Body
Good Practices for Effective Prompts
A good place to begin experimenting with generative AI in your work is with this four-step process. As you learn more, you can refine the process to get the most out of your AI interactions.
Use the C.A.R.E. Model to Construct High-Precision Prompts
- Context: Be specific about your role and the audience for your work. The more context it has, the better it can identify appropriate content and tone.
- Action: Be as specific as you can about the exact task you want the AI to perform, and the type of output you want. The more specific you are, the better it can fit output to your needs.
- Restrictions: Tell the AI what you want it to avoid. That includes types of writing (like excess use of em dashes), types of output (no tables), and types of content (no data from a particular source).
- Examples: Provide a template or similar prior documents to help the AI gain a clearer view of what you want.
Preserve Privacy and Confidentiality
Use the redaction protocol outlined on the AI at Haverford website:
- Make a working copy of the document.
- Read deliberately and mark confidential information.
- For each piece of confidential information you find, you may:
- Remove or delete it.
- Redact and replace with a placeholder.
- Do a second read-through and ask yourself: could the remaining data used to reconstruct deleted information? If so, redact further.
- Check the metadata to remove confidential information in author names, revision histories, and comments in their file properties.
Iterate and Chain Your Prompts
Chatbots produce better results when you iterate and revise, and when you break complex tasks down into sequential prompts. For example:
- Turn 1: Find me 10 sources from scholarly journals and industry publications that will extend my knowledge.
- Turn 2: Based on what you found, create an outline that prioritizes [specific key ideas from the texts].
- Turn 3: Revise the outline based on [list of changes you want to make].
Validate Your Results
You are accountable for all AI-generated outputs. Take this two-step approach to validating your work:
- Ask in a fresh chat: For example, “Here is a copy of a spreadsheet. Please take on the role of an expert in Google Sheets and double check every formula for errors. Make me a list of each error and how I can fix it.”
- Check the work yourself! Your expertise and professional judgement are indispensable in creating high-quality output. Edit output for citation accuracy, voice and tone, and knowledge of your specific workplace context.
Examples of Effective Prompts
Gaining Fresh Perspective
I am a faculty administrative assistant writing a welcome email to new faculty hires who are coming into our department this fall. Attached is a copy of the email. Stage a dialog between three recipients of the email: 1) a new faculty member coming in straight out of their PhD program; 2) a senior professor who has held several posts at other colleges, but none this small; 3) and the provost of the College. In four rounds of back-and-forth, have them discuss what they found helpful about the letter, what they would have wanted to change or add, and how they felt after reading it. Make sure their reactions are realistic and not just what I want to hear. Then read back through the dialog and make a list of changes I should make, prioritized by their likely impact.
Talking an Idea Through
I am working on an idea for a new student tracking system that will help faculty in our department better calibrate their teaching by understanding the academic backgrounds of students taking their courses. I would like it to be able to produce basic demographic breakdowns: by major, year, etc. But I’d also like it to give me more granular insights. For example, computer science majors, stats majors, and physics majors may all have a background in computer programming. That would be good to know. Students majoring in history, English, and a foreign language may all have special expertise in the close reading of texts. I am attaching a list of our enrollments for the last three years with all the personal, identifying information removed. Read through the data we have then 1) suggest a data structure that could help me see these connections; and 2) in non-technical prose, suggest a way to build a system that does this without buying expensive new software.
Cross-Referencing Data
I am leading a committee that needs to make a recommendation about the contents of our introductory survey course. I have attached a national report on outcomes for students coming out of their freshman year, as well as syllabi and evaluation data for the past three years. Read the full text of the syllabi and the evaluations, and cross-reference them against the national report, to identify gaps in the course. Output a table that lists each gap, how important it is to address in terms of student success in later courses (high, medium, or low), and a 1-sentence explanation of how you identified the gap. Then recommend three changes to the course that will make the most difference.
Before using generative AI in your work, please consult the College policy on administrative use, as well as the data classification guidelines.