Data and AI Training

Home | Power BI | Excel | Python | SQL | Generative AI | Visualising Data | Analysing Data

GenAI Prompt Guidelines - Tips to get better responses from AI tools

We talk with AI tools by providing a question or instruction (called a prompt). The AI tool responds and then we can prompt again and the AI tool will respond again – and so on. Often the AI tool user interface may suggest a few example prompts.

A basic instruction alone may not result in a good response. We can help AI tools help us with more relevant, accurate responses if we provide more tailored prompts. Here are a few pointers to obtain better responses from AI tools.

• Make the instruction more specific.
• Tell the AI tool what role you want it to play.
• Provide context and relevant background.
• Describe any constraints.
• Provide an example of the response required.
• Refine your prompt and continue the conversation.
• Create a new chat and ask a new question on a different topic.

A novel technique is to turn the tables on the AI tool. Ask it to help you with your prompting

Can you give me some ideas of good prompts on the topic of …

It is often useful to know when the model was trained since it will not know about more recent events unless it specifically searches the web. One option is to ask it directly.

What is your knowledge cut-off date?

You may want run lots of experiments: test different prompts or different variations of the same prompt.

Ask the AI tool to adopt a persona

Ask the AI tool to adopt a persona or act in a role at the start of a conversation. For example, the initial prompt for a finance-related conversation could be

Act as an accountant with 10 years experience in the finance sector who has an ACCA qualification

Or as a software development-related conversation

Act as a software development manager with 20 years experience in the technology sector who has an Scrum certification

Or as a project management-related conversation

Act as a project manager with 30 years experience in the health sector who has a PRINCE2 qualification

Chain of thought reasoning

Some models now have chain-of-thought reasoning. Even if your model is not one of these, these prompts may encourage it to give a more considered or thoughtful response.

Think step-by-step
Please take your time and report back once you have considered thoroughly

Reverse the roles (flipped interactions)

We can ask the AI tool to elicit the information from us in a series of targeted questions, for example:

I am going to Rome for a holiday. Can you ask me a series of question to find out my preferences. Stop after each question and wait for my answer. Take that answer into account in your subsequent questions. Once you have enough information from me, provide your recommendations.

OpenAI’s advice on prompt design

OpenAI provides some comprehensive advice and suggests six strategies for better results here

Google’s advice on prompt design

Google provides some prompting 101 advice and examples here. In short

1. Give clear instructions: Help the model understand your request.
2. Define the task: Be specific about the desired output.
3. Specify constraints: Set limits on length, format, or style.
4. Define the response format: Tell the model how to structure its output.

They provide some examples in their prompt gallery here.

The RICE Framework

One framework to help us with this is RICE: this stands for Role, Instructions, Context (and Constraints) and Examples.

• Role: ask the AI tool to act in a certain role e.g., an expert travel advisor, and tell it how to respond e.g. brief or verbose, friendly or formal. Also, for example, tell the AI tool to provide a link to a relevant web page with each response.
• Instructions: provide specific, concrete instructions to the AI tool. (This is the most important piece of advice.)
• Context (and constraints): provide background information that will help the AI tool tailor its response.
• Example: provide an example of the sort of response that we would like to see.