It has never been easier to get started

coding with AI. But that's not the

reason to start. You should start

because this skill is becoming

nonoptional. Not just for coding, but

for interacting with AI in any way. The

gulf is really widening between the

people that understand how to interact

with these tools and the people that

don't. And the people that don't, I

guarantee you, will get left behind.

Maybe you've already played around a

little bit with Copilot or ChatgPT or

even Cursor and you've had some success

or maybe all failures, but you really

didn't know how to get from the toy

stage or the not entirely successful

stage all the way over into the complete

win column. Well, that stops today. I'm

going to show you exactly how I use

these tools, the real workflow. Every

step along the way, you can build right

along with me. I'm going to share a full

repo to make this super simple and I'm

going to point out the actual critical

elements that make the difference. This

is not going to take long. It's easier

than you think. But you have to get

invested in your own future. That's what

this is really about. All right, let's

get in there. But first, let's talk

about the tools. Okay, you need to get

armed. We're starting with Claude Code.

I'm going full opinionated here. So what

I'm going to give you is the best tools

in the best pattern that I've found.

That's where you're starting today. So

this is Claude Code. That's the tool. It

means an anthropic subscription. That's

20 bucks. I understand there's real

money, but for one month, invest in

yourself. And when you do, think about

using this every single day. Claude Code

runs in your terminal. Use the terminal

version first. I guarantee you this is

the best path, and that's why I'm

starting you here. Now that we have this

in your terminal, I want to tell you I

am going to be using something called

Visual Studio Code, which is just an IDE

that allows me to also see the files.

You'll see me poking around in a minute

in the files here. It will be important.

This to me is really a very large

unlock. Okay, next is slash commands. If

you hit slash, you'll see the commands.

This is like the menu on any application

typically. I'm not going to read through

this. You don't need to know most of

them. You'll learn them as you need

them. There are a few that are really

important and the first one takes us to

the last thing that I need to tell you.

So if you did /mod, it will take you to

a list of the models. Now you will get

Opus. This one here is Opus45. That's

their biggest model. You'll also get

Sonnet. It's only January 2026. In the

last 2 months, something massive has

actually happened using Clawude Code as

the harness and Opus 45 and planning

mode. It has actually changed things in

a meaningful way and it really gaps

everything else out there. So use Opus.

All right. So these are the tools. Now

that you're armed, let's build. Okay.

I'm going to share everything with you.

There's a new project that is out on my

GitHub. I will share this. This link

will be in the show notes below that you

can take a look at them. There are

different branches in here. These

branches correlate to the different

steps and those different steps are

simply the requests that I put in. So

you can see here that I basically put

seven or so requests into the system.

That's the different checkpoints. You

can start at any one of them. We're

going to be actually starting at the 01

initial one in just a second. All right.

One other thing I want to share is we're

going to be using two 2D or 3D assets

and one 2D asset that were made by

community artists that shared their

information and their their skills,

their art with a a creative common

license. So, they're okay to use, but I

wanted to call them out anyway. We have

Fairy Floss Lord 7 Ephuren. I'm

[laughter] sorry if I don't have that

right. I'm really trying. That gave us

our basketball. And Demetrius uh is the

one that gave us our wood. You'll see

all of this in a second. I just wanted

to call this out. Now, how did I get to

this? I will say SketchFab is a place

that shares 3D assets that artists go

out and share this stuff. I did not know

of this before. So this really takes us

to one of the first most fundamental

aspects. You need to use chat GPT or now

that you have an anthropic subscription.

Congratulations by the way. You can use

claude. Either one. They're perfectly

fine. So I knew I was going to build a

little application for us that used a

little bit of 3D in it. Not so much that

I had to worry about it, but I actually

came to Chat GPT and said, "Okay, I am

going to use 3D. I can either let the

agent build it, which is a very

complicated process. whether or not they

get that right, it's pretty low low

likelihood. So, let me see if there are

some assets that I might be able to use

for anything. Even if you don't know

that there would be assets, this is the

kind of question you should be using.

And I will share with you the first time

I built this application, I did not do

this step and I just pushed forward into

the model and said, "Build me these

things." And it was a complete face

plant. It was terrible. Okay, I wanted

to tell you about that face plant. Like,

it was honestly real. So, I had built

this uh during a previous runthrough of

the video. I got to a point to where I

thought, "Okay, I'm going to show off

what it built." And it really was

tremendously bad. It was trying to draw

a net. It was trying to draw a hoop. The

ball was too big. Nothing had collision

detectors. It was a train wreck. And I

mean, like irreoverable. And I thought I

was going to have to give up. I want to

say this because I think a lot of you

feel this as well. Every now and then

you you feel like maybe my ambition was

too big or maybe I don't know what I'm

doing or I don't know how to ask for

these kinds of things. I run into this

as well. I use these things all of the

time and I hit this what I call the 70%

wall. It gets you to a place and you

feel like I thought you were going to

handle most of this. Now I don't know

what to do because where we're at I feel

incapable of moving it forward. In this

case my answer was go back to chat GPT.

I scrapped it all and I had a

conversation and said, "Look, here's

what I'm trying to do. Here's what I

think went wrong or why it might be

difficult for the thing to just go out

and build it. What might some options

be?" And ChatGpt was the one that said,

"Oh, great. You might be looking for

some assets. There's places that you can

get those." And I said, "Where where

might I get those? Bob's your uncle."

So, I just wanted to share that. It's

something that we all bump into and and

really I did on this project. And boy,

I've been doing this for a long time

already. Okay. And one last thing that I

want to tell you about chat GPT and it

really does take us into the process.

We're now into building. We're one step

away, but I also start in chat GPT to

have a conversation about the things

that I want. I'm looking to build

something that looks like this and feels

like this and works like this. What I

need to build is a PRD. That those

letters really matter. All of these chat

bots understand the language of PRD

these days. All it basically means is a

product requirement document that kind

of describes what you're looking for. It

doesn't necessarily go into the details

of how this is an important step to

define. So what you want to be able to

do is describe what you're actually

looking for. What does done look like?

What is good? And that's what you're

describing in one of these. And I have a

conversation. It's usually three or four

back and forths with Chad GBT. And quite

frankly, I actually record myself on

what I call a walk, which is kind of a a

brain self- brainstorming session that I

have. Those can be up to 10 minutes.

Sometimes they're 25 minutes. I take the

transcript from that and just throw it

into the chat GPT and say, "This is what

I'm thinking about. Help me refine

something down to a PRD level that I can

feed to an agent. At the end of that, we

end up with our PRD. And now it's time

to dive into the building. Let's really

dive in and take a look at how we use

the output that we would get from

something like ChatGpt." Okay, so here

we are in what is Visual Studio Code. As

I was mentioning, we have Cloud Code

open so that we can talk to Claude here.

And this is a list of the files that are

in the project. It is important to note

that inside of this docs folder, that's

where I put the initial PRD. So if we go

into this PRD, I won't go through this

here. It's checked in. Again, it's in

the repo. Everything's in the repo. This

is basically going to be a daily tracker

that uses a 3D system to kind of track

the daily elements that you've enacted

or enabled. So, as you turn something

on, a 3D interaction occurs. We'll see

all of that in just a second. But it the

nice thing about having this file here

is I can just come to Claude and then

say, take a look at the initial PRD

planned and execute. That's it. And

these words aren't necessarily all that

important. You probably could just put

in a link to the initial PRD and say

build. It doesn't really matter. It

knows what to do. But here is the next

critical element here in cloud code. If

I hit shift tab, I'm going to hit it a

couple times so that you can see me

toggling through it. It goes through

different modes. And one of those modes

is planning mode. This planning mode is

something special and I can't say enough

about it. This is essentially letting

the system think about the problem

multiple times. Get into planning mode

with Opus. Give it a good PRD. your

results will absolutely be much better.

And in fact, all of the PRDs that you'll

see that I've shared in all of those

branches, each one of the changes, I did

them in planning mode. I posed it as a

PRD to planning mode, let it think about

the problem, and then executed on it.

Okay, here we are. It came back and it

had some questions for us. It's finished

with the plan, the crate model, and

sound effects. We submit the answers,

and what it's going to do is then write

an actual technical plan. that technical

plan basically you'll see as a a series

of check items that it needs to execute

on. You can also look through the

technical plan at this point to

understand what it's going to build. I

do highly advise looking through this to

get a clue of are they going the right

direction? You should be able to eyeball

a few things from what you asked for

coming in. Does this feel like it's

hitting the mark correctly? You can

always reverse things later, but it's

better to get it right here. This is the

best place to kind of correct this

because it could create a whole bunch of

artifacts that you're going to have to

kind of reverse on some of them in the

future. That gets pretty hard. So, at

this point, you can say yes, auto accept

edits, and it'll go off and build. So,

fantastic. Let's see what it builds.

Okay. And here was the first output. Uh,

it's very interesting. If I drag and

scroll a little bit, you can see that we

have milk crates moving around inside of

them. You can see Liipuchian

basketballs, which is kind of

interesting. And if I select one of

these, oh, the basketball jumps in over

there. And if I select it again, it

leaves. Okay. And somewhere, who knows

where these basketballs are. So,

definitely has some problems, but these

are very easy to overcome. This is just

it not fully understanding what we were

going for or the 3D world that it put

together. This is easy stuff, but it's

actually a little bit difficult to feed

this kind of visual information into the

model. So, let's talk about how do I

tell the model what's going on? I can't

just say, "Hey, fix it." If you've done

this, you know, this is the moment you

kind of go, "Uh, they're not going into

the right baskets." That is not nearly

enough context. So, let me show you a

trick. So, the changes that we want, I

created another PRD. Honestly, again,

using chat GBT just by saying, "Here's

what my problems are, XYZ." I sent all

of my screenshots I'll show you in a

second to chat JBT as well and said

here's the problems I'm experiencing

blah blah blah. Can you write a simple

PRD that will describe what I'm actually

trying to ask for? And it does a very

good job of baking it down. Okay. So the

real trick is these models cannot see

anything. So what did I do? I took a

bunch of screenshots, four different

screenshots and illustrated them a

little bit and annotated them by saying

I clicked here and the ball entered

here. And then you can notice that the

light is lit up on the basket. That was

one of the things that I asked for. So

that's correct. But if we go to

screenshot two, you'll notice the light

stays where it is as we scroll to the

side. And so I just said the light

stayed when we scrolled. Okay. You go to

screenshot three. I'm also illustrating

that the ball is too small. And I go to

screenshot four. And I talk about the

different month labels and the day

labels that are missing. All of that is

annotated very clearly in individual

screenshots. Don't overload one

screenshot with 50 different things. It

will not remember all of them. So this

way it can work through them. Remember

this claude code agentic system knows

how to go through things progressively.

So it will hit each one of these. I also

asked chat GPT. I don't usually do this

but I just decided I'm putting so many

changes in at once. What I wanted to do

is get chat GPT to say can you create me

a little markdown file that represents

all of the changes that I asked for in

those screenshots. Don't don't don't try

to understand them and give me an actual

definition of them. Just say what I said

in them basically. So it's kind of a

transcript of what I annotated inside of

these screenshots. The reason I did that

is I wanted to make sure that that the

building model or the planning model in

fact in this case had an opportunity to

say oh I might have forgotten that I

read screenshot one two or three and

while I'm reading four it can easily get

lost in that way. But with text it has

kind of a progressive. You can see image

one, image two, image three, image four

and blocks of information that it can

keep track of very well. This keeps it

in line. Very often I'm trying to give

it guidance. You have to understand

where these systems don't do incredibly

well. And as you start learning small

tricks on, okay, if you're not going to

keep track of everything, I can give it

to you in a way that's more trackable. I

can give you a series of check items

that you have to check off as you work

or numbered lists that you know you can

read through or at the end I can ask you

can you enumerate by using the numbers

that I used in the document what you're

going to do for each one of those that

will force it to check its work. So I

just want to call out that sometimes the

real important thing that you're

learning here is not how to code with

these systems like I said in the

beginning it's how to work with these

systems. That muscle, that skill, that

absolutely is the thing I'm talking

about here. I am not talking about how

to code better. I am absolutely saying

use coding as a foil because it can be

fun and really expansive itself. But use

it as a foil to help yourself learn how

to work with these models and these

systems. This is applicable everywhere.

How we deal with images or video or any

other generative these days. All of

these techniques matter. Okay. And at

this point, just like we did last time,

I will shift tab into planning mode and

say look at changes one. And inside of

that document, all of the screenshots

are referenced, but I would here

generally also say and all other files

in the one changes folder, whatever. And

think about it, plan and execute, and

we're off. And literally, that's all

that needs to be done. So if you're

following along, you can kind of just

say, hey, take a look at this PRD and

move on. And it's possible that I didn't

get these branches exactly right when I

saved them. I happen to be on the 03

branch to be able to run these changes

correctly. I don't know if they're on

the O2 branch or not. I apologize for

that. That would be my mistake. But in

any case, step forward and see if you

can find everything. I guarantee you by

the end they're all in there. So, at

least you can go to the last one. All

right. So, I want to talk to you about

two remaining concepts. They're both

pretty simple, but I think they're

important enough for you to understand

how I found this process. how I

understand what makes these models work

better because that's what you're going

to want to tease out. You don't want to

have to come back to people like me to

figure out what to do next or how to

make them fly. You need to learn how to

look at these things yourself. Things

have changed. If you haven't looked at

this deeply in the last 2 months,

everything you know about it is probably

actually old. So 6 months ago, we were

saying you're going to be deeply

involved. 3 months ago, we were still

saying you really need to kind of be

involved. We're moving away from that as

well. Now, I'm telling you, you need to

figure out how to unplug yourself. Your

goal is not to learn how to code. Your

goal is not to learn how to build a

website. You do have to take an active

role. Just watching and listening. That

is not the same these days. It takes a

long long long time, I believe, to get

really versed on how to work with these

systems. You have to understand that and

be willing to put the work in. Expect it

to be months. The second one that I

wanted to talk about is more functional.

We have something on the screen here.

Very simple. This was our initial build

if you recall. And if you look up, of

course, this is the whole plan that it

worked through and everything else.

There's a lot of data here. This is the

context. I just want to talk to you

about this one really briefly because I

think all of us are aware, but you need

to hear it. Clear your context. When

you're at a point of saying, "Okay, we

finished this step. I'm moving on to the

next step." and it does not need a whole

bunch of information from what I was

just doing or it will be able to

discover it again. I have faith that

it'll look around and discover it, then

do slashcle. So that's the other slash

command that of course you're really

going to want to know is /clear. It is

basically just new chat on chatgbt or

something like that. But it is

incredibly important. It has been shown

that the bigger context you have, the

more filled context you have, the lower

in performance the model has. And that

leads me to kind of the very last thing

I wanted to say and it's it's pretty

small but it's worth hearing. These

models are not going to remember you.

You know that from Chad GPT. They are

now remembering us a little bit here and

there. They kind of feel like they are

starting to understand us. That story

will be told and get better. But down in

the coding environment, it feels like

y'all are so smart and doing so much

that I should be able to rely on you to

understand that last time I asked you to

run the tests and you ran them. I

cleared my context and this time you're

not running the tests with that. There

is one special file and it is a very

important file. If you do a slash innit

and I know I said we didn't need any

more slash commands. When you do a slash

innit what it will do is it will look

through the project and it will create a

new file called clawed md in your root

directory. So it's just going to look

around figure out what the project is

about and write a file. You don't have

to really worry about what's in there.

Feel free to remove it. There's nothing

magical. more important is that you have

a file here cl called claude MD and if

you want those tests to run every time

that's the file the the contents of that

file will be put into every new chat

that you have. So, if you have things

that you keep thinking, gosh, why do you

keep forgetting that I don't want to use

that kind of math? I want to use this

kind of math. Or why do you keep putting

the colors inside of the markup? I want

it out in my style system. That's great.

That's what this file is literally for.

That's all I wanted to share with you.

This is a very, very important concept.

It's used across almost everything these

days that builds. So, they will have

their own kind of what we call memory

file. It's an important one. Okay. After

all those updates, the PRDS, the

screenshots, planning mode sessions,

let's take a look at what we end up with

this application. And by the way, this

is what's checked in. So if you just use

main and you run the application, you'll

see this, but I'll say work your way

backwards and see if you can get back to

this point or can you take it elsewhere?

So this is the ending application that

we have. We have the basketballs that

know that they're the right size and

they know the basket that they're

supposed to go into and we try to get

the basketball to go in and sometimes it

goes in. So, just part of the fun, but

that's okay. We've got streaks and goals

and all kinds of other things and has

all kinds of uh normal kind of 3D

interactions in it. This is a fun little

version of application that you can

build for a daily tracker or something

else. take this further, turn it into a

game, whatever you might like. You can

see the steps forward in how it was

built. That's the intention. Hopefully,

this has helped you understand how I go

through building something to this

point, cuz really, you have almost all

of the steps that I took to get to here

now. Okay, let me show you one of the

coolest things I built that really

helped. As I talked about earlier, these

systems can't see what you're doing.

They have no eyes. They don't take in

video. That's not one of their

modalities at this point. If you're

doing animations or anything that

quickly changes state or you're seeing

things flop back and forth and it says,

"No, I looked at it and it's fine or I

think I've fixed it." It's probably the

fact that it took one picture and it

took another picture if it even bothered

to look and it can't tell the delta. So

here I'm sharing the screen and what I

did in this case because this is 3D.

It's actually a completely different and

more challenging problem that it doesn't

always actually understand where cameras

are pointing or where lights are

pointing or oriented. All of these

things are a challenge. And so I'm

showing you the final product here for a

reason. Each one of these has a light

here right above the basket itself when

there's a ball in it. And the light is

missing when it's not in it. There's

also a spotlight up here somewhere

that's pointing down on it that's kind

of in a forward- facing orientation to

create this spotlight cast that will

kind of blend together. This was very

difficult to do. Not to mention the

balls themselves were not falling into

the basket correctly. They were falling

out and just rolling all over the place.

And so the collision detector was off.

So what I decided to do, I'm going to

hit the D key. You can do this as well

on the repository near the end. It's

maybe at the six or so spot uh

checkpoint where I started adding this

debug panel. So, I'm going to hit the D

key and you'll see the debug panel comes

up. So, the debug panel on the left is

all about physics so that we can change

physical elements and what's on the

right is all about the lights. So, for

example, I can pull down the intensity

of certain lights. I can change their

angles. I can change the ball radius.

Let's create the create the crate depth

to be bigger or smaller. You can see

that bounding box is shrinking backwards

and forwards or even the crate height.

So if we make the crate height much

shorter and then click on one, you can

see that the ball can escape very easily

because it's actually only interacting

with those yellow walls there. So that's

just this is a a technique that I use

from time to time where if I think that

what it's trying to set up or toggle or

mess around with has kind of particular

values it can't see very well, the model

being it. I then will say let's create a

debug panel that has all of these config

values pulled out. And what is cool

about this. So now that we've done this,

you can see I updated our bounciness. Um

once I get the values set, what I've

told it to do is to create a JSON

structure down here that I can copy and

just give it back to the model and it

will take all of these values and spread

them out into the locations they need to

go. Hopefully they're all in one config

property. Same thing for the light. As I

change these lights, I can just copy

this at the end and say, "Good. I set

this up. It looks good." So, I'm just

advocating for help yourself, right? You

are there as an active partner in some

cases when it gets into debug places

from a performance standpoint, other

things like that. You will need to

become active and creative on the

solutions that you come up with. I just

wanted to share this because I'm not

sure that most people would think to go

this route and it's really been helpful.

All right. So, in the end, all I really

wanted to leave you with is it is

important to use the best tools right

now. If you're down on a lower level

model of any kind and wondering how do I

get the best performance out of this,

all these people are raving about it and

saying it's so great, but that's not my

experience. I can guarantee you it's

because you're not using the best tools.

If you're not using cloud code, you're

not using planning mode, and you're not

using Opus. It is my personal feeling

right now, you are absolutely not using

the best tools that we have at the

moment. Like I said, next month we'll

see. I hope there's new ones coming out,

but really this thing is several steps

ahead of where it was only a few months

ago. So, I highly advise giving it a

shot. All right, that was a lot. I

really appreciate you coming along for a

ride on this one, and I'll see you in

the next one.