Prompt Engineering: Strategies and Techniques to Optimize Outputs from Large Language Models

This guide shares strategies and techniques for obtaining better results from large language models (such as GPT-4). These methods can be used individually or combined to achieve better results.

Six Strategies for Better Results

1. Write Clear Instructions

   Explanation: Clear instructions help the model understand your needs more accurately, providing more relevant answers.

   • Include detailed information for relevant answers Example: Instead of asking, “How to add numbers in Excel?”, ask, “How to automatically sum a whole row of dollar amounts in Excel and place the total in a column labeled ‘Total’?”

   • Request the model to adopt a specific role Example: “When I ask for writing help, respond as a humorous writer, including at least one joke or quip per paragraph.”

   • Use delimiters to clearly distinguish different parts of the input Example: Use triple quotes, XML tags, or section headers to separate different parts of the text.

   • Specify the steps required to complete the task Example: “Please follow these steps to answer user input: Step 1 - Summarize the text; Step 2 - Translate the summary into Spanish.”

   • Provide examples Example: Give examples of questions and answers to show the model the style of response you expect.

   • Specify the desired output length Example: “Please summarize the following text in approximately 50 words.” or “Please summarize the following text in 3 bullet points.”

2. Provide Reference Texts

   Explanation: Providing relevant references can help the model generate more accurate and well-founded answers.

   • Instruct the model to use reference texts to answer Example: “Use the provided article (separated by triple quotes) to answer the questions. If you can’t find the answer in the article, write ‘I can’t find the answer.’”

   • Instruct the model to cite content from the reference text Example: “When answering questions, please quote relevant paragraphs from the provided document. Use the following format for citations: {‘Quote’: …}”

3. Break Down Complex Tasks into Simple Subtasks

   Explanation: Breaking down complex tasks can improve accuracy and clarify the process.

   • Use intent classification to identify the most relevant instruction for user queries Example: First classify customer service queries (such as billing, technical support, etc.), then provide appropriate handling instructions based on the classification.

   • For applications requiring very long conversations, summarize or filter previous conversations Example: Regularly summarize conversation history, using the summaries as new context to maintain conversation coherence.

   • Summarize long documents in sections, then recursively build a complete summary Example: Divide a long document into several parts, summarize each part, and then summarize these summaries to get a concise overview of the entire document.

4. Give the Model Time to Think

   Explanation: Allowing the model to think more deeply can produce more accurate and comprehensive answers.

   • Instruct the model to formulate its own solution before reaching a conclusion Example: “Before evaluating the student’s solution, please solve this problem yourself.”

   • Use inner monologue or a series of queries to hide the model’s reasoning process Example: Require the model to place its reasoning process in a specific format (such as within triple quotes) so it can be filtered out later.

   • Ask the model if it missed any previously passed content Example: “Did you miss any relevant excerpts? Please check carefully to ensure you don’t repeat already mentioned content.”

5. Use External Tools

   Explanation: Combining external tools can compensate for the model’s weaknesses, improving the accuracy and practicality of the answers.

   • Use embedding-based search for efficient knowledge retrieval Example: Use text embeddings to retrieve the most relevant document fragments for the query, then provide this information to the model.

   • Use code execution for more accurate calculations or calling external APIs Example: “You can write and execute Python code using triple backticks. Use this feature to perform calculations.”

   • Allow the model to access specific functions Example: Provide API documentation or code examples so the model learns how to use specific external functions.

6. Systematically Test Changes

   Explanation: Systematic testing can help determine if changes truly improve model performance.

   • Evaluate model output against gold standard answers Example: Use a test set with known correct answers to compare the model’s output against the standard answers.

Conclusion

Prompt engineering is a key technique for optimizing outputs from large language models. By employing the six strategies and their corresponding techniques introduced in this guide, users can significantly improve their interactions with AI models:

1. Writing clear instructions helps the model understand needs more accurately, providing more relevant answers.
2. Providing reference texts enables the model to generate more accurate and well-founded answers, reducing the likelihood of fabricated information.
3. Breaking down complex tasks into simple subtasks enhances processing accuracy and clarity.
4. Giving the model time to think produces more in-depth and comprehensive answers.
5. Combining external tools can compensate for the model’s limitations, improving the accuracy and practicality of the answers.
6. Systematically testing changes ensures that optimizations genuinely improve model performance.

These strategies can be used individually or flexibly combined for the best results. As AI technology continues to evolve, the importance of prompt engineering will become increasingly apparent. Mastering these techniques will help users better utilize AI models, achieving more efficient and accurate human-machine collaboration.

However, it is important to note that prompt engineering is a process requiring constant practice and adjustment. Each specific application scenario may require different strategy combinations. Therefore, users are encouraged to remain innovative in practical applications, continuously optimizing prompt strategies based on specific needs and feedback to achieve the best AI-assisted results.

You can learn more about this topic in detail through this article.