Prompt Engineering
Key University AI Guidelines (Click to expand)
- No uploading copyrighted articles, book chapters, or subscription content.
- No uploading student work or personal data.
- AI hallucinations apply to all tools — verify everything.
- Use Library‑integrated AI tools for summarising academic articles.
- Use reference managers (EndNote, Mendeley, Zotero) — AI citations are unreliable.
- Gemini: connecting Gmail/Docs/Drive exposes personal data.
- Follow IMPS guidance for research and teaching.
Generative AI models are highly literal. They do not know what you want unless you explicitly tell them. Learning to structure your prompts allows you to reliably produce formatted outputs, reducing errors and saving you time.
The CREATE Framework
A simple prompt like "Write an email about the new IT policy" will yield a generic, likely unusable result. Instead, use a framework to build a robust set of instructions.
Context & Persona
Tell the AI exactly who it is acting as and who the audience is.
"Act as a senior IT communications officer. The audience is university staff who are not highly technical. We are rolling out a new mandatory cybersecurity module."
Request & Format
State clearly what it needs to produce and how it should look.
"Write a 200-word introductory email. Use a welcoming tone. Format the email with a clear subject line and use bullet points for the instructions."
When to Provide Examples
If you want the AI to write in a specific style or format data in a precise way, give it examples directly in the prompt. This drastically improves accuracy compared to relying on instructions alone.
- Tone matching: Paste in two previous emails you've written, and instruct: "Analyze the tone and vocabulary of these examples. Now, write a new piece matching this exact style."
- Data extraction: Show the AI exactly how you want a summary formatted (e.g., providing a blank template with headers) before asking it to process a new document.
Advanced Techniques
- Chain of Thought: Add the phrase "think step-by-step" to the end of a complex prompt. This forces the model to break down its logic before answering, drastically reducing logical errors in data formatting.
- Negative Constraints: Tell the AI what you don't want. For example: "Do not use marketing jargon, do not exceed 300 words, and do not use the word 'delve'."
- Iteration is Key: Treat the AI like a conversational partner. If the output isn't right, do not start over. Reply with a targeted correction: "Make the second paragraph less formal," or "Expand on bullet point three."
Fact-Checking and Hallucinations
Generative AI predicts the next logical word. Even with perfect prompt engineering, AI can occasionally present false information or fabricated citations as absolute fact. Always verify specific statistics, dates, and claims against trusted university sources.
These pages are currently in beta. To give feedback, please contact DTS at ai.training@reading.ac.uk.