0% found this document useful (0 votes)
2 views

Prompt guide 3

This document provides guidelines for writing effective prompts when using generative AI tools. It emphasizes the importance of clear communication and context to elicit useful responses, offering examples of questions and practices to enhance interactions with AI. Additionally, it suggests seeking advice from prompt engineers and utilizing prompt databases and apps to improve prompt quality.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
2 views

Prompt guide 3

This document provides guidelines for writing effective prompts when using generative AI tools. It emphasizes the importance of clear communication and context to elicit useful responses, offering examples of questions and practices to enhance interactions with AI. Additionally, it suggests seeking advice from prompt engineers and utilizing prompt databases and apps to improve prompt quality.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 11

HDR HELPFUL HINTS

From the Dean of Graduate Research, Simon Moss.

Effective prompts in AI

SUMMARY

To use generative AI tools effectively, you need to know how to write your questions or prompts
appropriately. This document presents some key guidelines to help you achieve this goal. The following
infographic summarises these guidelines. For more insights on how to use generative AI tools effectively,
you may even seek the advice of prompt engineers—individuals who offer advice on how to optimise the
questions.

Graduate Research School



University of Wollongong, NSW 2522 Australia [email protected]
In practice, you do not need to master all the principles in this document. Instead, learn to interact with Chat
GPT or other AI tools like you would interact with a smart robot—such as C3PO in Star Wars, Data in Star
Trek, TARS in Interstellar, and so forth. That is, do not treat AI tools like Google, but instead converse as
you would with a smart robot or even a fairly literal human or alien. Indeed, on many attributes, AI tools
resemble humans more than perhaps they resemble machines:

• AI tools and machines do not excel on the same tasks,


• For example, machines excel on executing routines consistently or performing complicated calculations.
• In contrast, AI tools excel more on tasks on which humans are proficient: coding, chatting, or even lying.
• In one famous example that Mollick (2024) reported, the tool, when told that AI does not feel emotions,
disagreed vehemently and even indicated that humans are merely cyborgs as well, because they use
computers to store information, phones to communicate, and so forth.

SAMPLE QUESTIONS

To use generative AI, such as Chat GPT, you need to enter a question together with some information. The
question, coupled with the information, is called a prompt. Here is an example of some typical questions
that HDR candidates may write. These examples revolve around one topic—the causes of diabetes—but
could readily be adapted to other topics.

TOPIC TYPICAL QUESTION


Clarify the • What are the main sources of controversy around the causes of diabetes—
research question controversies that researchers have yet to resolve?
and review the
• What have researchers identified as the main limitations in the research on
literature
the causes of diabetes?
• What are the main topics I should consider on the causes of diabetes?
• What keywords should I enter into databases, such as Google Scholar?
• And what are the most recent advances in research on the causes of
diabetes?
• Which academics are studying and publishing on the causes of diabetes?
• What theories, models, or frameworks could help me understand the
causes of diabetes?

2 | EFFECTIVE PROMPTS IN AI
Design the • Which qualitative methodologies could I consider to explore the causes of
methods diabetes?
• What are the benefits and drawbacks of these methodologies?
• Which tools can I use to measure the risk of diabetes?
• What is the cost of each measure in US dollars?
• Which measure is the quickest to administer?
• Which measure is the most valid or accurate?
• Can you suggest some publications I should read to support these claims?
• When applying these methods, what are the most common mistakes that
people commit? How can I avoid these mistakes?
• Which statistical technique should I use to determine whether age affects
the relationship between duration of exercise and incidence of diabetes?
• Can you provide an example to illustrate this technique?
• What software or tools should I use to complete this technique?
• What are some mistakes that researchers sometimes commit when they
complete this technique?
• What python or R code should I include to complete this technique?
Improve your • Here is an excerpt of my writing. What are some helpful words that I have
writing overlooked?
• Which words could be more precise?
• Which words or phrases could I delete or condense?
• What are some principles I should consider to improve my writing? Can
you illustrate these principles with reference to my paragraphs?

INFORMATION THAT PRECEDES QUESTIONS

If you enter these questions into an AI tool, like Chat GPT, the answers may not be especially informative.
That is, the answers may not be tailored to your needs and circumstances. To illustrate, if someone utters
“We should get together”, depending on the relationship, this sentence could imply

• this individual wants to visit the other person or


• this individual never wants to visit the other person.

3 | EFFECTIVE PROMPTS IN AI
To overcome this limitation, you should precede the question with some helpful information about your
needs and circumstances. This information diminishes the likelihood the AI tool will produce generic
answers that dismiss your specific needs. The following table outlines how to achieve this goal.

PRACTICE EXAMPLES
Clarify your circumstances and • Describe your goals, such as “I am writing a scientific
preferences as comprehensively report and want to write the report professionally and
as possible. concisely” to generate the right style of output.
• Outline your study.
• Clarify your challenges or limitations, such as “I need to
complete this thesis within 6 months” and “My supervisor
is too busy to assist at the moment”. These personal
challenges tend to promote more helpful responses.
• Describe your circumstances. “I study in Australia but am a
citizen of Indonesia”.

Highlight the importance of this • “Can you please answer this question carefully and
question to you comprehensively because this matter is important to my
career”.

Prompt the AI tool to assume or • “When answering this question, can you play the role of a
play a specific role. pedantic examiner?”
• “When answering this question, can you play the role of an
experienced researcher?”
• “When answering this question, can you play the role of a
health practitioner or manager in this field with expertise in
diabetes?”

Even subtle words can shift the responses of AI tool. For example

• if your prompt commences with the words “I want a serious debate about…”, the AI tool is likely to
respond defensively and sometimes aggressively.
• if your prompt commences with the words, “As a teacher, how do you feel about…”, the AI tool is likely
to respond more objectively and cooperatively.

Some generative AI tools enable you to choose one of several pre-existing personas or roles. For example,

4 | EFFECTIVE PROMPTS IN AI
a paid version of Chat GPT includes a feature called “my GTPs” that enables users to choose one of several
personas or roles the AI tool can adopt.

INFORMATION THAT FOLLOWS QUESTIONS

To generate suitable answers, you should also follow questions with helpful information. Specifically, you
should generally

• illustrate examples or features of the answers you are seeking to guide the tool,
• constrain the answers, such as the number of words.

The following table outlines how you illustrate examples and constrain answers effectively. Some of these
practices can significantly improve the answers you receive.

PRACTICE EXAMPLES
Illustrate potential answers
Enter examples of options you • “When suggesting topics to read, possible examples could
like—sometimes called few-shot include the impact of exercise on diabetes or the effects of
prompting (Min et al., 2022). relaxation on diabetes?”

Clarify why your examples are • “Here is an example of how you might complete this maths
valid (Wei et al., 2022). equation… The rationale behind the first step is…”

When users divide tasks into subtasks and explain the rationale
of each subtask—called chain-of-thought prompting—

• AI tools can answer complicated questions,


• AI tools can generate more diverse, novel answers (e.g.,
Meincke et al., 2024).

Prompt the AI to recall some For example, suppose you want to know whether meditation
relevant knowledge that could might affect back pain. Rather than merely ask “How does
be useful to answer a question, meditation affect back pain”, you might include the following
called generated knowledge prompts:
prompting (Liu et al., 2022)
• What are the main consequences of meditation?
• What are the main causes of backpain?
• How does meditation affect back pain?

5 | EFFECTIVE PROMPTS IN AI
Specify the features of answers • If you write “Do not include technical terms”, the reference
you want rather than do not to technical terms might, paradoxically, prompt some
want. generative AI tools to include technical terms in the
answer.
• Instead, you may write “Include only simple, everyday
language”.

Constrain the answers and


scope
Constrain the scope of your • Rather than “can you explain climate change?”, you could
question. write “can you explain how climate change may damage
the economy over the next 20 years?”

Encourage AI tools to answer as • “Can you ensure your answer is accurate. That is, can you
accurately as possible—a prevent hallucinations and report only valid information
prompt that seems to diminish and facts?”
the likelihood of inaccuracies.

Commence your question with • Apparently, this instruction has been shown to diminish the
the words “Please start your likelihood that Chat GPT will generate misleading or
answer with the phrase ‘My best limited responses.
guess is…’ and answer step-by-
step.”

Impose specific limits to the • “Can you confine this answer to 50 words”.
answer. • “Can you confine this answer in 20 bullet points”.
• “Tell me about the assumptions and limitations of this
theory” rather than “Tell me about this theory”.

For some AI tools, you can also • “Can you produce a list”?
specify the format of answers • “Can you produce a table to summarise the answers”?

Occasionally, ask the AI tool to • “As part of your answer, can you clarify how you interpret
demonstrate an understanding of this question. That is, what do you think I am asking?”
your question.

6 | EFFECTIVE PROMPTS IN AI
INFORMATION THAT FOLLOWS ANSWERS

Usually, you will not merely ask one question and then shift to the next task. Instead, you should generally

• divide each task into multiple subtasks and ask questions about each subtask in sequence,
• write politely, but not obsequiously or sarcastically, such as “Thank you for your answer. Can I ask
whether…”

Polite questions tend to promote more accurate and extensive answers (Yin et al., 2024). Nevertheless,
sometimes, the answers may not satisfy or please you. In these circumstances, consider the following
practices in your next prompt.

PRACTICE EXAMPLES
Repeat a variant of the question. Here are some variations:

• What are the causes of diabetes?


• Please list the causes of diabetes?
• Please summarise the causes of diabetes?
• Please explain the causes of diabetes using simple
language?

You should typically use open-ended questions—questions that


prompt an extended response rather than limit the answer to a
few possible options.

Prompt more information, more • “OK, but can you also define the terms?”
examples, or a better response. • “OK, but can you provide more examples?”

Ask the tool to appraise and to • “Can you please evaluation your response. Then, based on
improve the response. The AI this evaluation, can you generate a full, improved
tool will then use more data to response”.
improve the answer. • “How can you improve this answer?”
• “Rate your response, from a scale of 1 to 100, on the
degree to which the answer is valid and comprehensive. If
below 70, can you answer the question again, step-by-
step”.

7 | EFFECTIVE PROMPTS IN AI
If the AI tools refuse to answer a • “I appreciate some of these limitations. But can you
question, inform the tools they provide the best guess you can. Apparently, you have
are capable to answer (Mollick developed the capability to answer this question”.
& Mollick, 2024).

ADVANCED PRINCIPLES

To optimise prompts, prompt engineers also utilise a range of other techniques, as illustrated in the following
table.

PRACTICE EXAMPLES
ReAct prompting. Rather than merely ask Here is an example of a prompt you may enter.
a tool to complete some complicated task,
you also list a series of thoughts and actions “Decision task: I need to plan my conference trip to
about this task. If you apply this approach, Berlin.
the tool is likely to generate rational
suggestions that are more sensitive to the Thought 1: I need to gather information about this
circumstances. This approach is especially city.
useful if your tool can access other
databases or apps. Action 1: Search top attractions in Berlin.

Thought 2: I need to book accommodation.

Action 2: Search hotels or other accommodation in


Berlin that is less than $200 a night.

Thought 3: I need to check I have fulfilled the


necessary requirements, such as visas.

Action 3: Search travel requirements for Berlin”

Thought 10: I need to summarise the plan.

Action: Finalise the trip plan”.

8 | EFFECTIVE PROMPTS IN AI
You could then prompt the tool to identify more
considerations and then expand this list
considerably.

Delimited with XML tags—useful if your Here is an example.


prompts contain lists or other extensive
information. Specifically, you can write a I have presented my students with many ‘facts and
prompt that refers to some tag or label and principles’ <delimited with XML tags>. Can you
then define this tag or label later. generate 10 multiple-choice questions to test
whether students have learned these facts and
principles.

<facts and principles>


A cloud weighs around a million tonnes
Giraffes are 30 times more likely to be hit by
lightning than are people.
Fingerprints differ between identical twins.
Smells can pass through liquid
Murder rates increase in summer
On average, Mercury is closer to Earth than is
Venus.
Cognitive behavioural therapy can reduce
psychopathy.
</facts and principles>

In this instance

• the tag or label is called “facts and principles”,


• the phrase in brackets, delimited with XML tag,
reminds the tool that facts and principles is a tag
or label that will be defined later,
• the tags <facts and principles> and </facts and
principles> indicates where the material in this
tag begins and ends.

PROMPT DATABASES AND TOOLS

Prompt databases

9 | EFFECTIVE PROMPTS IN AI
You may also use apps or databases to improve your prompts. For example, to write more effective
prompts, you could access a range of databases. These databases offer examples of prompts that are
especially useful and effective in specific circumstances and include:

• PromptBase: includes prompts to generate more creative products,


• PromptPerfect: includes prompts to write stories,
• Prompt.chat: includes prompts that are most relevant to conversations,
• Prompt.Hero: includes a broad range of prompts for many writing tasks,
• MidLibrary: includes prompts to improve the style or format of outputs,
• AllPrompts: includes prompts that are relevant to a variety of purposes.

Prompt apps
Some apps have been designed to improve prompts. For example

• Hamming’s Prompt Optimizer uses AI to generate and refine your prompts, saving you significant time,
• you simply type the inputs or information you can provide, as well as the outputs or responses you are
seeking, and the tool will generate a suitable prompt.

Prompts that improve prompts


Finally, you can also utilise your own AI tools, such as Chat GPT, to improve your prompt. You would first
insert the following prompt—although you might swap the phrase “Chat GPT” with the tool you are using.
You would then answer the questions the tool asks.

I would like you to help me write a prompt. Your goal, as a prompt engineer, is to help me craft the best
possible prompt that aligns with my needs. Here is the process:
Your first response: Your first response is to ask me about the topic or subject of the prompt. I will
answer this question. Then, in the following steps, with your assistance, I will gradually iterate this topic
or subject to generate an effective prompt.
Your subsequent responses: Your second response will comprise two sections. First, you will use my
previous answer to generate a possible prompt that is clear, concise, and understand by Chat GPT.
Second, you will ask relevant questions to gather more information you need from me to improve the
prompt further.
Next responses: You will then repeat this subsequent response iteratively, gradually updating the prompt
and asking questions to achieve this goal. We will repeat this cycle until I confirm that we have reached
the desired prompt.
So, to reiterate, begin by clarifying the topic or subject of the prompt. Then gradually ask questions to
improve the prompt.

10 | EFFECTIVE PROMPTS IN AI
When uses apply AI and other algorithms to improve prompts, the answers that AI tools generate are more
accurate and comprehensive. These improved prompts enhance the responses by about 50% (Ark Invest,
2024).

11 | EFFECTIVE PROMPTS IN AI

You might also like