Practical Prompt Processes
There was a moment in this week’s AI Literacy Knowledge Café where it became clear: Many people are still treating AI like search tool.
It’s a useful starting point. But it does explain why results can feel hit-and-miss.
Lots of people have tried to convert the PICO (Problem Intervention Comparison Outcome) Framework into searching and found it isn’t very effective. We discussed prompt frameworks at the knowledge café. Those that had tried some found that it was a good starting place for teaching good practice but soon they found themselves organically developing their prompt techniques.
Do we need a framework?
The useful thing about prompting is you can tell the Generative AI model who to be, so you can analysis the content and tailor it to the intended audience or generate new insights from developing different personas to use in testing.
AI ROCKS is a good framework to start on. It helps to provide clarity to your instructions. This is good for safer responses and the precision is good for the environment, too.
· Role – Give the AI a persona of who is doing this work
· Objective – what are you wanting it to do
· Community – who is the intended audience
· Key – What is the style and tone, also think about filters, e.g. UK English
· Output – What are you asking it to create
Whilst a useful tool for providing a good level of skill, the best way to prompt is through prompt chaining. This helps to reduce processing power required by AI, you can also see how it evolves and stop it going astray. Just like a database search, you need to break your prompt into separate lines, asking it to process one thing at a time, for example draft, analyse, tailor, refine.
There are several services creating prompt libraries and commercial ones are now available to help people get used to prompting. I’ve started to add into everyday library products; you may have noticed a prompt library appear in the Awards Bulletin that I co-produce with Sinead Stringwell.
In Higher Education, there is a move away from prompt libraries to tailor custom Generative Pretraining Transformers (GPTs) which can be shared across the organisation. Others are learning effective prompting to really get to grips with Agentic AI.
What was also noted was a shift from telling AI what to do ‘push prompting’ and asking it to help refine ‘pull prompting’. Asking AI to how to ‘improve this prompt’, ‘highlight inconsistencies’, ‘deduplicate’ and ‘what is missing’ are powerful tools in building your skills and getting more accurate returns.
However, you are using prompting, one thing we need to do more is adding guardrails. We ask AI what we want it to do, but we need to remember to set the boundaries, too.
Think of it like exclusion criteria in research of filtering a search. This is useful for personalising your Copilot experience: Did you know you can go into your Copilot settings and give it custom instructions?
…This is mine…
Do not use icons in results unless requested. Provide citations. Don’t speculate. Include ideas I am missing. Suggest guardrails. Recommend how to improve this prompt. Use UK English and provide explanation of acronyms with acronyms in brackets.
There are plenty more you can add: Tailor it to your style.
Even with good prompts, chaining, and guardrails, outputs are not guaranteed to be correct. Verification remains essential, especially in clinical or evidence-based contexts. Always have a human in the loop to check the content and make sure you cite your use of AI.
This article was created from a summary of LIS-AI-Literacy meeting notes on prompt frameworks created 22/05/26. Using the notes the prompt was “draft a blog post in the style of https://northernlightsblog.net/“). This was modified to focus the article and use only examples supplied by the author.
To learn more on AI Literacy, join the LIS-AI-Literacy Mailing List to attend the Knowledge Cafes. Previous notes can be found on the Current & Emerging Tech Group or in the mailing list archive.
Susan Smith
Knowledge & Library Manager
Mid Cheshire Hospitals NHS Foundation Trust
