5 Ways Your Prompts Need To Change To Get The Best of GPT-4o

Summary

This video script provides essential tactics for optimizing prompts to achieve the best results from Chat GPT. It emphasizes the importance of clear details, breaking down complex tasks, utilizing larger prompts with delimiters, referencing texts for specific information, and using one-shot or few-shot prompting with examples. The script illustrates these tactics with practical examples, demonstrating how to structure prompts for detailed outputs, organize virtual conferences, synthesize research findings, and write abstracts in a specific journal style.

Takeaways

• 📝 Be specific and clear in your prompts to provide the AI with more context, leading to more accurate predictions and relevant outputs.
• 📚 Include detailed instructions in your prompts, such as specifying the desired output length, focus, audience, and tone, to guide the AI effectively.
• 🔍 Use examples to illustrate the task, such as one-shot or few-shot prompting, to help the AI replicate the desired outcome more effectively.
• 📈 Break down complex tasks into simpler subtasks to make it easier for the AI to provide precise and helpful responses for each step.
• 📑 Utilize larger prompts with clear delimiters to structure complex interactions and maintain the AI's understanding across a large context window.
• 📘 Provide reference texts and specific questions to extract accurate information, ensuring the AI cites the document and does not fabricate information.
• 🔑 Use delimiters consistently to separate different parts of your request, helping the AI understand where one part of the input ends and another begins.
• 📝 When providing a large prompt, structure it clearly with instructions and examples to guide the AI in generating the desired output.
• 🔬 Test the AI's ability to handle complex instructions by attaching articles and giving detailed tasks, such as summarizing and crafting sections based on those summaries.
• 📋 Demonstrate the effectiveness of using examples by showing how the AI can mimic the style of abstracts from a journal when given two or more examples.
• 📈 The video script emphasizes the importance of prompt engineering techniques to optimize interactions with AI and achieve better results.

Q & A

What are the main tactics discussed in the video for getting the best results from chat GPT?

The video discusses five main tactics: 1) Include specific and clear details in your prompt, 2) Split complex tasks into simpler subtasks, 3) Use larger prompts with delimiters, 4) Provide a reference text, and 5) Use one-shot and few-shot prompting with examples.

Why is it important to include specific details in a prompt for chat GPT?

Including specific details provides the model with more context, which helps it to predict and generate more relevant and accurate subsequent words and tokens.

Can you give an example of a bad prompt according to the video?

A bad example prompt would be something like 'write about AI in healthcare', which is too general and lacks specificity.

What is the purpose of splitting complex tasks into simpler subtasks?

Splitting tasks simplifies the process for chat GPT, making it easier to provide precise and helpful responses for each step of the task.

How does using delimiters help in structuring large prompts for chat GPT?

Delimiters help the AI understand where one part of the input ends and another begins, making it easier to process and respond to complex and detailed prompts.

What is the advantage of using a reference text in prompts?

Using a reference text allows chat GPT to extract specific information and answer questions accurately, ensuring the response is based on the provided document.

How does one-shot prompting differ from few-shot prompting?

One-shot prompting involves giving one example to illustrate the task, while few-shot prompting involves giving two or more examples to help the model replicate the outcome more effectively.

What is the recommended number of examples to use in few-shot prompting according to the video?

The video suggests that using two to five examples in few-shot prompting yields better results.

How does chat GPT handle a question if the reference text does not contain the information needed to answer it?

If the reference text does not contain the necessary information, chat GPT is instructed to respond with 'insufficient information'.

What is the significance of providing examples when asking chat GPT to write in a specific style, such as an academic abstract?

Providing examples illustrates the desired style and structure, enabling chat GPT to replicate it accurately in its response.

What is the next step the video promises to deliver in terms of prompt engineering tactics?

The video promises to cover five more tactics in the next video, encouraging viewers to subscribe for updates.

Outlines

00:00

📘 Optimizing Chat GPT Prompts with Clear Details

The first paragraph introduces the video's focus on tactics for crafting effective prompts for Chat GPT, referencing Open AI's prompt engineering guide. The key tactic discussed is the importance of including specific and clear details in prompts to provide the AI with sufficient context for accurate predictions. An example is given to illustrate the difference between a vague prompt and a detailed one, showing how specificity improves the relevance and quality of the AI's output.

05:02

📚 Breaking Down Complex Tasks for Chat GPT

The second paragraph emphasizes the strategy of splitting complex tasks into simpler subtasks to facilitate more precise responses from Chat GPT. It demonstrates how breaking down a request, such as organizing a virtual conference, into actionable steps results in a more detailed guide. The paragraph also shows how further details can be obtained by specifying particular steps, leading to a more comprehensive and step-by-step approach.

10:02

🔍 Utilizing Delimiters for Structured Prompts

The third paragraph discusses the use of delimiters to structure large prompts and help the AI understand the different parts of a request. Examples of delimiters are provided, and a detailed example is given where XML tags are used to structure instructions for summarizing research papers and developing sections based on those summaries. The output from Chat GPT is shown to be well-structured and adhering to the provided guidelines, highlighting the effectiveness of using delimiters.

📝 Extracting Information with Reference Texts

The fourth paragraph introduces the technique of providing a reference text enclosed in triple quotes for Chat GPT to extract specific information from. It outlines a method where the AI is instructed to answer a question using only the provided document and to cite the passages used. An example is given where the AI successfully extracts a definition from a text and cites the source, and also correctly indicates 'insufficient information' when the question cannot be answered by the provided text.

🎯 Enhancing Prompts with One-Shot and Few-Shot Examples

The final paragraph of the script explores the power of providing examples through one-shot and few-shot prompting to illustrate tasks for the AI. It explains that giving clear examples helps the model replicate the desired outcome, with research suggesting that few-shot prompting (two to five examples) yields better results. An example is provided where Chat GPT is instructed to write an abstract in the style of existing journal abstracts, demonstrating the AI's ability to mimic the style effectively.

Keywords

💡Prompt Engineering

Prompt engineering refers to the art of crafting input prompts for AI systems to elicit desired responses. In the video, it is the central theme, focusing on how to effectively communicate with AI like Chat GPT to achieve optimal results. The script provides various tactics to improve prompt engineering, such as including clear details and splitting complex tasks into simpler subtasks.

💡Tokens

In the context of the video, tokens represent individual elements, typically words, that an AI model considers when generating responses. The script explains that providing specific details in prompts helps the model to predict subsequent tokens more accurately within its context window.

💡Context Window

The context window is the set of tokens or words that an AI model uses to understand the current session and predict the next tokens. The script emphasizes the importance of this concept in creating effective prompts, as it directly influences the AI's output generation process.

💡Actionable Steps

The script suggests breaking down complex tasks into actionable steps to simplify the process for the AI. This approach helps in getting more precise responses from Chat GPT, as demonstrated by the example of organizing a virtual conference on AI, where the AI is first asked to list the steps involved.

💡Delimiters

Delimiters are symbols or words used to separate different parts of a prompt, helping the AI to understand the structure of the input. The video script mentions various types of delimiters, such as triple quotes and XML tags, and advises staying consistent with them for clarity in complex prompts.

💡128,000 Token Context

This term refers to the capacity of the AI model to retain and understand information across a large number of tokens, approximately 100,000 words. The script highlights this as an advantage for conducting detailed and complex interactions with the model, allowing for longer and more comprehensive prompts.

💡Reference Text

The script discusses the use of reference text to extract specific information from documents or data sets. This is done by enclosing the text in triple quotes and providing a specific question for the AI to answer, ensuring accuracy and adherence to the provided material.

💡One Shot and Few Shot Prompting

These terms refer to techniques where the AI is given one or a few examples to illustrate a task, helping it to replicate the outcome. The script explains that providing examples, especially in few shot prompting, can significantly improve the AI's performance in tasks such as writing abstracts in a specific style.

💡Research Papers

The script uses the term 'research papers' in the context of summarizing and synthesizing academic findings. It illustrates how to instruct the AI to create sections based on these summaries, such as research questions and theoretical frameworks, adhering to specific guidelines and examples provided.

💡Insufficient Information

This phrase is used in the script to indicate a response from the AI when the provided reference text does not contain the information needed to answer a question. It demonstrates the AI's ability to acknowledge limitations in the data, rather than fabricating answers.

💡Abstracts

Abstracts are brief summaries of research papers, and the script discusses how to instruct the AI to write them in a specific journal's style by providing examples of existing abstracts. This demonstrates the effectiveness of using examples in prompt engineering for academic writing.

Highlights

Introduction to the most important tactics for effective prompt engineering with Chat GPT.

Tactic 1: Importance of including specific and clear details in prompts for better predictions.

Explanation of the model's context window and how it uses tokens to predict subsequent words.

Example of a good prompt for writing an article on AI in healthcare with specified focus and audience.

Demonstration of the model's output adhering to the detailed prompt, showcasing six paragraphs and a conclusion.

Tactic 2: Splitting complex tasks into simpler subtasks for more precise responses.

Example of breaking down the task of organizing a virtual conference into actionable steps.

Tactic 3: Utilizing larger prompts with the 128,000 token context limit for detailed interactions.

Use of delimiters to separate different parts of a request for clarity in long prompts.

Example of structuring a prompt with XML tags for synthesizing research findings and drafting sections.

Tactic 4: Providing a reference text enclosed in triple quotes for extracting specific information.

Method of answering questions using only the provided document and citing passages.

Tactic 5: Power of providing examples using one-shot and few-shot prompting for task illustration.

Research showing better results with few-shot prompting using two to five examples.

Example of writing an abstract in a journal's specific style using two abstract examples.

Output demonstration of an abstract written in a similar style to the provided journal examples.

Upcoming video teaser on five additional tactics for prompt engineering with Chat GPT.

Transcripts

hello everyone and in this video we're

going to be looking at the most

important tactics that you need to use

in your prompts to get the best results

out of chat GPT and I'm going to be

showing you this with some practical

examples for each one and this is based

on open ai's official prompt engineering

guide which I'll leave the link for you

below in case you want to go into

further detail so tactic number one

include specific and clear details in

your prompt and I know this seems

straightforward but it's important to

understand why we need to do this and

this is because the model considers all

the previous tokens or let's just call

them words and uses it to predict the

next words within its current context

window which is basically the session

and the words that you're working with

right now so by providing that extra

detail in your prompt you give it

significantly more information to work

with and to accurately predict all the

subsequent words and tokens that are

coming next so the more you're able to

provide in terms of uh context the more

relevant your output is going to be so

if you were to give a bad example you'd

say something like write about AI in

healthcare which is completely General

but a good example would be to write a

prompt like this so write an article

composed of six paragraphs so you can

see I specified the amount of the output

that I want on the impact of AI in

healthcare focusing on how it enhances

personalized patient care okay so I've

said specifically what I want the focus

of my article to be the article is

intended for healthc Care Professionals

and administrators so I've said who my

audience is and where it's going to be

published in a leading Healthcare

Journal the tone should be informative

and professional specifi the tone the

way it should be written and it's

suitable for an audience familiar with

medical and technological advancements

telling it that you can use some

technical terminology as well so I'm

going to enter this

prompt okay so if we look at our output

we can immediately see it's given us six

paragraph output that we wanted and it's

also added a conclusion at the end uh if

we look at the actual writing we can see

that it's focused on the specific areas

that told it to focus on which is the

area of personalized patient care and I

can already see that it's given me some

examples of specific AI tools and if I

read through it I can see that it's

adopted the tone that I want it has

technical language and it sounds

professional as well so again being H

specific and clear in my in my prompt

will help me get a better output so

tactic number two split complex tasks

into simpler subtasks so instead of just

asking for help in general terms break

down your request into specific

actionable steps and this approach makes

it easier for chat GPT to give you a

more precise and helpful response for

each step so I'll give you an example

instead of prompting it by say help me

organize a virtual conference on AI you

can first ask chat GPT to give you the

steps involved so we will put a prompt

here that says what are the steps

involved in organizing a virtual

conference on the advances of AI okay so

I'm going to enter

that and you can see that it's come back

with some steps that tell me what I need

need to do in order to organize the

conference I've got find the conference

goals and so on got almost um 12

different steps that I would need to go

through and now I can then follow up

this prompt that says provide a detailed

guide on organizing a virtual conference

on the topic of advancements in

artificial intelligence and machine

learning and then the guide should

include all the above steps

and now you can see it's come back with

a detailed guide and for each step it's

now broken it down into further substeps

that tell me exactly what I need to do

for each one and it's given me a lot

more detail that it initially did in the

first part so helping it to think in

these in terms of these different steps

will get more detailed precise responses

and what you can do is you can even ask

for further detail

step number one consider any other

variables that should be added okay and

I can specify one specific step to go

into further detail and then I will

enter

that okay and you can see just for that

step again um it's gone into a lot more

detail in terms of the topics and themes

and so on so when we get the model to

think in terms of steps we're usually

getting a more precise detail detailed

output than if we just ask it to think

of a topic as a whole so tactic number

three is to use larger prompts with the

limits and with chat GPT for you can now

use much larger prompts because it has a

128,000 token context which is roughly

about 100,000 words and this means that

GPT 4 can retain and understand

information across all those words and

you can use this to your advantage to

have more detailed and complex

interactions with the model so to get

the best results from these long prompts

you need to use the limits clearly to

separate different parts of your request

and when we talk about Del limits

they're basically symbols or words that

help the AI understand where one part of

your input ends and the other begins so

some examples are as we can see here you

can use triple quotes you can use three

dashes you can use angle brackets you

can use XML tags really it doesn't

matter which delimiter you use the most

important thing is that you stay

consistent using the same delimiter for

the type of instruction that you're

giving so let me give you an example

just to clarify a bit and I probably

don't need to use all these delimits

here but I have quite a large uh prompt

that I'm going to be inputting and I

want to structure it um so that it's

very clear in terms of what it needs to

do and in terms of the output that I

want it to give me so if you look here

I've put my instructions um in XML tags

so showing the beginning and the end and

I've said you're an expert academic

writer tasked with synthesizing research

findings your primary Duty is to

summarize provided research papers and

then develop crucial sections based on

these summaries please adhere to the

following guidelines while drafting your

summary and section and then the next

section I've given it is the research

papers and I've said I have attached the

three research papers for summarization

which I'm going to attach them in the

prompt and then I've got some section

constraints so specific things that I

want it to stick to while it's

responding so maintain a clear and

concise academic tone limit each section

structure each section um clearly and

engagingly and so on and then I've got

the summary of the research papers so if

you look at the instructions above I

wanted to summarize the papers first and

then I've got a section that says

research paper sections to craft okay

and here so the first section I wanted

to come up with is the research

questions and I've also given an example

that I wanted to follow when it comes up

with the research question that it knows

when it comes back gives me some

research questions it will follow the

guidelines or the structure of this

example and then got my research

questions section and then this is where

it's going to insert the research

questions here and then I've got a

section on the important theoretical

Frameworks and it's going to insert the

theoretical Frameworks based on the

articles that I provided here so you can

see there's a lot of sections had I just

written it all as one prompt it might

not have come back with an accurate

response you don't definitely don't need

to include all that in every prompt so

I'm going to go back to chat GPT and

what I'm going to do is I'm going to

insert this entire prompt prompt over

here and what I'm also going to do

attach the Articles to this prompt so

after I've attached my articles I'm now

going to enter the

prompt okay and if we look at the output

it's come back with exactly what I've

specified it's first summarized the

research papers that I've provided and

you can see it's got three clear

summaries of the three research papers

the next section is the research

questions it's now come up with three

different research question questions

that are based on the um articles that I

have submitted and it's also following

the style of the example that I had

given and then the next section we can

see is the important theoretical

Frameworks again it's specified the

three theoretical Frameworks from the

above examples and you can see I was

able to get the exact output and the

structure that I wanted so tactic number

four is to provide a reference text and

you won't use this all the time but this

is really useful when you're focusing on

extracting specific information from an

article or a reference or a data set and

to do this you provide a document and

enclose that in triple quotes and a

specific question for chat GPT to answer

so we can add a prompt here that says

you will be provided with a document

delimited by triple quotes and a

question your task is to answer the

question using only the provided

document and to cite the passages of the

document used to answer the question if

the document does not contain the

information needed to answer this

question then simply write insufficient

information okay so basically we're

telling it don't make up the information

just say that it doesn't have the

information and then if an answer to the

question is provided it must be

annotated with the citation and use this

format to give me the uh citation okay

so I've added the prompt here now what I

need to do is I'm going to open my

triple quotes and I'm going to extract

the reference text from an article so

I'm going to add that in and I'm also

going to add like I mentioned the uh

question that I wanted to answer which

is what definition is provided for

entrepreneurial action and I'm going to

close off the section with my triple

quotes again and I'm going to enter that

and you can see the response that it's

given the definition provided of

entrepreneurial action emphasize the

importance of both opportunities and

individuals it's given me where it's

extracted the information from and the

citation here's the citation in the

format that I specified um above and

then it carries on giving me the uh the

rest of the information here as well so

what if I asked that something that

really isn't part of the text will it

respond as I asked it to to just say

insufficient information um so I'm just

going to ask it something that

definitely not in that reference text

I'm going to

say how many companies were used in this

sample and you can see it's come back

with insufficient information this is a

really good way to get accurate

information from a reference text and

also to build on tactic number three you

can see using the limits here to

identify the part of the text that we're

submitting is also really useful so

tactic number five is to provide

examples using one shot and few shot

prompting now providing examples is such

a powerful technique in prompt

engineering and the main idea here is to

illustrate the task by giving clear

concrete examples which can then help

the model to replicate the outcome and

this technique is sometimes called One

Shot prompting or few shot prompting

depending on how many examples you give

so obviously one shot prompting would be

just giving it one example and then F

shot prompting would be giving two or

more examples and research has actually

shown that when you use f shop prompting

so anything from two to five examples uh

you actually get much better results so

I'm going to use two examples in this

prompt and we're going to apply this to

getting chat GPT to write an abstract

based on existing abstracts in a journal

so say for example I want to uh submit a

paper to this journal and and they

already have a specific structure to the

way they write the abstracts I'm going

to copy two different abstracts exactly

the same and I'm going to give chat GPT

an instruction here that says okay use

these two abstract examples to write an

abstract in a similar style for a

research paper and then I've given the

title of the paper that I wanted to

write this is the end of the

instructions and you can see I've used

the limits here as well and then I've

put the abstract example one write an

abstract for research paper and I've

taken the exact name of the research

paper okay and then I've added here the

actual abstract that was written for

that paper with the same title and then

I've given another example also from the

same Journal using writer an abstract

for research paper and then I've found

another paper and given it the title and

then I've added the abstract so I'm

going to enter

that okay and now you can see from the

output it's come back with an abstract

that really is in a similar style to the

ab racts from the journal that I've

chosen and if we read through it you'll

find that it actually sounds really

academic it talks about the Gap that's

been addressed which is a key feature of

all the other abstracts that we have it

talks about the findings obviously this

one is uh chat GPT making these things

up have to input additional material in

your prompt to make it accurate but

you'll see that illustrates the findings

and and then it goes on to talk about

the discussion elements in the app so

really well uh written done an amazing

job in copying style of the previous

examples so in this video I've focused

on five main tactics that you can use to

really get the best out of your prompt

engineering there are five more that I'd

like to talk to you about but I'll keep

them to the next video so please

subscribe and I look forward to seeing

you in the next video