flowchart LR
A["Define the task"] --> B["Get a first structured answer"]
B --> C["Interrogate assumptions or gaps"]
C --> D["Refine format for use"]
D --> E["Verify against source or data"]
🔧 The emphasis is still on practical exercises; short teaching segments provide prompt structure, validation habits and workflow ideas to support each activity.
This workshop builds on the principles introduced in the last session. Responsible AI use requires transparency, verification, documentation, equity and continuous monitoring. Participants will practise constructing clear and specific prompts and learn to record prompts and outputs for transparency. Remember not to include confidential or proprietary information in prompts.
A good policy when writing research prompt is to usually include:
âť“ A useful extra instruction is to flag uncertainty rather than guessing.
Participants are welcome to form up into groups or work alone
https://m365.cloud.microsoft/chat/
Formulate a simple prompt such as:
📝 “Summarise the University of Nottingham guidance on generative AI for researchers in six bullet points. Use only the linked guidance page. Flag any area that remains ambiguous or context-dependent.”
Each group or individual will investigate one AI policy area:
👤 We will examine different policy domains to compare how organisations are responding to AI.
Using Copilot, collect information about your assigned policy area. Document the outputs in a shared sheet or board.
Groups should prepare to:
Groups present initial findings and discuss prompt strategies and observed weaknesses in AI responses.
A safer workflow is:
🔎 Use AI after you have the source in hand: let it help interpret official documents, not replace reading, checking, and defending the evidence yourself.
These are the official policy documents:
⚖️ Basically dont trust the AI summary at face value, test it against the official policy text and see where it is accurate, incomplete, or misleading.
| Output | What to check |
|---|---|
| summary or explanation | source accuracy, date, missing caveats |
| quoted rule or citation | exact wording, existence, relevance |
| interpretation | scope, ambiguity, whether it follows from the source |
| agent output | consistency, gaps and whether checks were actually applied |
âś… Validation should be built into the workflow, not bolted on afterwards.
flowchart LR
A["Define the task"] --> B["Get a first structured answer"]
B --> C["Interrogate assumptions or gaps"]
C --> D["Refine format for use"]
D --> E["Verify against source or data"]
🧠A staged prompt chain is less about cleverness than control: it helps separate drafting, critique, refinement, and verification so the final output is more reliable and easier to defend.
Use Copilot to compair its initial claims with the actual policy text.
For at least two AI-generated statements, verify:
🔍 Hint Note possible reasons for inaccuracies: outdated training data, vague prompts, or ambiguous policy language.
The tone and content of Copilot’s output can vary depending on the role specified. Asking Copilot to answer “as a researcher”, “as a policy adviser” or “as a science journalist” can change the depth, formality and focus of the response. Prompt engineering strategies emphasise that the quality of outputs depends on the clarity and context provided.
🎠Role prompts can be useful for shaping tone and audience, but it should still not be confused with expertise, evidence, or authority.
Re-ask the same policy question used in Task 1 with three different role instructions.
🧾 Record differences in tone, structure and certainty.
Good candidates for a reusable workflow have:
🧠If you are doing the same low-risk task repeatedly, ad hoc prompting becomes inefficient.
AI agents can automate a repeatable, multi-step workflow for policy review.
Instead of prompting for each step separately, the agent can be configured to:
âś… Once this workflow is defined, you only need to provide the topic or organisation and the agent should cary out the process consistently each time.
| Good agent candidates | Poor agent candidates |
|---|---|
| policy lookup from public websites | final interpretation of research findings |
| turning repeated notes into a consistent template | authorship or disclosure decisions |
| monitoring public guidance updates | any task based on unclear or sensitive evidence |
🤖 Agents work best when the task is narrow and easy to check.
A short record after each substantial use can include:
If you are using an agent you can ask it to include this information!
Create an AI agent configured to:
Test it on a few organisations.
Note where it succeeds, where it struggles and where human oversight is required. Keep a brief note of the agent instructions and the checks you applied.