RCT and 5P Prompt Configurations
Research Series on AI Benchmarking
This research note is part of a series on AI benchmarking and cross-pollinated methods. The note addresses different prompt configurations for role-context-task (RCT prompting) and prime-persona-privacy-product-polish (5P prompting) tasks.
I was first exposed to the basics of formalized prompt engineering at a great Cognit DX workshop in #Dubai on #AI in the workplace. It was targeted at white collar workers and designed to dispel anxiety over workplace disruption. Its focus was practical, introducing attendees to basic concepts, available tools and use cases, and the essentials of prompting.
Cognit-DX Managing Director Alaa Dalghan really drilled home the three essential elements of an effective AI prompt: role, context, task. Let's call it Alaa's RCT approach (TM). Here's what RCT prompting looks like.
Role: In your prompts, orient the AI by telling it what its role will be in this particular instance. Something like, say, "You are a lawyer specialising in AI #law and #regulation."
Context: Follow up with context. Something like "I'm developing an AI tool that does X. I think X is a completely new thing, but I don't know what laws and regulations potentially apply to it or what licenses or permissions I might need."
Task: This is what you want the AI to do - for example, "Please prepare a list of laws and regulations that apply to my innovation." Prompts that provide more detail, like how the task output should be organised (ie. in a table format, include dates, specify for a specific jurisdiction, and so on), yield better results.
That's it in a nutshell. Simple, straightforward, effective. This is, I guess, old hat among the developer crowd. I'd hazard that it's also pretty intuitive for any trained and experienced researcher, in any field. That's been my experience over the last few weeks, at least, while testing out #ChatGPT, #Claude, and a handful of other closed source models.
Some next-gen AI appears to be doing away with the need for all this prompt structuring and specificity. I think it's probably still a good idea for lay users of AI - the non-techie consumers and end users, not just the specialist engineering and tech teams that design and build these things - to think through and communicate effectively what they want their tools to do.
Which is what IMHO makes the content of Josh Kubicki's latest BrainyActs newsletter so useful. In it some of Josh's students outline the "5Ps Framework" for prompting, which is not the same as the 5P framework for marketing. This is what it involves in prompting terms: prime ("start with context"), persona ("set the tone and expertise"), privacy ("protect sensitive information"), product ("define your goal"), polish ("refine for precision").
There's some overlap between RCT and the 5PF, but I think they complement each other really nicely. A joined up RCT and 5PF approach starts with the simple efficiency of the former, then expands with the latter on how to think about each RCT element, and includes some additional points and how to think through them.
RESEARCH SERIES ON AI BENCHMARKING.
Exploration and discovery of AI, from a user perspective.
© Craighead Kellas 2024. All Rights Reserved.