What are Prompts?

Prompts are textual inputs or commands provided to the Gen AI model for a multitude of
tasks to trigger a response or action. examples:
Simple question: "Who is the president of the US?"

What is Prompt Engineering?

a relatively new discipline for developing and optimizing prompts to efficiently use
language models (LMs) for a wide variety of applications and research topics. Prompt
engineering skills help to better understand the capabilities and limitations of large
language models (LLMs).
Prompt Engineering is a process used by:
-AI professionals: works one or several tasks are converted to a prompt-based dataset
that a language model is then trained to learn.
-Researchers: improve the capacity of LLMs on a wide range of common and complex
tasks such as question answering and arithmetic reasoning.
-Developers: to design robust and effective prompting techniques that interface with
LLMs and other tools

Types of Writing Prompts: Instruction, Continuation and In-Context Learning

1. Instruction: This is the most common type of writing prompt where we expect an
output from the model based on the instruction given.
Example:
E.g. Prompt :“Write a story of a poor hard-working farmer
who became a successful business with his hard work.”

2. Continuation: A continuation type of prompt is a specific format or style of prompt

where a partial sentence or phrase occurs in the prompt, and the language
model continues or completes the prompt by generating the rest of the text.
E.g. Prompt: “The following is a step-by-step process of
preparing chicken salad
Step 1: Cut the Veggies AI Response
Step 2: Toss the chicken nuggets
Step 3: Mix everything together”

3. In-Context Learning (ICL): This type of prompt is used for few-shot learning
where few examples are provided in the prompt and based on those examples,
we expect a response in the completion. The goal of In-Context Learning is to
enable models to learn from a few examples and generalize to new data.
E.g. Prompt: “Sustainability is a social goal about the ability of people to co-exist
on Earth over a long time. Specific definitions of this term are difficult to agree on.
 They have varied with literature, context, and time. Experts often describe
sustainability as having three dimensions (or pillars). These are environmental,
economic, and social. Many publications state that the environmental dimension
is the most important.”
Topic: Sustainability

What are Delimiters?

- A delimiter is a special symbol or character that separates different parts of the
prompt.
- It helps organize the information, making it easier for the language model to
understand and generate responses.
- Delimiters show the start or end of specific sections or instructions in the prompt.
Each completion should finish with a fixed stop sequence to signal the model
when the completion ends.
- A stop sequence could be a new line, hashtag, colon, semi-colon, or any other
token that is present in actuality.

-
- Delimiter Refers to a specific symbol or character used to separate or mark
different parts or components of the prompt.
- Helps to organize and structure the information provided within the prompt.
- Used to indicate the beginning or end of specific sections or instructions within
the prompt. Each completion should end with a fixed stop sequence to inform the
model when the completion ends. A stop sequence could be \n, ###,::,;;, or any
other token that doesn't appear in any completion.
- Helps in controlled task focus: eg. the “==” delimiter separates the translation
task instruction from the actual text to be translated.
- Helps in clear conversations by distinguishing user inputs from ai outputs.

- Helps in Task Separation, context preservation, customization and flexibility, and

simultaneous prompts.

Elements of Prompts: Instruction, Context, Input data, Output

1. Instruction : A specific task or instruction you want the model to perform.
2. Context: External information or additional context that can steer the model to
better responses.
3. Input Data: The input or question that we are interested to find a response for.
4. Output: The type or format of the output you expect from the model.
Key Processes in Prompt Engineering: Prompt Engineering involves crafting
instructions or queries to elicit specific responses from language models. It includes
formulating clear and precise prompts to guide the model in generating desired outputs.
There are three main steps to follow in prompt engineering:
1. Tailoring for the task: diff questions or instructions may be needed to get
answers, generate content, or perform various functions.
2. Contextual awareness effective prompts: in context of the conversation, or the
problem being addressed.
3. In-context learning: provides worked out examples in the prompt to clarify what
the LLM should produce.
Designing Prompts for different tasks: todays LLM can perform al kind of advanced task
that range from text summarisation and mathematical reasoning to code generation.

Properties of prompts:
1. Factual prompts: Factual prompt engineering involves crafting prompts that are
focused on presenting accurate and factual information. This can include
providing specific details, relevant context, and verified sources within the
prompts.
- Include ground truth, like an article paragraph or Wikipedia entry, to
reduce fabricated text.
- Decreasing probability parameters and prompting the model to admit "I
don't know" when it lacks an answer, to decrease response diversity.
- Example of factual prompt: Describe the major events that led to the
American Revolution, What is the population of USA?
2. Non-factual prompts: Non-factual prompting refers to the practice of crafting
prompts that do not adhere to factual accuracy.
- Non-factual prompts are often used to explore imaginative or subjective
responses from the model, allowing it to generate creative or speculative
content that is not necessarily grounded in real-world facts.
Imagine a world where gravity doesn't exist.
How would people move around?

• Tell me a story about a talking animal that becomes

a superhero

• If you could travel to any planet in the solar

system,
which one would you choose and why?"

Types of prompts: open ended and close ended

 1. Open ended prompts: requires the model to generate an answer with few
constraints or options given by the prompt.

Prompt type Description Example

Generate This entails the model "Write a blog on Terrace

producing unrestrained Gardening."
textual replies, which can
be described as tasks This is a non-factual
involving "writing," prompt.
"generating," or "creating"
text.

Open ended QA Posing a specific inquiry to "Which countries were a

the model that does not part of the Allied Powers
have the answer explicitly during World War II?"
stated within the given
prompt This is a factual prompt.

Open ended classification Requesting the model to "What subject matters are
categorize a given text into covered in this article?"
a specific class without
providing predefined This is a factual prompt.
classification options.

Chain of thought(CoT) Enables the model to "Sarah has 4 boxes of

break down a complex chocolates. Each box
problem into smaller, contains 8 chocolates. If
manageable steps that can she receives 3 more
be solved individually, boxes, each with 6
rather than attempting to chocolates, how many
solve the entire problem chocolates does she have
all at once in total?"

This is a factual prompt.

Brainstorm Seeking input or guidance "Provide me with a list of

from the model to assist the top 7 essential steps to
with generating an idea take in order to excel in a
job interview."

This is a factual prompt.

Chat Requesting the model to Human - Hello there, who

produce a chat-like are you? AI - I am your AI
response that incorporates assistant. How can I help
 background information you? Human -I'm having a
from earlier parts of the bit of a rough day, to be
conversation. honest. AI - …

This is a factual prompt.

2. Close ended prompts: asks the model to complete an action on a text, given
specific parameters and instructions.

Prompt type Description Example

Rewrite Modify the existing content Infuse a cheerful tone into

to suit different variations this blog.
such as form, tone,
persona, etc This is a factual prompt.

extract Extract a specific block of Identify the sentence

text verbatim from the where the doctor
given content recommended the
treatment plan.

This is a factual prompt.

Summarize Involves summarizing a Summarize the key tasks

block of text while from this conference
avoiding the use of exact report.
wording from the original
text. This is a factual prompt.

Close ended QA This entails extracting How frequently did Alice

information from a written discuss her new project in
text, which can be either this conversation?
extractive or abstractive in
nature This is a factual prompt.

Close ended classification Requesting the model to Categorize the content as

categorize a given text either Travel, Fashion, or
based on a provided Entertainment.
selection of classes.
This is a factual prompt.

Open ended instruction prompts:

1. Generate: expect the language model to produce new text using the provided
input. These prompts commonly involve completing a sentence, paragraph, or
 story or generating text on a specific topic. Keywords like "write," "create,"
"generate," "describe," "list," or "produce" are often used in such prompts.
Additionally, the prompt may specify the type of text desired, such as a short
story, recipe, list, or description.
2. Types of Generate prompt:
- Completion: these prompts involve presenting a partial sentence or phrase
and requesting the model to generate the subsequent words needed to
finish it.
Example: complete the following sentence: “The old house at the end of
street was…”
- Dialogue generation: these prompts require a model to create a
conversation between two or more characters.
Example: “Create a conversation between a detective and a suspect
during interrogation”
- Poetry Generation: these prompts require the model to create a poem
responding to a given prompt.
Example: “Create poem on river”
- Text Continuation: these prompts offer a sentence or passage of text and
request the model to extend or complete it.
Example: “The four friends went on an adventure where they
encountered…”
- Instruction: these prompts require a model to create a poem in response
to a given prompt.
Example: “write a step by step process of saving document in excel”
3. Generate Sub-Tasks:
- Knowledge Acquisition: prompts of this type include a guide to any skill
acquisition.
Example: “Explain principle of physics in human life and its importance”
- Informative: renders information on a particular topic in the form of essay,
articles, blog, research, educational material.
Example: “The causes, effects, and potential solution to global warming”
- Creative: includes creative which may include stories, personal narratives,
poetries, fiction in specific style.
E.g., "Imagine a world where humans can communicate with animals.
Write a short story about a young girl who loves to stay in such a world."
- Social/Promotional:Creative content on social media posts,
advertisements, and product descriptions in the product website.
E.g., "Craft a tweet promoting a new environmentally friendly product that
helps reduce household waste."
 - Persuasive:This type of prompt includes persuasive, opinion-related
content.
E.g., "Write a persuasive article advocating for the importance of
renewable energy in combating climate change."
- Collaboration: This prompt type includes debates, competitions, reviews,
critics, etc., and must be original.
E.g., "Present a debate if an electric car is better or a petrol car."
4. Code generation: LLM models are quite effective in performing code generation
tasks and co-pilot is a great example of this.
5. Chain of thought: The chain of thoughts method enables models to decompose
multi-step problems into intermediate steps.
With the chain of thought prompting, language models of sufficient scale (~100
billion parameters) can solve complex reasoning problems that are not solvable
with standard prompting methods.
It pertains to the series of ideas or concepts one's mind traverses while
contemplating a specific subject or issue.

6. Open ended QA: Open-ended question-answering prompts are designed to elicit

responses without restricting them to predefined choices or specific types of
answers.
They foster creativity exploration and accommodate a broader spectrum of
responses.
The model is then relied upon to generate a response based on its own
comprehension.
7. Open ended classification: Open-ended classification prompts involve classifying
a specific set of inputs, typically text, into various categories determined by the
researcher.
Unlike close-ended classification prompts that provide predetermined options for
classification, open-ended classification prompts do not have predefined
categories.

8. Brainstorm: Brainstorming prompts encompass a category of prompts that

prompt the model to produce a compilation of concepts or recommendations
associated with a particular topic or theme.
Unlike generate-type prompts, these prompts are more general, aiming to inspire
the model to generate a diverse array of ideas that may not be directly linked to a
specific task or objective.
Typically characterized by their open-ended nature, these prompts require the
model to generate imaginative, distinct, and varied ideas.
Brainstorming prompts are valuable tools for generating fresh concepts,
exploring uncharted possibilities, and uncovering novel perspectives on a given
topic.
9. Chat: Chat prompts can take three different formats: "single instruction-based,"
"paragraph-based instruction," and "instructions with examples.“
When creating chatbot prompts, several aspects can be considered: Defining the
persona or description of the chatbot as an entity.
Specifying the knowledge base or conversational topics the chatbot should
possess. Determining the desired format or specific characteristics of the
generated output.
Alternatively, a comprehensive prompt can be crafted as a paragraph,
incorporating all of these elements.

Close ended instruction prompts:

1. Rewrite: Rewrite prompts are questions that ask to rephrase or restate a
given text or sentence.
These prompts can manifest in diverse formats, such as
“Rewrite this sentence for conciseness",
"Correct the grammar errors in this paragraph",
"Enhance the persuasive tone of this text".
The model is expected to grasp the meaning and structure of the input text
to generate appropriate and coherent changes.
Rewrite sub-tasks:
- Proofreading/Editing: Proofreading prompts are requests or
instructions designed to help someone check written content for
errors, clarity, or improvement.

- Content enhancement: This includes rewriting the original content

by adding extra information. It encourages adding more
information, examples, explanations, or details to enhance the
content's quality, relevance, or comprehensiveness.

- Content Repurposing: This means changing or adapting the

content for different audiences or mediums or purposes. It often
involves reshaping original content to serve a new purpose or reach
a distinct audience.

- Style Transfer: This includes changing the content by adapting the

style of a writer, poet, character of any existing narratives. Style
transfer prompt typically involves the transformation of a piece of
content or artwork to adopt the characteristics, aesthetics, or style
of another.
2. Extract: Extract prompts are questions that ask to extract specific
information from a given text.
They require a person to read the text carefully and identify specific
details, facts, or pieces of information that are relevant to the prompt.
These prompts might require the model to recognize entities, concepts, or
relationships within the text or extract structured data such as dates,
numbers, or addresses.

3. Summarize: requires the language model to briefly condense a given text

or information while ensuring that the main points and information are
retained in a more concise version.
The prompts may require the model to generate a summary of textual
content, articles, or documents or to extract the main points from a
provided text.

4. Close ended QA: Closed-ended question-answering prompts entail a

defined and limited set of answers.
-These prompts require the model to furnish precise answers or choose
from a predetermined list, rather than generating a response based on its
own understanding.
-Typically, these prompts are limited to a specific context or adopt formats
such as multiple-choice questions, true/false questions, or queries that
necessitate a yes/no response.
-When constrained within a context, these prompts typically contain a
specific answer or a constrained set of answers that can be located in the
provided text to extract that information accurately.
 5. Close ended classification: entail posing questions that involve
categorizing a given set of inputs, often textual, into a predetermined set
of defined categories.
-These prompts usually require a straightforward classification without the
need for extensive explanation.

Open ended continuous prompts:

1. Open ended Text Generation”: In "Generate" Continuous Prompts, the focus lies
on seamlessly crafting text with an open-ended approach while ensuring
continuity throughout the generated output. It's about fostering a smooth flow,
coherence, and logical progression between sentences and paragraphs. Within
this framework, certain aspects of the prompt are intentionally left open,
prompting the model to fill in the missing components and produce a
comprehensive and interconnected response. This process allows for the
creation of cohesive narratives and structured content, enhancing the overall
quality and coherence of the generated text.

2. open-ended QA continuous: prompts the generation of text in response to a

question, ensuring a continuous flow throughout. This format encourages the
generation of detailed, contextually relevant answers that surpass simple
 one-word or limited responses. The continuous nature of the prompt allows for
the evolution of the conversation over time, with the model consistently providing
detailed and contextually appropriate responses to the user's queries, fostering
deeper engagement and understanding.

3. open-ended classification continuous prompts: The term 'open-ended

classification continuous prompts' embodies a synthesis of various concepts
revolving around text generation employing an open-ended approach, executing
classification tasks, and ensuring continuity in the resulting output. Within this
framework, a segment of text related to classification is furnished in the prompt,
with the expectation that the model will complete the classification within its
response. This approach not only stimulates the generation of comprehensive
classifications but also facilitates a seamless and coherent flow in the generated
text, thereby enhancing the quality and relevance of the model's responses.

4. Chain of thought continuous prompts: Chain of thought continuous prompts

encapsulate a series of prompts or statements aimed at fostering the generation
of text in a cohesive and uninterrupted manner. Their objective is to sustain a
continuous flow of ideas or information, thereby preserving the coherence and
connectivity of the generated text. By promoting a seamless progression of
thoughts, these prompts facilitate the development of coherent narratives or
responses, enhancing the overall quality and engagement of the generated
content.
5. brainstorm continuous prompt: Its goal is to foster creativity, encourage
exploration of various possibilities, and prompt the creation of diverse ideas
relevant to a specific subject or issue. In such prompts, the initial idea originates
within the prompt and seamlessly transitions to further elaboration within the
response. This open-ended approach aims to sustain the momentum of
brainstorming by prompting individuals to delve deeper into and expand upon
their initial concepts, thereby facilitating a continuous flow of imaginative thought.

6. Continuous chat prompts: Continuous chat prompts are tailored to sustain an

ongoing and interactive dialogue with an AI chatbot or virtual assistant. Their
objective is to uphold the conversational flow, enabling users to contribute input,
pose inquiries, or engage in a dynamic exchange with the AI. The model
 seamlessly progresses the conversation, ensuring logical continuity and
resolution.

Close ended continuous prompts:

1. rewritten Continuous Prompt: the model continues generating output based on
the partial information given in the prompt. The context and purpose are
presented in the prompt but left incomplete, allowing the model to complete it
cohesively in the response.

2. Extract: In an extract prompt, text and partial information is provided in the

prompt and based on this partial information the model generates the response.
3. Summarize continuous prompt: the text for the summary and partial summary is
provided in the prompt and the remaining part of the summary is created by the
AI model as a response. The process leverages natural language processing
capabilities of the AI model to ensure a coherent and contextually relevant
summary.

4. continuous close ended QA prompt: in a continuous close ended QA prompt the

text options and partial context is provided in the prompt based on which the
model continues to generate response which adheres to the text given in the
prompt. Unlike open-ended prompts that allow for more diverse responses,
closed-ended prompts guide the model to generate a specific type of response
based on the provided information.

5. continuous close ended classification: In a continuous close ended classification,

the text and the classification is given in the prompt with partial response
 providing enough context for the model to generate logical response based on
the given text. This format guides the model to understand the context, category,
or class associated with the text and then continue the response in a manner that
aligns with the given classification.

##Open Ended In Context Learning Prompts:

1. Generate ICL: The objective of In-Context Learning (ICL) is to empower models
to learn from a limited number of examples and apply that knowledge to new
data. In the ICL Generate prompt, a few examples are provided, guiding the
model in generating responses. This approach allows the model to generalize
from the provided examples and produce coherent and contextually relevant
outputs.

2. open-ended ICL QA prompt: In an open-ended ICL QA prompt, the question is

designed to be open-ended, allowing for a free-form response from the model
based on the provided context. In this type of Open-Ended QA prompt nature will
be provided to the model expecting the model to generate an accurate response
of the same type.
 Prompt: [k=4] :
What is the currency of India
Response: The currency of India is Rupees###
What is the currency of Japan
Response: The currency of Japan is Yen###
What is the currency of USA
Response: The currency of USA is US Dollar###
What is the currency of UK?
Response: The currency of UK is Pound sterling###
What is the currency of Germany
AI Response :
The currency of Germany is Euro

3. Open ended ICL classification: The "ICL Classification" component introduces

conditional instructions guiding the model's response by prompting it to classify
or categorize information in a specific manner. Within this framework, a few
prompts on Open ended classification are provided, directing the model to
generate responses of a similar type.
Prompt: [k=3] :
Give the genre of the following novel
Novel: Wuthering Heights
Genre: fiction
Novel: Frankenstein
Genre: horror
Novel: The Da Vinci Code
Genre: mystery
Novel: To kill a Mockingbird
Genre:
AI Response :
Drama

4. Open-Ended ICL Chain of Thought prompt: An Open-Ended ICL Chain of

Thought prompt is a question or task meant to inspire creative and critical
thinking within a meaningful context. With no predetermined response, the
prompt aims to spark a connected sequence of ideas or a "chain of thought" as
individuals engage in the learning process. This prompts a dynamic and evolving
conversation with the model, where responses progressively build upon each
other in a coherent manner.
##Closed ended ICL prompt:
1. Rewrite ICL: In a Close-Ended ICL prompt, a "rewrite" instruction usually entails
requesting the model to offer an alternative or rephrased version of a provided
input text. This type of prompt assesses the model's capacity to generate varied
yet contextually equivalent responses. To grasp the functionality of rewrite within
the context of a Close-Ended ICL prompt, multiple rewrite prompts are included
as examples within the prompts.

2. Extract ICL: In a Close-Ended ICL prompt, an "extract" instruction typically

involves directing the model to pull specific information or details from a given
input text. This type of prompt is designed to test the model's ability to
comprehend and selectively retrieve relevant content from a larger context.
-In this, few examples from the extract category is provided to the model so that
the model can produce response based on the extract category based on the
instruction given.
Prompt: Extract the adjectives from the passages:
 Passage 1: “The radiant sun slowly dipped below the horizon, casting a golden
glow across the tranquil ocean.”
Adjectives: radiant, golden, tranquil
Passage 2: “The bustling city streets were alive with vibrant energy. Skyscrapers
towered above, reaching towards the
endless blue sky.”
Adjectives: bustling, vibrant, endless, blue
Passage 3: “The serene meadow was a haven of tranquility. Soft blades of
emerald-green grass caressed bare feet as one wandered through the expansive
landscape.”
Adjectives:
AI Response: Serene, soft, emerald-green, bare, expansive

3. Summarize ICL: A "summarize" instruction typically involves directing the model

to condense and provide a concise overview of the information presented in a
given input text.
-This type of prompt is designed to test the model's ability to understand the main
points of a passage and articulate a brief summary that captures the essential
details.
Prompt: Provide a summary of the below articles in not more than 20 words.
Passage: “At the most basic level, nutrition is about eating a regular, balanced
diet. Good nutrition helps fuel your body.
The foods you eat supply the nutrients your body needs to maintain your brain,
muscle, bone, nerves, skin, blood
circulation, and immune system. Proper nutrition also helps protect you from
illness and disease, such as heart disease,
diabetes, cancer, and osteoporosis.”
AI Response: Summary: Nutrition is vital for overall health, providing necessary
nutrients for body functions, protecting against diseases like heart disease and
diabetes.

4. Close ended QA ICL: A close-ended QA prompt is designed to elicit a specific

and concise response from the model. This type of prompt is structured to obtain
a definitive answer without requiring additional context or explanation.
-In a close ended QA, sample QnAs from a given passage are provided and the
model then generates the answer from the passage in a similar fashion as
provided as an example in next slide.
Prompt:
While details may vary from diet to diet, all healthy eating plans have four basic
principles in common:
 1. Abundant plant foods, rich in vitamins, minerals, fiber, and phytochemicals,
provide various health benefits, including antioxidant and anti-cancer activity.
2. Essential protein intake is highlighted, with plant sources (beans, lentils, soy,
nuts, seeds) and fish offering optimal health benefits.
3. Opt for minimally processed foods to prevent weight gain and adverse effects
on blood sugar and cholesterol.
4. Adhere to U.S. Dietary Guidelines, limiting saturated fats and added sugars to
less than 10% of daily calories, and sodium to 2,300 mg or less per day.
Given the passage, answer the following:
Q: What are the four basic principles in healthy eating plans?
A: The four basic ……
Q: Which food items should have minimum consumption?
A: Processed food, saturated fat, added sugar and sodium should be consumed
less.
Q: Which food items should have maximum food consumption?
A:
AI Response
Lot of plant-based food, and adequate protein should have maximum
consumption.

##Properties of Prompts:
1. Factual Prompts: Factual prompt engineering involves crafting prompts that are
focused on presenting accurate and factual [Link] can include
providing specific details, relevant context, and verified sources within the
prompts.
-Include ground truth, like an article paragraph or Wikipedia entry, to reduce
fabricated text.
-Decreasing probability parameters and prompting the model to admit "I don't
know" when it lacks an answer, to decrease response diversity.
Example: Describe the major events that led to the American Revolution
What is the population of USA?
● Close ended factual prompt: The prompt "close-ended QA" (Question and
Answer) is typically associated with factual prompts. It refers to questions
that have specific, predetermined answers, often in the form of "yes" or
"no" or a short response. These questions seek factual information or
confirmatory answers rather than open-ended discussion or subjective
interpretation.
● Extract:
Example:
Context: Earlier,the Prime Ministers of India lived in their own
house or house allotted to them through Parliament, allotment by
virtue of being an MP. Jawaharlal Nehru took up residence in Teen
Murti Bhavan, which used to be the residence of the
Commander-in-Chief of the British Indian Army in British India; it
was formerly called the Flag staff House. After Nehru's death in
1964, the building was converted to the Nehru Memorial Museum
and Library in 1984.
Prompt: From the above passage extract the name of th residence
where Jawaharlal Nehru used to live after becoming PM of India.
Response:Teen Murti Bhavan and Flagstaff House

● Chain of thoughts: This particular prompt is factual as it revolves

around a concept that can be proven.“
The example prompt presents a scenario with red and blue marbles
and asks for the maximum number of equal groups without leaving
any marbles behind. This engages the model in a logical process to
arrive at a factual answer."

● Summarize:
● Open ended classification:Unlike traditional open-ended questions,
this type of prompt introduces a fact or seeks a definitive answer. It
encourages a more exploratory and analytical approach, fostering
discussion around a broad topic or concept.

● Rewrite: Rewrite is particularly useful when presenting information

to a diverse audience or those unfamiliar with specific terminology.
2. Non-Factual Prompts: Non-factual prompting refers to the practice of crafting
prompts that do not adhere to factual accuracy.
-Non-factual prompts are often used to explore imaginative or subjective
responses from the model, allowing it to generate creative or speculative content
that is not necessarily grounded in real-world facts.
● Generate: "Generate" prompts do not seek a specific factual answer but
encourage the generation of new and creative responses, often based on
individual interpretation, imagination, or problem-solving skills.

● Brainstorm:
● Open ended QA:both factual and non-factual responses

● Close ended classification: