Today we'll explore prompt engineering

and understand why it's such an

essential skill for AI engineers.

Traditionally, a user provides a query

and this query is called a prompt. It is

passed to the LLM which processes it and

generates an output. Prompt engineering

improves this process by crafting clear

effective instructions and adding them

to the user's query before sending it to

the LLM. The instruction defines rules

to follow, the communication style, the

purpose, and the overall goal. This

greatly improves the quality of the

output produced by the LLM. So, we can

think of prompt engineering as an

intermediate step before sending the

user's query to the LLM. In this step,

we apply various techniques to enhance

the query with clear and effective

instructions.

So, what are the main techniques used in

prompt engineering? Let's go over the

most important ones. One, few shot

prompting. In few shot prompting, we

include a few input output examples in

the prompt to guide the LLM's behavior

and help it understand the desired

pattern. For example, if we want to use

an LLM as an English to French

translator, instead of just sending the

query, where is the trainer station, we

can first provide a few English to

French translation examples and include

them in the prompt as part of the

instruction. This way, the LLM can

follow the same format, tone, and style

we desire by following the format of

provided examples.

Two, zero shot prompting. In zero shot

prompting, we only provide an

instruction for the translation task

without including any specific examples.

For example, we can give an instruction

like translate the following sentence

into French. This makes the task clear

and helps the language model understand

exactly what to do. Three, chain of

thought prompting. In this technique, we

ask the model to reason a step by step

before giving the final answer. In its

few shot setup, we encourage a

step-by-step reasoning by including

examples that demonstrate this reasoning

process. In a zero shot setup, we ask

the model just to reason a step by step

before answering. This often improves

the quality of the LLM's output.

Four, role a specific prompting. In role

a specific prompting, we instruct the

model to take on a specific role or

persona. For example, we can instruct a

model by saying, "You are a financial

advisor." This helps the LLM generate

more accurate and context appropriate

responses.

Five, prompt hierarchy. In prompt

hierarchy, we establish different levels

of authority within the instructions.

A typical setup includes a system

message where hidden instructions define

the model's behavior, guards, and highle

goals. This instruction is known as the

system prompt. A developer prompt

contains instructions from the

application developer that define

formatting roles and customize the LLM's

behavior. And finally, a user prompt,

which is just the user's direct input or

question. Together, these three levels

form the complete prompt that is sent to

the LLM. While these are the most common

prompt engineering techniques, there are

many others as well, such as negative

prompting, where we include do not

instructions that specify what the model

should avoid doing, iterative prompting,

which relies on trial and error to

refine and discover the most effective

instruction.

and prompt training where we break a

complex task into a smaller more

manageable steps and guide the model

through them in sequence. For example,

instead of directly asking an LLM to

determine if a product is healthy from

an image of its ingredients, we can

first ask the LLM to extract the text

from the image and then decide if the

product is healthy. What are the

principles and best practices?

Here are some key principles and best

practices to keep in mind when designing

your prompts.

One, start simple. Begin with a

straightforward prompts, then gradually

move to more complex ones. For example,

start with a simple request like

summarize this article. Once that works

well, expand it to something more

detailed, such as summarize this article

in two sentences for a startup founder.

Two, break down tasks. Divide complex

tasks into a smaller manageable steps.

For example, instead of asking write a

full research proposal on education,

start by prompting a step by step. List

three main challenges of the education

system. Then suggest possible research

directions for each challenge. And

finally, combine these ideas into a

short proposal. Three, be specific.

Clearly state what you expect in terms

of format, style, and desired outcomes.

For example, say write a two paragraph

summary in a formal tone. Use bullet

points for key facts and conclude with

one recommendation.

Four, include necessary information.

Adjust the prompt length to strike a

balance. Be concise but detailed enough

for the model to understand the task

clearly. For example, instead of saying

write a report, provide essential

context like write a one-page report

summarizing the key findings from the

attached experiment results, focusing on

model accuracy and failure cases. That

wraps up a quick overview of prompt

engineering, what it is, how it works,

and how to use it effectively. To learn

more about prompt engineering, check out

the links in the description.