Prompt Generation
The “P” in ChatGPT stands for pretrained. This means ChatGPT, and other LLMs, are not using a keyword search method to search relevant literature for information that best answers your question, like a librarian or researcher might do. Rather, they are pretrained in human language patterns which they use to string together a response designed to sound human. ChatGPT’s answers are also limited to the information it is trained on. It gets its information from the internet (pre-2021), and also learns from user prompts.
The prompts you create impact the answers you get. The more specific and detailed your prompts, the more likely you are to get a precise and useful answer. Here are some tips and resources to help you create effective prompts in ChatGPT or other LLMs.
Tips
- Provide detail and background information
-
- Instead of simply asking, “How do I add numbers in Excel?” Try adding more detail, “I am a graduate student at a state university and I am creating a budget for a two year grant for an on campus composting system for community food waste. How do I create a two column spreadsheet in Excel for supply expenses and future income. Each column should add up automatically with totals appearing at the bottom right of each column." (ChatGPT 4.0 will even create a file of the spreadsheet for you to edit.)
- Use clear language and break down complex requests into smaller ones
- Instead of asking the model to write an entire cover letter, break the request down into smaller prompts, such as, "Write a cover letter introduction paragraph with a professional introduction for the following job..." and input the specific job description. (You can also include an example cover letter that you want the model to emulate and give it information about your previous work experience and credentials as well as the tone you would like the model to write in, more on that later.)
- Be sure to use clear language too. For example, rather than asking, "What words are similar to scholarly?" ask, "What are five synonyms for the word 'scholarly'?". If you use the term 'similar' instead of 'synonym', the model may look for words that sound similar rather than words that share the same meaning.
- Refine your prompts
- Try using different wording or iterations of the same questions.
- You can also ask follow up questions or assign additional tasks in a separate prompt. The model will remember your previous prompts and reply accordingly. For example, if you first asked the model, "What is the habitat, food requirements, and breeding habits of a goldfinch?" it will respond with basic information about goldfinches. You can then ask it a more specific question related to this such as, "Based on this information, what are the best places in the state of Wisconsin to watch for goldfinches."
- Identify your audience
- One example of this is asking the model to summarize or rephrase complex information for a different reading level. For example, "Summarize the following article for an 8th grade reader," and copy and paste the article in the chat box. By identifying your specific audience the model is better able to complete the task.
- You can also ask it to write emails or give feedback to a specific person or audience, but be careful about how much personal information you give to LLMs. Remember your prompts are saved and used to help the model learn. Instead, consider asking it to write for a specific demographic or profession the person falls under. Then, you can always edit and personalize it more later, outside the model. (Also, be aware of potential bias or lack of information the model may have about certain demographics such as minorities and women.)
- Include examples
- You can and should include examples of things you want the model to follow. For example, you can say something like, "Based on this example undergraduate syllabus, write a new syllabus for a graduate level course on the same topic, and include the following statement about the use of generative AI." (and, insert your statement)
- You can even ask the model to emulate tone or voice by copy and pasting links to articles or blog posts of an author or researcher you want it to write in the style of. (Keep in mind however, you are responsible for how you use, share, and edit the content the model creates. To avoid plagiarism, be sure to give credit where credit is due.)
