7 ChatGPT prompting strategies recommended by OpenAI
Last modified
With the recent popularity of conversational AI models such as ChatGPT, it has become important to write effective prompts to get the results users want. OpenAI suggests seven prompting strategies to get the most out of these conversational AI models.
1.
Write clear instructions
It should be concise but contain the necessary details so that the model can accurately understand the user's requirements. It is helpful to use delimiters to clearly separate each part of the input, to describe specifically the steps the model should perform, and to specify the desired output length.
Example: "Please write a summary in English of approximately 200 characters based on the following content: [original content]"
It should be concise but contain the necessary details so that the model can accurately understand the user's requirements. It is helpful to use delimiters to clearly separate each part of the input, to describe specifically the steps the model should perform, and to specify the desired output length.
Example: "Please write a summary in English of approximately 200 characters based on the following content: [original content]"
2.
Provide reference text
Language models generate answers based on given text, so providing reference text can help the model provide more accurate answers, especially when asking about difficult topics or URLs.
Example: "Read the article below and explain how climate change affects agriculture. [article URL]"
Language models generate answers based on given text, so providing reference text can help the model provide more accurate answers, especially when asking about difficult topics or URLs.
Example: "Read the article below and explain how climate change affects agriculture. [article URL]"
3.
Divide complex tasks into subtasks
If you ask the model to process a complex task at once, it may show a high error rate. Therefore, it is better to divide the complex task into several simpler subtasks and synthesize the results of each subtask.
Example: "Please write a recipe using the following steps: 1) List the ingredients needed, 2) Describe the cooking process in order, 3) Suggest ways to plate the finished dish."
If you ask the model to process a complex task at once, it may show a high error rate. Therefore, it is better to divide the complex task into several simpler subtasks and synthesize the results of each subtask.
Example: "Please write a recipe using the following steps: 1) List the ingredients needed, 2) Describe the cooking process in order, 3) Suggest ways to plate the finished dish."
4.
Give your model “time to think”
AI models like ChatGPT also need time to think. If you give them about 17 to 28 seconds, the model can infer more accurate answers on its own.
Example: "Please think deeply about the problem presented and suggest a possible solution after 30 seconds."
AI models like ChatGPT also need time to think. If you give them about 17 to 28 seconds, the model can infer more accurate answers on its own.
Example: "Please think deeply about the problem presented and suggest a possible solution after 30 seconds."
5.
Request a description of the model's inference process
You can understand the reasoning process by asking the model to explain for itself how it reached its conclusions. You can use the internal dialogue technique, or you can start by asking for a summary of the user's query and then gradually build up to a longer conversation.
Example: "Please analyze the given data and explain the analysis process and results in detail step by step."
You can understand the reasoning process by asking the model to explain for itself how it reached its conclusions. You can use the internal dialogue technique, or you can start by asking for a summary of the user's query and then gradually build up to a longer conversation.
Example: "Please analyze the given data and explain the analysis process and results in detail step by step."
6.
Using external tools
To complement the model’s performance, you can leverage text search tools (RAG, search alignment generation, etc.). OpenAI’s Code Interpreter can also help your model perform math/code-related tasks more accurately.
Example: "Run the Python code below and explain how it works and what the results are. [Python code]"
To complement the model’s performance, you can leverage text search tools (RAG, search alignment generation, etc.). OpenAI’s Code Interpreter can also help your model perform math/code-related tasks more accurately.
Example: "Run the Python code below and explain how it works and what the results are. [Python code]"
7.
Systematically testing changes
When modifying prompts to improve model performance, you should define a comprehensive test suite and establish performance evaluation criteria. Keep in mind that while performance may improve for some examples, it may degrade for others.
Example: "Please compare and analyze the performance of the old prompt and the new prompt using various examples."
When modifying prompts to improve model performance, you should define a comprehensive test suite and establish performance evaluation criteria. Keep in mind that while performance may improve for some examples, it may degrade for others.
Example: "Please compare and analyze the performance of the old prompt and the new prompt using various examples."
If you use ChatGPT with reference to the above 7 strategies and examples, you will be able to get more effective and accurate results. The key is to clearly convey the user's intention, maximize the model's strengths, and find ways to complement its limitations. I hope this article will be helpful to those who are having difficulty using ChatGPT.
Subscribe to '오늘배움'
Get the latest on education trends, AI-based learning know-how, and corporate talent development insights at Learn Today. Receive useful news as we explore the future of digital learning together.
