Every Essential AI Skill in 25 Minutes (2025)

I have learned all the AI things for

you. So, here's the cliffnotes version

of everything you need to know about AI

in my opinion in 2025. We'll be going

from beginner to intermediate to

advanced and I'll be giving you a crash

course on each topic as well as

providing more resources for you if you

want to dig deeper into any of them. By

the end of this video, you will know

more about AI than like 99% of the

population. But not if if you don't

actually retain that information. So

there will be little assessments

throughout the video. Now, pay

attention. Let's go. A portion of this

video is sponsored by Retool. Here's the

structure of the video. First, we're

going to go over the basic definitions

of AI and how they work. Then, we'll be

covering prompting, followed by agents

very hot these days, followed by AI

assisted coding. We're building

applications through what is called vibe

coding, and finally looking at some

emerging technologies going into the

second, half, of, 2025., All right,, let's

get started by first defining artificial

intelligence. Artificial intelligence

refers to computer programs that can

complete cognitive tasks typically

associated with human intelligence. Now

AI as a field has been around for a very

long time. And some examples of

traditional artificial intelligence

which back in the day we used to call

machine learning, include things like

Google search algorithms or YouTube's

recommendation system for recommending

you content like this video. But what we

typically refer to as AI these days is

what is called generative AI which is a

specific subset of artificial

intelligence that can generate new

content such as text, images, audio

video, and other types of media. The

most popular example of a generative AI

model is one that can process text and

output text otherwise known as a large

language model or LLM. Some examples of

large language models include the GPT

family from OpenAI, Gemini from Google

and the Claude models from Anthropic.

These days there are so many different

types of models now and many models are

also natively multimodal which means

that you can input and output not only

text but also images, audio and video.

Your favorite models like GPD40 or

Gemini 2.5 Pro are all multimodal. Okay

great. Now you know some of the basic

key terms that is used in the AI world.

So now I'm going to put on screen a

little quiz for this section. Please put

it in the comments below your answers to

these questions. Also, if you want more

details about these Genaii models

including a deeper dive under the hood

of these models, how they're being used

in our workplaces, as well as how to use

AI responsibly, I recommend that you

check out this video, which I'll link

over here, where I condense Google's

8-hour AI essentials course into 15

minutes. But for now, let's move on to

the next section on how to actually get

the most out of these AI models through

prompting. Let's first define prompting.

Prompting is the process of providing

specific instructions to a Genai tool to

receive new information or to achieve a

desired outcome on a task. This can be

through text, images, audio, video, or

even code. Prompting is the single

highest return on investment skill that

you can possibly learn. It's also

foundational for every other more

advanced AI skill. And this makes sense

because prompting is how to communicate

with these AI models. Like you can have

the fanciest models, the fanciest tools

the fanciest whatever, but if you don't

know how to interact with it, it's still

useless. So, if you want to get started

and practice prompting as a beginner

the first step is just to choose your

favorite AI chatbot. That could be Chad

GBT or Gemini or Claude or whatever it

is that you like. Next, I have two

pneumonics for you, which if you can

remember and implement will make you

better at prompting than 98% of the

population. The first one is what I call

the tiny crabs ride enormous iguanas

framework, which stands for task

context resources evaluate and

iterate. When you are crafting a prompt

the first thing that you want to think

of is the task that you want it to do.

What do you want the AI to do? For

example, maybe you want the AI to help

you make some IG posts to market your

new octopus merch line. You could just

prompt it, create an IG post, marketing

my new octopus merch line. And with

that, you'll probably get some okay

results, but you can make the results

much better. First, you can add in a

persona by telling the AI to act as an

expert IG influencer to make the IG

post. This allows the AI to take on the

role of an IG influencer and use some of

that more specific domain knowledge to

make a better IG post. Then you can also

add in the desired format of the output.

The default right now is a generic

caption with some hashtags, right? But

maybe you want something that's a little

bit more structured. You can ask it to

start the caption with a fun fact about

octtopi, then followed by the

announcement and ending with three

relevant hashtags. Great. This is now

already looking much better, but there

is still so much more we can do. The

next part of this framework is context.

The general rule of thumb is that the

more context that you can provide to the

AI, the more specific and the better the

results are going to be. The most

obvious piece of context that we can

provide right now is some pictures of

the actual merch that we're selling. We

can also add in some background about

our company. Like our company is called

Lonely Octopus, where we teach people AI

skills, like our recent AI agents boot

camp, which by the way, we sold out last

time within just 40 hours through the

wait list. So, thank you so much for

that. And we're actually going to be

opening up a new cohort soon. So, do

sign up for the weight list if you're

interested. I will link it over here

also linked in description. Anyways

some additional context that we can give

the AI is that our mascot, which is what

is on the merch here, is called Inky. We

can also be more specific about our

launch date and our target audience for

the merch, like people between the ages

of 20 to 40, mostly working

professionals, something like that. With

this context, your results are going to

be so much more precise and specific to

what you want. But we can do even

better. That's where the next step of

the framework comes in, which is

references. This is where you can

provide examples of some other IG posts

that you like. This way, the AI can take

inspiration from this example. Providing

examples can be so powerful because you

can describe things with words as much

as you like. But, you know, if you just

provide it with an example, there's like

so much there that you can capture the

nuances that you can incorporate into

the results. And voila, you press enter

and here is your IG post. Now, you want

to evaluate. Do you like it? Is there

anything that you want to tweak or want

to change? If so, you go into the final

step of the framework, which is to

iterate. When interacting with AI

models, it is a very iterative process.

So even at the first time it doesn't get

what you want, you can tell it like

tweak a little bit about this, add

something over here, change the color of

something, and you work alongside AI to

get the result that you finally want.

Tiny crabs ride enormous iguanas. If you

can remember this pneummonic and how to

use it, you would be better than 80% of

people at prompting. Let's call it 88

because that is a lucky Chinese number.

But if you want to be better than 98% of

the population, I have one more

framework for you. This is when you do

the tiny crabs ride enormous iguanas

framework and you feel like the results

are still not quite there. Well, you can

elevate this even further using the

ramen saves tragic idiots framework.

First part of the framework is just to

revisit the tiny crabs ride enormous

iguanas framework. See if you can add in

something else, maybe a persona. Be more

detailed about the output, more

references. Also consider taking out

something. Is there any conflicting

information in there that could be

confusing for the AI? Second part of the

framework is to separate the prompt into

shorter sentences. Talking to AI is

similar to talking to a human if you

just like word vomit all over them and

just say like a bunch of things. It can

be confusing for the AI. So you can

consider splitting what you're saying

into shorter sentences to make it more

clear and more concise. So instead of

just being like blah blah blah blah blah

blah blah blah blah blah blah blah blah

blah all over the place, you could just

be like blah then blah then blah. Make

sense? Third part of the framework is to

try different phrasing and analogous

task. For example, maybe you're asking

AI to help you write your speech and

it's just like not quite there, you

know? It's just like not really hitting

it. So maybe you can reframe this.

Instead of saying, "Help me write a

speech," say instead, "Help me write a

story illustrating whatever it is that

you want to illustrate." After all, what

makes a good speech is a compelling and

powerful story. Hello. So, this is Tina

from the future. I have just gotten back

to Hong Kong from Austin, and it seems

like in my jetlegged state, I have

forgotten to record the last part of

this framework. So I'm going to do that

now which is introducing constraints. Do

you have one of those friends where you

know maybe you are that friend when

someone asks like hey what do you want

to get for lunch and they're just like

oh anything. Yeah not very helpful.

Similarly if you feel like the output

from your AI is just like not quite

there. You can consider introducing

constraints to make the results more

specific and targeted. For example maybe

you're making your playlist for a road

trip that you're going on across Texas

and you know you're just really not

quite vibing with it. You can introduce

a constraint like only include country

music in the summertime. Much more

suitable, vibes., All right,, now, back, to

pastina. Got that? Ramen saves tragic

idiots. With these two frameworks

together, you'll be better than 98% of

people at prompting. By the way, I also

just want to say that I didn't just make

up these frameworks myself. I only take

credit for the cool pneumonics. The

actual framework comes from Google

itself. So, if you want to dive even

deeper and be better than like 99% or

even 100% of people at prompting, I

recommend that you check out this video

over here, which I'll link, in which I

summarize Google's prompting course

which is the best general prompting

course that I found so far. Also, I

would recommend checking out some of the

prompt generators for specific models

like this one from OpenAI, this one from

Gemini, and this one from Anthropic.

These are helpful for generating a first

draft and for getting the most out of

specific models. For anybody that thinks

that prompting as a skill is going to

become obsolete, think again. Especially

for more advanced applications like

building agents and coding, prompting is

getting more important than ever. It's

like the glue that holds everything

together to make sure that you get the

results that you want consistently. Now

speaking of more advanced skills, let's

now move on to the next topic, which is

agents.

AI agents are software systems that use

AI to pursue goals and complete tasks on

behalf of users. When we refer to AI

agents, we usually refer to it as an AI

version of a specific type of role. For

example, a customer service AI agent

should be able to receive an email maybe

of somebody being like, I forgot my

password and I can't log in. And it

should be able to reply to that email

and should be able to reference the

forgot password page on the website. As

of today, it can't do everything and it

can't handle all of the queries that a

customer service person should receive

but it can handle a lot of these kind of

generic or common questions that people

may have all autonomously. Similarly

for a coding agent, if you prompt it

well and you tell it to build like a web

application, it should be able to come

back with an MVP version of that web

application. Still got to like add on a

bunch of things and tweak it for sure

but it can write the code for the first

version of it. AI agents is a space

where there's a lot of interest and a

lot of money that is being poured into

it and I really expect them to get

better and better over time and

incorporate into all sorts of products

and businesses. In fact, the most golden

piece of advice that I have ever heard

about AI agents was from this YC video

which is for every SAS software as a

service company there will be a vertical

AI agent version of it. Every company

that is a SAS unicorn you could imagine

there's a vertical AI unicorn

equivalent. So what exactly makes up an

AI agent? Well, there are a lot of

frameworks out there, but the best one

that I've seen so far comes from OpenAI.

They list six components that make up an

AI agent. The first one is the actual AI

model. Can't have an AI agent without a

model. This is the engine that powers

the reasoning and the decision-m

capabilities of the AI agent. Second is

tools. By providing your AI agent with

different types of tools, you allow it

to be able to interact with different

interfaces and access to different

information. For example, you can give

your AI agent an email tool where it's

able to access your email account and be

able to send emails on your behalf. Next

up is knowledge and memory. You can give

your agent access to say like a specific

database about your company so that it's

able to answer questions and be able to

analyze data specific to your company.

Memory is also important when it comes

to specific types of agents. Like say if

you have a therapy agent and you have

like a really great session with it and

then next time around it just like

completely forgets what you're talking

about. That probably wouldn't be great.

So that's why you want to allow your

agent to have access to memory. So it's

able to remember all the different

sessions that you've had previously.

Then we have audio and speech. This

gives your AI agent the capability of

interacting with you through natural

language like being able to just to talk

to it in a variety of different

languages. Then we have guardrails. Be

no good if your AI agent goes rogue and

starts doing things that you don't

intend it to do. So we have systems for

that to make sure that your AI agent is

kept in check. And finally, there is

orchestration. These are processes that

allow you to deploy your agent in

specific environments, monitor them, and

also improve them over time. After you

build an AI agent, you don't just run

away and hope that it works by itself.

Speaking of AI agents, Retool just

launched its enterprisegrade agentic

development platform. Right now, there's

still a big gap between building AI

demos and AI that actually does useful

stuff in your business. Retool allows

you to build apps that connect to your

actual systems and take real actions.

You can use any LM like Claude, Gemini

OpenAI, whatever you want. Your agents

can actually read and write to your

databases, not just chat with you. It

also has endto-end support, including

test and emails to track performance

monitoring, access control, and a lot

more. These are all things that are not

flashy, but really crucial to real

implementation in your business.

Companies that are using retool plus AI

are already seeing really genuinely

impressive results. For example, the

University of Texas Medical Branch has

increased their diagnostic capacity by

10 times. Over 10,000 companies already

use Retool. So if you want to build AI

that is actually useful instead of just

look impressive, do check out

retool.com/tina

also linked in description. Thank you so

much retool for sponsoring this portion

of the video. Models provide

intelligence, tools enable action

memory and knowledge informs decisions

voice and audio enables natural

interaction. Guard rails ensure safety

and orchestration manages them all. I do

also want to point out that prompting is

also really really important when it

comes to agents, especially if you're

building multi- aent systems where

you're not just having a single agent

but you actually have networks of agents

that are interacting with each other.

Your prompts need to be very precise and

produce consistent results. So, how do

we actually build these AI agents like

what are the technologies for this?

There are quite a few currently

available for no code and low code

tools. I personally think nend is the

best for general use cases and gum loop

is great for enterprise use cases. If

you do know how to code, I recommend

checking out OpenAI's agents SDK, which

does have all these components built

into it. Or if you want something that

is free, there is Google's ADK agent

development kit. There's also the Claude

Code SDK, which is specific for coding

agents. Honestly, these different

technologies implementation methods are

going to keep changing over time, and

I'm sure within the next few months

there's going to be even more agent

builders for you to build agents with.

That's why I really recommend that you

actually focus on this fundamental

knowledge about the components of AI

agents, what are the different protocols

and the different systems because this

foundational fundamental knowledge is

not going to change so quickly and it's

going to be applicable to whatever new

tool and technology comes out. So, if

you do want to dive a little bit deeper

into AI agents, I have a video over here

that I made about AI agent fundamentals.

And if you want to get started in

building your AI agents, I also have

another video called building AI agents

which you can check out over here as

well. And I go into a lot more detail

about AI agents. So these are the

components that make up a single AI

agent. But often times you may also want

to build multi- aent systems in which

you don't have just one agent, but you

could have a system of agents that are

working together. And the reason for

this is kind of like if you have a

company and you just have like one

person trying to do everything in the

company, it's probably going to not be

great, right? That person is going to

get very confused trying to manage

everything at the same time. So it's

much better to have people with specific

roles that make up that company. Very

similar with agents. If you just have

one single agent trying to do

everything, then it's going to get

confused. there's going to be like a lot

of stuff that's happening. So, it's

often good to break it down into

different sub aents that have specific

roles and work together in order to get

the result that you want. If you want to

learn more about multi- aent systems

Anthropic has a really great article for

that and I'll link it in the

description. By the way, I'll link all

the resources that I'm referring to in

the descriptions. You may also have

heard about MCP, which is what a lot of

people are talking about these days.

This is also developed from Anthropic

and it's basically a standardized way

for your agents to have access to tools

and knowledge. You can think about it

like a universal USB plug. Prior to MCP

it was actually quite difficult to give

your agents access to certain tools

because all the different websites and

all the different APIs, they do it in a

different way and databases as well.

They're all configured slightly

differently., So,, it, was, kind, of a, pain

in the ass trying to like connect that

with your agent. But with MCP, because

there's a universal USB plug, you're now

able to give your agents any type of

tool and any kind of knowledge very

easily, assuming it follows the MCP

protocol. All right, here is a little

assessment on this agent section. Write

the answers in the comments. Next up

let's move on to using AI to build

applications, aka AI assisted coding

aka vibe coding.

In February of 2025, Andre Kaparthy, the

co-founder of OpenAI, made a viral

tweet. He says, "There's a new kind of

coding I call vibe coding, where you

fully give into the vibes, embrace

exponentials, and forget that the code

even exists. It's possible because the

LMS are getting too good. You simply

tell the AI what it is that you wanted

to build and it just handles the

implementation for you. And this, in my

opinion, is the new way of incorporating

AI into your products and your workflows

using vibe coding to build things. For

example, you can simply tell an LM

please create for me a simple React web

app called Daily Vibes. Users can select

a mood from a list of emojis.

Optionally, write a short note and

submit it below. Show a list of past

mood entries with a date and a note. And

you just click enter. And the LLM writes

the code for you. and generates this

app. And voila, there you go. But it

doesn't just end there. There still are

skills, principles, and best practices

for how to work with AI in order to vibe

code properly and produce products that

are actually usable and scalable. Let me

present to you now a five-step framework

for vibe coding with the pneummonic tiny

ferrets carry dangerous code. dangerous

code because if you don't do it

properly, you could potentially end up

like this guy over here who vibe coded

an app and then lost all of it because

he didn't understand something called

version control. Tiny ferrets carry

dangerous code stands for thinking

frameworks checkpoints debugging and

context. Thinking, as it sounds, is

about thinking really hard about what it

is that you actually want to build. If

you don't even know exactly what it is

that you want to build, how do you

expect AI to be able to do so? The best

way of doing this, in my opinion, is to

create something called a product

requirements document or a PRD. This is

where you define your target audience

your core features, and what it is that

you're going to use to build the product

with. I'll link an example PRD in the

description, but basically, you just

want to spend significant amount of time

thinking through what it is that you're

trying to build. Next up is frameworks.

Whatever it is that you're trying to

build, there has probably been very

similar things that have been built

before. So instead of just trying to

reinvent everything and telling the AI

to figure everything out, it's much

better to point the AI towards the

correct tools for building your specific

product by telling it to use React or

Tailwind or 3.js if you're making 3D

interactive experiences. But Tina, you

may ask, how am I supposed to know what

to tell the AI to use if I don't even

know what it's supposed to use? Great

question. AI can help you with that

too. When you're building your PRD, ask

the AI directly. I'm trying to build

something that's like, you know, like

this and it's very 3D animationheavy

for example,, and, I, want, it, to, be, a, web

app. What are the common frameworks for

building something like this? When

you're asking in this way, you're also

learning yourself what are the common

frameworks for building specific things.

And over time, you're going to have a

much better grasp of what you need to

use as well. In the era of vibe coding

you may not need to code everything by

yourself, but it still serves you very

well to understand the common frameworks

that are used for building different

types of applications. You should also

know how different parts and different

files in your project are interacting

with each other. This is going to help

you out so much as you're building more

and more complex features into your

product. Third step of the framework is

checkpoints. Always use version control

like Git or GitHub or else things will

break and you will lose your progress

and you will feel very very sad like

this guy who vibe coded an entire

application and then lost all of it

because he didn't understand version

control. Fourth step, debugging. you are

probably going to spend more time

debugging and fixing your code than

actually building anything new. That is

the reality. Be methodical and be

patient and guide the AI towards where

it is that it needs to fix. When you're

debugging, if you understand the file

structures and what's happening, then

you're much better at providing specific

instructions for where in your codebase

the AI should be debugging. The first

place to start when you come across an

error is to copy paste the error message

directly into the AI and tell it to try

to fix it. If it's something visual that

needs to be fixed, also provide a

screenshot for the AI. The more details

and the more context that you give the

AI, the better it would be at figuring

out how to fix the problem. And speaking

of context, the final part of the

framework is context. Whenever you're in

doubt, add more context. Generally

speaking, the more context that you

provide to AI, whether you're building

or debugging or you're doing whatever

the better the results are going to be.

This includes providing the AI with

mockups, examples, and screenshots. The

pneummonic to remember for this

five-step framework is tiny ferrets

carry dangerous code. thinking

frameworks checkpoints debugging and

context. A helpful way of thinking about

how these principles of the framework

work well together in the process of

vibe coding is to realize that there's

only two modes that you're ever in.

You're either implementing a feature

where you're debugging your code. When

you're implementing features, you should

be thinking about how to provide more

context, mentioning frameworks, and

making incremental changes. You always

want to approach building new things one

step at a time. Implement one feature at

a time as you build your product. When

you're in debugging mode, you should be

thinking about the underlying structure

of your project, where it is that you

should be pointing the AI towards

changing as well as providing more

context like error messages and

screenshots. So, we now know the

fundamentals of what makes good vibe

coding. So, what are the actual tools

that we use? There are a full spectrum

of development tools available. On one

of the spectrum is for complete

beginners, people who have no

engineering background and no coding

background. Some popular

beginnerfriendly vibe coding tools

include lovable, vzero, and bolt. Then

slightly more intermediate, we have

something like Replet. This is still

very beginner friendly, but it also

showcases the codebase, so you can

actually dig into a little bit more and

understand the structures of the

projects. Then a little bit more

advanced, you have something like

Firebase Studio. Firebase Studio has two

modes to it. It has the very

user-friendly prompting mode as well as

a full ID experience, which stands for

integrated development environment, an

interface that is specifically designed

for writing and working with code. In

this case, it was built on top of VS

Code, which is a very popular ID. With

Firebase Studio, you can alternate

between the no code prompting view and

decoding mode. Firebase Studio also has

the benefit of being free. Now, moving

on to the more advanced vibe coding

tools. This will include AI code editors

and coding agents like Windsurf and

Cursor. Everything that we talked about

earlier was all web- based, so the setup

is really easy. The environment is

isolated and it takes care of a lot of

things for you. But if you really want

to produce productionready scalable

code, then you generally need to start

migrating to using something like

windsurf and cursor. Development is

going to be on your local machine. So

the setup is going to be a little bit

more complex but you also have access to

a full suite of development tools and

different features for Windsor and

cursor. You just directly have that

coding environment that IDE. Then on the

most advanced side of the spectrum you

have command line tools like cloud code

for example., These, are, tools, that, live

directly in your terminal in the root of

your computer. With these tools you need

to be comfortable working in the

terminal or the command line. But it

does give you so much more functionality

and you can use it with any type of ID

of your choosing. Something like cloud

code really begins to shine when you're

working on complex code bases. But the

expectation here is that you do really

need to know how to code and know your

way around a computer and have a deep

understanding, of, software., All right,

that is a crash course on vibe coding.

If you do want to dig into this more, I

made a full video called Vibe Coding

Fundamentals where I go into a lot more

detail. I also made a video specifically

about Firebase Studio which I'll link

over here and another one where I talk

about the cloud for models and cloud

code which I'll link over here too. Now

I will put on screen a little assessment

to see if we have retained information

about vibe coding. Final section out.

What are things looking like going into

the future?

In the AI world, we don't measure things

in terms of years or even months. We

measure things in terms of weeks. And

the timelines are just getting more and

more compressed. When I was at the code

with cloud conference, Daario, the CEO

of Enthropic, made a really good

analogy. He says that it's basically

like being strapped on a rocket that is

going through time and time and space

are warping so that everything is

speeding up faster and faster and

faster. And especially because of this

if you're just trying to keep up with

all the AI news, all the things that are

coming out, all the new models, all the

new tools, all the new technologies, you

will never be able to catch up with

everything and probably get really

stressed along the way, too. So that's

why my advice is to not pay too much

attention to all the new things that are

coming out, but instead focus on the

underlying trends that are happening.

And I think there are three major

underlying trends. The first one is

integration into workflows and existing

products. 2025 is definitely the year in

which people are taking the AI and

actually integrating it into their

existing workflows. Prime example of

this is Google itself. I was at their

Google IO conference and they are

putting a lot of effort into just making

Google products better by integrating AI

throughout. And I think this should be a

model for all companies. Think about how

do you improve your processes by

incorporating AI to have a better user

experience and also to reduce your cost.

And when it comes to implementation of

this, there's massive productivity boost

if you learn how to do AI assisted

coding or vibe coding. With this full

spectrum of coding tools, there's a

dramatic decrease in barrier of entry

for people who want to build things and

who may not know how to code. But

there's also a big push towards

increasing the productivity of

developers. After experiencing command

line tools like cloud code, I can

absolutely see the massive benefits of

tools like this. And I think there's

going to be massive focus of developing

and improving command line tools. So I

think if you are technical or if you're

someone who's willing to learn technical

things, learning command line tools like

cloud code is going to be where it's at.

And finally, the focus on AI agents is

not going away at all. In fact, there's

more and more interest in building AI

agents because AI agents have so much

potential in improving existing products

and for building new products as well.

AI agents allow experiences to be

personalized, available 24/7, and at

much much lower cost. Like Weissy said

for every SAS unicorn company, there

will probably be an equivalent AI agent

company. I'm sure in the coming few

months, there's going to be more and

more tools that will allow you to

implement and build agents even more

easily. So, if you want to build

something, build a business, do a

startup, whatever, I would recommend

looking into AI agents. All right, that

is all I have for you today. Here is a

final little assessment. Please answer

these questions in the comments. Thank

you so much for watching till the end of

this video. I'm so excited to see all

the things that you guys are going to do

and build using AI. I really hope this

is helpful and good luck on your AI

journey. I will see you guys in next

video or live stream.