Mastering Effective Prompting Techniques: Andrew NG's Guidelines
Table of Contents
- Introduction
- Principles for Effective Prompting
- 2.1 Write Clear and Specific Instructions
- 2.2 Give the Model Time to Think
- Examples and Tactics for Effective Prompting
- 3.1 Using Delimiters for Clear Instructions
- 3.2 Asking for Structured Output
- 3.3 Checking Whether Conditions are Satisfied
- 3.4 Using Few-Shot Prompting
- 3.5 Specifying Steps for Task Completion
- 3.6 Instructing the Model to Reason Out Its Solution
- Limitations of Language Models
- 4.1 Hallucinations and Fabricated Ideas
- 4.2 Reducing Hallucinations with Relevant Quotes
- Conclusion
Introduction
In this article, we will explore the effective techniques and principles of prompting to guide you in achieving the desired results when working with language models. We will discuss two key principles that play a crucial role in prompt engineering: writing clear and specific instructions, and giving the model sufficient time to think before producing a response. By understanding these principles and implementing the tactics outlined in this article, you will be able to prompt language models effectively and obtain accurate and relevant outputs.
Principles for Effective Prompting
2.1 Write Clear and Specific Instructions
To guide the model towards the desired output, it is essential to provide clear and specific instructions. Longer prompts often provide more Clarity and context, leading to more detailed and relevant outputs. A helpful tactic for clear instructions is to use delimiters, such as triple backticks or quotes, to indicate distinct parts of the input. Delimiters can be any clear punctuation that separates specific pieces of text. By utilizing delimiters, you can avoid prompt injections, where conflicting instructions from the user may affect the model's output. Asking for a structured output format, such as HTML or JSON, can also facilitate easier processing of model outputs.
2.2 Give the Model Time to Think
Rushing the model to a conclusion without allowing it adequate time to think can result in reasoning errors. By reframing the query and requesting relevant reasoning chains or series of steps, you can encourage the model to arrive at a correct solution. Specifying the steps required to complete a task can provide additional clarity and help the model understand the desired actions. Instructing the model to work out its own solution before rushing to a conclusion can also improve accuracy. By allowing the model to spend more computational effort and time on a task, you can mitigate errors caused by incomplete reasoning.
Examples and Tactics for Effective Prompting
3.1 Using Delimiters for Clear Instructions
One effective tactic for clear instructions is the use of delimiters. These delimiters can be any clear punctuation or tags that separate specific parts of the text from the Prompt. By enclosing the text to be summarized or processed within delimiters, you can provide clear indications to the model. Delimiters help the model understand the exact text it should focus on, reducing the chance of irrelevant or incorrect responses. Delimiters also serve as a technique to avoid prompt injections, where conflicting instructions from the user may misguide the model.
Example Prompt:
Summarize the following text delimited by triple backticks into a single sentence:
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Donec vitae eros eu urna rutrum luctus. Nunc egestas odio vitae tortor hendrerit eleifend.```
By utilizing delimiters, the model can accurately summarize the specified text, improving the relevance and specificity of the output.
### 3.2 Asking for Structured Output
Specifying the desired output format can assist in easier processing and interpretation of the model's responses. For example, requesting the output in JSON or HTML format allows for seamless integration into Python dictionaries or web applications. Furthermore, providing clear instructions on the required structure and keys in the output object can ensure uniformity and ease of handling the model's responses.
#### Example Prompt:
Generate a list of three made-up book titles, along with their authors and genres. Provide the information in JSON format with the following keys: book ID, title, author, and genre. Separate each entry with line breaks.
By asking for the output in a structured format like JSON, you can easily parse the model's response into a dictionary or list, simplifying further processing.
### 3.3 Checking Whether Conditions are Satisfied
When a task relies on certain conditions being met, instructing the model to check and indicate if those conditions are satisfied can prevent incorrect or suboptimal outputs. By explicitly asking the model to verify the validity of assumptions and consider potential edge cases, you can avoid unexpected errors or results. This tactic ensures that the model does not attempt to complete a task without the necessary prerequisites, enhancing the accuracy and relevance of the output.
#### Example Prompt:
You will be provided with text delimited by triple quotes. If the text contains a sequence of instructions, rewrite those instructions in the following format. If no instructions are found, write "No steps provided."
Today's tasks:
1. Preheat the oven to 180°C.
2. In a mixing bowl, combine flour, sugar, and salt.
3. Beat the eggs in a separate bowl.
4. Gradually add the eggs to the dry mixture and mix until smooth.```
Using this prompt, the model is instructed to identify and extract the steps or instructions from the given text. If no instructions are found, it should respond with "No steps provided."
### 3.4 Using Few-Shot Prompting
Few-shot prompting involves providing examples of successful executions of the desired task to guide the model. By showcasing examples before instructing the model to perform the actual task, you can familiarize the model with the desired style, tone, or format. Few-shot prompting enhances the consistency and accuracy of the model's responses, enabling more coherent and contextually appropriate outputs.
#### Example Prompt:
Inconsistent Style Fixing:
Child: Teach me about patience.
Grandparent: "Patience is like a tree that bends with the wind but never breaks."
Please answer the following questions using a consistent style:
- Teach me about resilience.
- Describe the concept of equality.
By establishing a consistent style through examples, the model can respond in a manner that aligns with the desired tone and content.
3.5 Specifying Steps for Task Completion
To guide the model effectively, it is crucial to specify the necessary steps required to complete a task. By breaking down complex tasks into comprehensible and distinct instructions, you provide the model with a clear action plan. This enables the model to produce accurate and relevant outputs, reducing errors and misunderstandings.
Example Prompt:
Perform the following actions:
1. Summarize the following text delimited by triple backticks into one sentence.
2. Translate the summary into French.
3. List each name in the French summary.
4. Output a JSON object with the following keys: French Summary, Num Names. Separate the answers with line breaks.
Text to summarize:
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Donec vitae eros eu urna rutrum luctus. Nunc egestas odio vitae tortor hendrerit eleifend.```
Specifying the steps required for task completion allows the model to understand the desired actions and generate outputs that align with the given instructions.
### 3.6 Instructing the Model to Reason Out Its Solution
Sometimes, instructing the model to reason out its own solution before reaching a conclusion can lead to more accurate and well-reasoned outputs. By asking the model to work out the problem or provide a series of relevant reasoning steps, you encourage it to think critically and evaluate different possibilities. This tactic enhances the accuracy and logical consistency of the model's responses.
#### Example Prompt:
To determine if the student's solution is correct, please follow these steps:
- Work out your own solution to the problem.
- Compare your solution to the student solution.
- Evaluate if the student's solution is correct.
By instructing the model to evaluate its solution and compare it to the student's solution, you encourage a more thoughtful and detailed response, minimizing errors caused by rushing to a conclusion.
Limitations of Language Models
4.1 Hallucinations and Fabricated Ideas
Despite their extensive training on vast amounts of data, language models can sometimes generate hallucinations, fabricating information that sounds plausible but is, in fact, incorrect. These fabricated ideas can lead to misleading or inaccurate outputs. It is crucial to be aware of this limitation and implement techniques to reduce hallucinations, such as asking the model to find relevant quotes from the text and using them to answer questions.
4.2 Reducing Hallucinations with Relevant Quotes
One effective approach to reduce hallucinations is to request the model to provide answers based on quotes from the text. By asking the model to use the provided quotes to support its responses, you can ensure that the generated outputs are grounded in the source material. This approach helps to maintain accuracy and reduces the likelihood of the model generating false or fabricated information.
Conclusion
In this article, we have explored the principles and tactics for effective prompting in language models. By writing clear and specific instructions, giving the model time to think, and implementing various techniques discussed, you can prompt models with improved accuracy and relevance. While language models have their limitations, being mindful of these limitations and employing appropriate strategies can mitigate errors and enhance the overall performance of the models.
Resources
FAQ
Q: Can I use shorter prompts instead of longer prompts for clarity?
A: While it may seem counterintuitive, longer prompts often provide more clarity and context for the model. They help guide the model towards the desired output by offering additional information or instructions. However, it is essential to balance the length of the prompt with the simplicity of the task to avoid overwhelming the model.
Q: How can I prevent prompt injections from affecting the model's output?
A: Delimiters and clear instructions play a crucial role in avoiding prompt injections. By utilizing delimiters to separate specific parts of the text and providing explicit instructions, you can guide the model towards the desired output, irrespective of any conflicting input from the user. Delimiters act as signposts for the model, ensuring it understands the intended task.
Q: What do I do if the model responds with hallucinations or fabricated information?
A: Hallucinations and fabricated ideas are known limitations of language models. To mitigate this, you can ask the model to provide answers based on relevant quotes from the text to ground its responses in accurate information. This helps reduce the likelihood of the model generating false or implausible outputs.
Q: How can I effectively specify steps for task completion?
A: Clearly outlining the steps required for task completion helps the model understand and execute the desired actions accurately. Break down complex tasks into manageable steps, ensuring each instruction is explicit and unambiguous. By providing clear and sequential instructions, you guide the model towards producing outputs aligned with the task's requirements.
Q: How can I encourage the model to reason out its solution before reaching a conclusion?
A: To encourage the model to reason out its solution, explicitly instruct it to work through the problem step by step before arriving at a final answer. By asking the model to provide reasoning chains or relevant intermediate steps, you can enhance the accuracy and logical consistency of its responses.
