? These are strategic questions and the more precise you know what you want to get out of a generative AI tool, the higher the chances are, that you will get it. Or in other words: if you can’t put in words what you need, it will never exist. Clear and specific instructions are thus the most common tip. Sometimes, that leads to longer instructions. So clear doesn’t have to mean short.
2. Tell ChatGPT its role
Large Language Models were trained with vast amounts of text. That’s one of their strengths and at the same time one of their flaws. LLM comprise perspectives and biases that were reflected in the training material. In your prompt, you can tap into them and set the context for the task at hand. By framing the context, you can tap into the right knowledge base and get the desired output. For example: act as a seasoned Instructional Designer.
3. Give it a task
This seems quite obvious and that’s where most users start. By entering what ChatGPT should do. You could also say it’s your command, so typically you’ll start with a verb. For example: come up with ideas for a class warm up.
4. Chose the output format
ChatGPT and other AI-Chatbots can deliver the content they generate for you in the type of content and length you need. For instance, a poem, a news article, a video script that has the number of words you need.
5. Set the topic
What should the generated output be about? You can build on the knowledge that’s baked into ChatGPT or give it raw material that you already have. For example, about the periodic table, the history of philosophy, about the following text, the following terms or concepts etc.
6. Determine the tone and style
How should the output read? A rap has a different sound than an instruction manual, a pirate and a lawyer pick different words, a fifth-grader will articulate concepts differently than a scientist would. Tonality and style allow you to tailor your content to your audience or gain a different perspective on it.
7. Mention your audience
A beginner has different needs and questions than a professional, someone in marketing wants to know more about other aspects than someone from IT. When you let ChatGPT know who the intended recipients of your content are, the output is more likely to speak their language. Stating the audience can drastically improve the content’s quality and relevance.
Advanced Prompting Tips
The basics will help you come close to the output you are looking for. To improve the generated results even further, take advantage of the following prompting tips:
1. Provide examples
Give the bot a reference to work with by providing successful examples of completed prompts. For instance, if you want it to write similar slide headlines to the ones in your previous presentations, or if you are looking for a specific structure for your Twitter Threads, or if (as a developer) you want the output in JSON format. This is also called one-shot or few-shot learning.
2. Build clear sections with delimiters
As mentioned above, a precise prompt can get quite long. To make it clear to the model, where a certain section of the prompt ends and where another one starts, you can use delimiters. For example triple backticks “` or “”” or — or <> or xml tags
3. Let ChatGPT “think step by step” or “think aloud”
A proven technique to increase the accuracy of the output and to gain insights in parts of the reasoning behind it, add “let’s think step by step” or “think aloud” to the end of your prompt. This is called chain of thought prompting.
4. Break down big questions
Despite their many capabilities, ChatGPT & Co. might not generate helpful output if you ask it for too much. Sometimes it helps to divide a bigger prompt in chunks. For example, you first let it describe a certain method or model, then you use the output in the following prompt.
5. Keep iterating
If you’re looking for useful output, chances are high, you need more than one prompt to get it. Keep going, adjust and enhance your prompts, evaluate the different results until you’re happy.
Think strategically
While there is an ongoing debate about AI killing jobs, there is also a strong argument that subject matter expertise will still be required in the future. While it’s true that it now only needs a few words and sentences to let AI generate text, images, and computer code, users need to be able to think strategically to get the desired outcome.
If they’re not able to put in words what they want, AI won’t generate it. Or in other words: customers need to be able to write clear briefings. So there’s also some relief with marketers or software engineers for example.
To get the most out of ChatGPT and similar AI tools, what you input really matters. If you want better results, focus on improving your prompts. Prompting is a skill that can really make a difference.
Further Reading