Your AI is literally suffocating right

now. I'm not even joking. Let me show you

something

that's actually going to change how you

use Claude Code and Cursor. If you've been

using

ChatGPT as well, any AI system will feel

sometimes it's getting dumber after just

10 minutes

of coding. And what I've basically done

is created this nice little visualizer to

kind

of show us how this stuff works. All you

do is just run a simple query and you have

Like a 200k token context window

Agentagine You target to create more

and then it does some tool calling as

well and then comes back and actually does

some more thinking to make sure it's the

right thing and then produces some code

output.

And so as it does code output, depending

on whatever type of code you're doing,

let's just say it puts a bunch of tokens

here.

Then you say, oh, can you go ahead and do

this type of thing here? It's , yeah, go

ahead. Cool.

And can you also grab this other files

and do these other things?

So now that it's grabbing files from your

code base, that starts to fill up.

and so on.

And then it starts to fill up.

And so the agent does another 8,000

tokens worth of thinking, does a couple of

tool calls

and stuff and does some more code output.

And then it starts to output the code or

output the changes.

Now this token window starts to actually

fill up pretty fast.

And we're basically at 96,000 tokens out

of 200.

And you'll probably see a little thing

that says , yeah, you're using 48%.

You're , okay, what if I just want to do

another quick request?

So then you start to grab some more stuff

from your database.

You start to now connect all the front

end and the back end.

Your agent does some thinking.

You say, I wanted to think ultra hard.

So then you basically pass in the most

amount of tokens this.

And you're just for three different, you

know, if you're thinking ultra hard,

you're going

to use 32,000 tokens.

So that's 8, 16, 24, and then one more

this.

And then it does some tool calls to grab

some additional things.

And then it's going to try to do the code

output and boom, just that, you know,

With a couple of more requests, now

you're basically hitting almost 98% right

at the very end.

And as you can kind of see, this type of

workflow starts to add up super fast, and

we've only done a couple of turns.

And this is the point sometimes where

maybe Claude Code or Claude Sonnet or

maybe even Cursor itself will start to

kind of say,

Hey, we're going to roll the context

window. You're , dude, I just started the

conversation. Why is the AI getting

dumber?

and it's maybe it's not that the AI is

getting dumber but I think it's more that

the

context windows are kind of limited here

and that a lot of these apps in the

background are trying

to do different things for you to try to

maybe not take up so much space and on top

of that

there's this interesting study that I

want to go ahead and show you this was

produced from ChromaDB

and in this study they did all this is a

really lengthy report and I definitely

would

recommend you check it out so they did

this comparison called the needle in a

haystack

and so on.

One of these things is you can have these

large context windows from Google

or these other companies, millions of

tokens.

You're , okay, if I throw everything in

there, it should be okay.

The thing that I learned about this and

what I really want to share with you that

I don't

hear many people talking about other than

in these really deep reports is that when

you have a single question and you have a

single thing for it to search at, that

accuracy

is extremely high.

And so what that means for you and your

codebase is that maybe if you have one

specific

narrow task and you have this giant

codebase, maybe just giving it that one

thing is going

to make it much more accurate.

And that actual thesis was proven through

their research and this paper here.

And so whenever you have basically

there's this similarity bias and as this

type

of thing kind of goes out, it starts to

get a little more distracted when you

start to

give it more tasks.

This is kind of where we're going to

approach this thing called subagents and

kind of why this is such a big movement in

this realm.

It's not only about the context window,

but it's also about accuracy and achieving

the nice things that you need to do in

your code.

So we can kind of unpack this.

This is going to be a little bit more

theory type of things, but I kind of want

to do this a little bit more with visuals

because there's a lot to unpack here.

and I feel as a live stream I can answer

some of your questions back and forth

maybe some of the

stuff you've been running into and I'm

hoping that this can sort of influence and

change a little bit

of your behavior for how you prompt the

models and then how you can get the most

out of them so that

that way you could just be really buzzing

and kind of going I feel sometimes a lot

of people it's

really easy to use an off-the-shelf

framework or an off-the-shelf MCP that

does all these tool

So, what you'll actually end up

discovering is that maybe some of this

type of stuff is abstracting it a little

too much for you, and maybe it's kind of

going off the deep end kind of a big old

machine gun and just spraying and praying,

as the old gamers used to say.

So, in this Haystack question, obviously,

this is kind of what they're saying is, if

there's a bunch of irrelevant context,

this is the needle.

and whenever you have something called a

distractor, something that sounds similar,

this is kind of where it starts to kind of

get thrown off a little bit.

So that's sometimes maybe some of your

queries, you're not really clear at the

very beginning about what you want.

And then if you kind of give two similar

things, that's kind of where it starts to

say, maybe he wants this one and maybe he

also wants this.

And that if you're as confused, the AI

will also just be as confused for the task

when you hand that off.

So they say here, the best writing advice

I got from my college classmate was to

write every week.

and then here it says the the I think the

best writing tip I received from my

college professor was to write every day.

See this is the needle and this is

the distractor. All in the same paragraph

when you asked you know what's

the advice I got from my college

professor it's gonna see these two

things. So it's gonna be oh I've seen

that in the training set I see

these two things and you're wait which

one should I give back as the

answer because the college professor

they're asking about the college

professor. I'm well they said I should

write every week but also the

and so on.

So, you know, that's the best writing tip

that I should write every day.

And so that's, you know, that type of

what we call hallucinations can happen.

And this is kind of why it's a little bit

more important.

And this is kind of why I want to share

this type of information is because not

only did

they do the study, they actually went

super ham and they're , how many

distractors

can we have?

And what's the actual accuracy pool,

right?

they did this whole study and it's just

blown my mind right now.

So check this out.

If you have basically what they call four

distractors, the input token length,

when you start to kind of add it in, look

how fast it degrades as far as accuracy.

It goes crazy if you have four or more

distractors.

And so think about this in your own code

base, think of them as distractors, right?

Whenever you have code that's repeated on

all these different components and you

say,

go and update this type of thing, you're

, and you just YOLO it, it's going to

update

and so on.

And then you have some type of layout,

you know, you're supposed to kind of sort

of have

some hierarchy and some components you

reuse.

But it's , oh, I'm going to make another

progress bar over here in the admin panel

because a user requested it.

And it's , well, I have a whole directory

of components that I use from ShadCN.

You should just initiate one of those and

make it there.

AI will be really happy to fulfill your

request and just starts to actually code

up the button

and everything that it wants to do in

that separate area for the admin panel. So

now you

actually have the user side and you have

the admin side that have literally a copy

of each other's

code. And this is what I call a

distractor. And this is how crazy it is

and how important it is.

And Chroma is actually showing, it's one,

the accuracy stays really high if you just

have one specific thing you're trying to

achieve and you add two and then you add

four. And then

Think about what people are saying for

AI. This is kind of why people make that

term AI slop,

is because as an engineer, you'll kind of

start to see these patterns develop.

You're ,

well, the AI isn't really good because it

just keeps rewriting code. And I feel most

of

the time I'm doing as an engineer is ,

I'm writing more rules to guide the model

right now.

There is some good news out of this.

Every single year, these models have been

getting more and more

intelligent, so you don't have to guide

them as much with as many rules than I

previously did

and so on.

So, we're, models are probably going to

get smarter and the tooling is going to

get a

lot better to support these types of

things so that , you don't have to worry

about

this too much, but as of right now, the

tooling that we have that I use a lot is

Claude Code

and subagents is one of those types of

keys.

So I want to show you what that looks in

my content visualizer because I think

that's

going to be the key here for kicking off

subagents and trying to basically narrow

that scope

and so on.

So, you can have very defined things that

you do in your codebase, generating code,

generating UI components, working with

your backend, understanding how those

types of

things flow together.

So when I hit reset and I go down, when I

basically do a code generation task, you

can

kind of see how many tokens you take up,

right?

44%.

But then now when you spawn a subagent

down below, you'll actually see this type

of thing

start to take place.

And so you say, I'm going to go ahead and

add a request.

Now that subagent starts to take on that

request.

And then if there's another concern that

that request takes on, you just do add

request

as well.

And you can kind of see how the subagents

really start to unlock this whole type of

thing.

So you're basically getting this main

request up here, and then you're handing

off these

individual components, right?

So in this case, in my situation, I

started to make one specifically around a

design system.

And then I'm going to now try to make a

subagent for my convex database, because

the convex

and so on.

So, I want to make sure that when I do

make a feature request that it does pick

these

types of rules up so I can actually learn

how to interact with my codebase in

smaller

segments so that I can keep that context

window really tightly focused and that

just ensures

higher accuracy.

Because I'm even noticing that some of

the code generation, even though I have

these

rules files, it may not pick them up and

it may not pick them up into those

sub-agents.

So, I want to kind of share this as a

primer for y'all because I feel this type

of

and so on.

So, you know, that visualization hasn't

really been talked about too much and it's

still

very confusing because I see lots of

graphs and things.

But the best thing is literally just this

type of thing.

You just do a tool call, agent does some

thinking, you do some more input and then

the code,

you know, comes out and so forth.

So this stuff adds up pretty fast, as you

can kind of see just from a 200k token

context window.

And that's why it's important to start

new chats.

When you start a new chat and you want to

do new code generation task, you know,

this

is the type of thing you'll do.

and so forth.

And that type of stuff happens.

Now with these thinking models, I'll show

you real quick.

They will basically do after the user

input, they'll do some thinking.

You may generate some code and the agent

will come back to do some more thinking to

understand what it just did.

And that's sort of the power of Claude

Code and this is the power of the new

Sonnet models with the additional thinking

and the tool calling.

and this is also very transparent in

Cursor as well.

So let me just go ahead and catch up with

the comments

because I know a lot of people have been

trying to catch up here

and wondering how everyone's doing here.

So, alright, so let's see what's going

on.

I need to dial in my subagents more

around their tool calling access.

I've given them a set to inherit all AI

tools

even though I only have two MCPs enabled.

Eating tokens.

Yeah, that's what's up, man.

Right, guess what? After some stress and

work, I need to recharge seeing your fire

cooking apps and producing some sharing

contents every day.

Thank you so much, Alan, for becoming a

member and also kind of catching up here.

What if we were to tell the AI to

reevaluate its solution? Does it still

hallucinate?

Yes it can still hallucinate especially

if you have all of this conversation where

it can be thrown off So in this context

the conversation is remember that needle

in the haystack stuff

So if you're continuing the same

conversation, if you take some of that

context and pass

it on into a new chat, or if you're just

continuing on, you're going to keep adding

these distractors.

And these distractors in your chat, if

you're telling it to fix itself, are still

going

to be in there.

and so you start to basically for every

single turn in the conversation so that's

every single

hey go fix itself okay just change this

one not this one so you've seen that

vibe coding meme this is basically what's

happening is you're creating more

opportunities

for the accuracy to continue to go down

so the further you get in your chat the

more

it keeps actually going down in accuracy

and so at that point you just got to

create a

new chat and pretty much just focus again

and try to have it rebuild its context

because

Now you're going to basically just be at

this point you're just gonna be you're

doing

some complex debugging task right you're

doing this type of thing and you're just

okay user

input this the agent does thinking it

does some tool calling some more code

output you're

no try again think harder and it's okay

at this point bro anything that you've

done

before you've just introduced massive rot

and that's basically where you're at

you're just

just crazy crazy amounts of rot but it's

important to know that why why that maybe

Maybe that's happening and then at every

type of conversation that what I call a

distractor is going to keep throwing the

language model off until you basically

start a new window.

In Claude Code you can hit escape twice.

So hitting escape twice in Claude Code

will basically let you go back to that

conversation and truncate things out.

But I would just sometimes recommend just

starting a new window. Right now Claude

Code doesn't have anything visually that

should show you how much of the context

window you're taking up, which I feel is

super important.

Cursor has recently just introduced that

feature into their agent and it's not I

don't think it's live for everyone yet

but you can kind of see down below it so

shows context usage for 83% so that is

super duper valuable because once you

start passing that mark I'll show you

really quick in in chroma's paper here

there's you know this is a performance

over three models with 500

over 500k tokens so this is those 500k

tokens once you start to pass that

and so on.

But the accuracy rate, once you start to

go up in the percentage, right?

50% is your golden zone, really.

But when you have more distractors, that

goes down really, really fast.

So you're talking about 30%, right?

So if you have one dedicated thing that

you're saying you're going to do, you're ,

in this admin panel, I want you to do

this specific thing on this component, on

this,

and just add a for loop or something,

it's going to be really, really good

because it's

just that one thing, right?

and so on.

So, that's just what it is.

But once you start to continue the

conversation, it hasn't been going well,

fix itself and

try to reimagine and rethink everything

that you possibly can so that you can be

the best

version of you, AI, and then that's not

going to help.

That's actually not going to help it at

all.

It's been really good to back up some of

the vibes that we've been feeling with

actual

research to better understand what this

whole vibe shift is really about, because

I feel

that's what's been happening now.

Yo, how's it going, Vlad? Thank you so

much for joining on the stream. I've

started playing

around with Gemini CLI and not going to

lie, I'm impressed how good it is at

reinforcing

or refactoring. Oh, I'll have to try that

out again. I'm using Google Code Assist

plan

at 20 something dollars a month. Yeah,

I'm going to have to check that out again

because

I do want to kind of compare some of the

notes as well. How does it compare to

Claude CLI?

Yeah, I'm just curious as well. I took

the Cursor Ultra package, but in 20 days

limit

I was not even using Opus. Wow, lots of

agentic coding probably. I'm just really

curious how that works.

So are you able to select other models?

Would love to know more about that

workflow and kind of what's going on.

Yeah. OK, that's interesting. Well,

there's a lot kind of going on here.

Have you considered developing any local

Mac apps too, or are you mostly focused on

developing on the web for mobile apps?

MobileApps. So right now I'm thinking

about doing a Mac app and a Swift app, but

at the same time,

, it's gonna be more on demand because

right now with my app RayTranscribes, for

those who

aren't familiar, this is basically my app

right now. So my app is basically to just

transcribe

a lot of my videos and I have a lot of

videos and a lot of live streams and stuff

so they can be

hours and hours long. I found a hack.

It's basically you can get discovered

literally

Just by uploading your transcripts and

putting them into YouTube.

Because when YouTube processes them, the

keywords that you say

will get picked up when people search for

them. So if I do Claude Limits,

I bet you my video shows up from

yesterday or something that. Yep, see?

There it is.

So it's the second video, right? And I

could even do a private mode or something.

and then you'll see Claude Limits shows

up it's the second video and so in this

type of

thing it's my video is only 21 hours ago

and that some of the top results and

because of you know having those

transcripts in there it's super important

but I also have

timestamps in there as well so that's

super nice because Google likes to chunk

out the

timestamps to help with their AI

overviews they basically you're helping

their rankings and so

So any type of reinforcement that you do,

you basically start to get ranked higher.

It's sort of a secret sauce.

I need to write an actual blog post about

it.

But my Transcriber app basically does

this type of thing.

And right now, this is extremely

efficient.

You can have your own custom dictionary.

So I have these different words that need

to be in there so that Claude Code, o3 are

actually in the transcripts specifically

for these type of keyword things.

And so, yeah, when it's processing the

audio, it's doing its stuff here.

So this is a multi-hour stream.

I think it's a four hour, three hour live

stream.

So it does this stuff pretty fast.

And you can see all these transcripts

that you're able to kind of do the

things.

So I've done hours and hours and hours of

stuff.

And this is basically my little app here.

And I think if I do want to do this so

that,

I'm almost done with my refactor, by the

way.

So this started out as a bunch of slop

and this is kind of why I started to get

around this context rot is because

when I'm done with the refactor,

I can now make the iOS app and because

I'm using a database by the name of

Convex, for those who

don't know you may hear this a lot but I

just I'm a huge advocate of this company

because they're

just amazing. They have really great

database software. So one of the things

that they have

here is fix the bug and set complete, say

true. Okay so it's a TypeScript database

so

you literally you're just basically

playing around and this is the actual

dashboard for the database.

So as you make changes here it's

basically happening in real time on the

back end. So

As you add these different things, it's

all happening. So someone on their iPhone

can be

uploading, doing a transcript. Someone on

the web can be doing that. Someone on an

Android phone can

be doing that. And so this is kind of why

I wanted to make sure that everything was

kind of settled

on my backend side. And so when I build a

Swift app, I could just hit this database.

I have a

clerk for my authentication. Clerk has an

SDK on the iOS side. Everything would just

be a smooth

experience and it would just be super

duper fast to hit my endpoint, send all

that stuff up,

and if they're on the web, because I know

a lot of people who do content creators,

they have stuff on the phone, right? They

have a short they just recorded. They want

to

be able to do it from their iOS app and

then they just want to be on the web and

just copy and paste

that transcript and put it into YouTube

or something that or whatever other

workflow

they do and they can do it because of the

way I'm setting everything up. So yeah,

long story short,

that's kind of the way I was going with

this. I have yet to try the latest

versions of Swift that

and so on.

So, I'll be spending some time probably

in the next couple of months studying

that, figuring

out what's going on with all the latest

APIs and animations, because they tend to

settle

the dust for that right before the fall.

So yeah, that's kind of why I've been

waiting a little bit.

But yeah, I don't know.

I was thinking about doing some iOS

stuff, but I've been kind of trying to

stay away

from it for now.

I don't want to get in trouble.

What's your marketing strategy for

RayTranscribes?

I think right now is just literally just

trying to code with it and then tell

people about it.

I want to reach out to more content

creators to try the app. And if you want

to try the app right

now, you can actually try it with the

code RayCooks. So if you put in the code

RayCooks,

what that will enable you to do is when

you go here to Ray Transcribes in my app,

let me just put this up again as well. So

if you start any of the plans today, go

start today,

That'll give you basically $15 off. So

you can get into the beta for $5 basically

for the first month or just get into the

starter plan for free for the first month

and you can start cooking.

And I give a very generous amount of

minutes. I mean, there's 6,000 minutes for

the beta and 3,000, which is way more than

enough than anyone has.

And then, yes, obviously you have all the

cool stuff that you have here and so

forth. So yeah. Oh my gosh, check it out.

Dennis just became a member.

Thank you so much, Dennis, for becoming a

member.

If you want to become a member, you can

secretly slide in through my Discord.

I still have a member post up.

And right now what's happening is

basically people are able to slide into my

Discord for $6.99 right now.

It's kind of a secret perk, but it's in

my member post.

If you go to my member post, there's some

instructions there that you can send to

get into the Discord.

There's now over 100-plus people in my

Discord because I finally reached my first

100 members.

and so my promise then was to double the

price so that you know we have this

exclusivity for folks

who are early. So if you're watching this

right now it's still early so this is your

time to sign

up. Once the price doubles I'm going to

use I'm basically going to use Polar.sh

and so Polar

has a nice Discord integration that you

can basically hook up and it's super easy

to code

and a lot of the community stuff will be

built using kind of Polar and and these

types of

things and I'll be showing you this stuff

along the way which would be really really

cool. So

So yeah, that's kind of where that's at.

Ray Transcribe is built with Next.js, or

do you use another?

No, Next.js.

So I built my app, Ray Transcribes, with

Next.js.

And yeah, it's basically using a route

handler.

I use the API from there.

And I have some client-side work that I

do.

So I do some processing on the client

side.

And I'm transitioning that to Vercel so

that everything is going to be handled

through Convex.

So a lot of the pre-processing stuff will

be kind of minimal now on the client side

and it's going to move a lot of it to the

back end side.

So a lot more updates on that. But yeah,

Next.js as far as the thing, ShadCN for

some of the frameworks.

I think I have, let's see, Tailwind CSS

v4 so that I have this cool tokenizing

system.

I'm using Clerk for authentication. The

back end database is Convex.

By the way, when I launched the app, I

had no database.

It was literally a v0 slap-on, , hey, I

want to get some paid users. I sent Stripe

links.

People signed up with the Stripe links. I

went into Clerk and I just added their

email

as part of their getting started. I was

doing it all by hand. And it wasn't until

I was on

vacation in Croatia and I lost three

users who paid. Three users paid for my

app and I couldn't

and so on.

I started to get to them in a couple of

hours because I was remote and in a

different time

zone and I was , man, I didn't want to

set up the Stripe database, all that

stuff.

But I'm glad I did it actually because

now it's much more scalable and I can

actually

do more things, which I plan to do

anyways, but it's starting to actually get

traction,

which is cool.

So yeah, I'm going to reach out to more

content creators and people who are , if

you do

You got a lot of great minutes for the

costs and I'm kind of wanted to pass it on

as well.

Hi from France you make DiscoverCondex

thank you I use it for an upcoming

personal project I and I think you get I

don know 20 free projects Something

ridiculously cool for free So it great

I've been using your transcribed

platform, Ray. Loving it. How is the

architecture behind it? Using Groq for

Faster Whisper?

Yes, Groq is one piece of the story. So

part of the reason why I was using Groq is

because it's just super duper fast.

Two is I was actually setting up this

cool thing and I prototyped it and that's

kind of how Ray Transcribes got started.

I wanted it to be even faster than any

other transcriber that's out there, and I

kind of broke the record because I set up

a web socket, and I would literally stream

everything as if it was being processed

and being sent up in chunks to be

processed.

And so I had this cool workflow that I

don't think anyone really optimized, but

as a nerd, I was , this is going to be fun

to do.

So, yeah, it's a pretty nice, simple type

of thing that you can set up.

and that's one piece of it but then I'm

moving into more features now so I

realized that speed isn't really the big

thing it's more about just kind of

workflows so I can do more features if

I'm able to maybe fast isn't

really the big deal is more of just about

getting that transcript and

then doing other cool things with it

adding timestamps maybe having some

stuff where it writes you and you know

collecting these types of things to

I was trying to set up a personal project

with Groq but did not work streaming to

Groq.

Oh, you may want to check out the AISDK

from Vercel. That's one of them.

The other thing too, I think, is with

Groq there's a limit. You have a 25

megabyte limit to stream a file up.

So you'll have to do some chunking with

the file. That took me a long time to

figure out.

and then once I figured it out then I

kind of understood what to do from there

but basically you have to keep track of

those files that you chunk up and then

when they come back down for the results

then you'll be able to actually put

them together as one type of thing it

sounds more complicated but it took a

long time to figure out I guess I don't

know whenever I try to start a web

project all these tools generate and

suggest super base lovable v0 is there

any tool that use convex to let you

connect the way they integrate with

and SuperBase. I think the best tool

right now is called Chef from Convex. So

chef.convex.dev. I'm

going to put this into the chat here and

I'm going to show you real quick. So if I

do chef.convex.dev

I'll show you. And so this is basically

their Vibe code version. I think they

forked it from Bolt

and you can sign in with your thing and

you can you can make a Slack clone,

Instagram clone,

Notion clone, everything you would

normally do and it hooks it up right into

your database.

They have their own auth stack, so Convex

has their own Convex auth. You can later

hook that up to Clerc, which I love, and

it allows you to do way more, in my

opinion.

So yeah, if I sign in, I should be able

to sign in. Cool. Yeah, so yeah, that's

kind of Chef. I might do some videos on

Chef, because I do this.

But actually, I code so much in Cursor

and Claude Code that I just take the rules

files and I just find I have more, you

know, I could do more with it and so forth

that.

There's also a template that you can use

to get started.

So GitHub, I forked one and I want to

show you this and I'll put this in the

chat so that you guys can have it too.

So one of them, it's called, what is it

called?

I based it off this.

It is this one.

So this is from this guy.

This guy is elite.

and so on.

And he basically made the Elite Starter

Kit.

And so go ahead and click this one.

And so this one actually has everything

set up.

And this is what we based off the...

I have a four-hour livestream for members

only, and we started with this actual

repo.

So we started this repo.

I walk you through getting this stuff set

up with Qlirk, getting the secret keys

from

Qlirk input in, getting the convex keys,

configuring those two, configuring a

webhook.

and so on.

But afterwards, what we actually got out

of it was the Stripe stuff.

So you can actually sign up with Stripe

and it's actually built right in.

It syncs to the database.

And then it has a dashboard which you can

actually use to, you know, this logs in

and

all that stuff that you can use to kind

of just add your features and various

things.

So yeah, this is a great starter kit.

This is pretty much my stack as well.

And when I saw this, I was , bro, this is

exactly my stack.

I want to keep building stuff off of

this.

and so I basically forked this and I'm

going to update my own and add my own

rules and everything that that I normally

would use.

And then I'm going to try to see if I can

make my own fork have my actual Claude

Code subagents, the design enforcer.

I want to have it so that has this convex

rules type of thing. So my subagents will

be more accurate because the goal here is

really for me to just start a repo,

not have to create these rules files all

the times. And I have this nice groundwork

to start from and just start to just say,

OK, I want to add this feature.

I want to do this. Okay, I want to

connect to the new images API and just add

something so that it generates images ,

you know, what I was doing for my

thumbnail.

So I want to do stuff that and I want to

I need to have a good base set.

And so this would have it all hooked up

and I'm just I just go.

So that's the end goal.

It's going to start with my members first

content and then eventually I'll roll that

out to the general public.

My members are kind of helping me shape

the feedback, which is really great.

And so right now you can join in as a

member, which is super awesome.

and it's pretty busy right now in the

discord it's really nice I have these

different channels and everyone's

participating so I definitely appreciate

that I spent my night watching yeah that

video yesterday oh it's a four and

a half hour video but it actually could

have been a couple of hours but that

the last half of the video is literally

taking that 380 page PRD that we

generated there was some good prompting

that happened in there and I'll give you

the long story short but basically the

director's cut is I in the prompt I

and so on.

And I said, make sure to include the UI,

UX trees for all of the different

components

that we're going to make.

So when you're making this PRD file,

we'll be able to understand that.

So then I gave it my sort of prompt of

what I was doing, and then I gave it this

prompt to generate a PRD, and that

generated the 300 something line PRD file.

We gave that to Claude Code and Claude

Opus for 33 minutes in that live stream

generated

the code.

So that was the last half of that stream.

So the last half of that stream is

literally just us waiting for that whole

thing to generate

because we weren't really sure.

I thought it was going to generate

everything in one minute.

It generated the full-blown app from

everything.

So we took that template.

On that template, it built everything on

top of that, all these different trees and

everything.

It spawned off four sub-agents, and each

of those sub-agents went in this

procedural order

from working on the UI, working on the

backend.

It referenced all these rules.

So that was a really deep stream.

I'm pretty sure there's a just so there's

so much in there and I feel that's a

perfect members only stream.

I did go live to the public and then it

just went to members only afterwards.

So I want to do more of those at least

once a week or once a month for members.

And, you know, that was actually by

request from several members who wanted to

just figure out something to get started.

So that was our massive thing.

Don't stop believing. Became a member.

Thank you so much, man, for becoming a

member.

I really appreciate that.

You guys are the real MVP.

I really need to have a watch of the 4

hour videos this weekend.

Yeah, I'm also, I just hired an editor

because my goal is actually to get my

first 100 members

and so I achieved that goal which is

really awesome.

And so you may see Dan in the chat.

So Dan is going to be cutting up some of

the videos and I got to speak with them to

figure

out what parts we want to kind of, how we

can trim that video down to make it

really great.

My goal might be just, I might just

re-record the whole thing myself just

offline and that

way we get that video.

So I think today is July 29th and I've

been streaming every single day from the

beginning

of July 1st and the growth has been

incredible.

I've grown almost 40% of my channel.

So I started off with around 10,000 subs.

Now we're at almost 14,000 subscribers

here on YouTube, which is a huge

milestone.

This is amazing.

Thanks to people you guys who have been

joining in, but also thanks to everyone

showing

and the rest.

So, I'm just going to be coming up every

day, making this chat really fun and

participating

and helping ask questions to drive some

of the discussions, which has been great.

The second part of it is I wanted to have

my first 100 paying members, and we've

totally

blown past that.

I think we're almost at 135 or 140 paying

members, so you get access to the Discord,

and that has actually helped basically

paying for Dan to do some of the video

editing, because

I want to save some of your time as well.

So the members will get to see those

videos edited first, and then they'll be

released

and the public later on. So yeah, there's

some really good meaty topics in there.

There's a lot of stuff I want to cover.

But today's video was my sort of

a dry run of my recording that I'm going

to do for a context rot video. And so you

got to see a little bit of that. So you

got to see a little bit of so for those

who don't know, I mean, this is how

serious I take my craft is that I started

this video off with wanting to do a

context window visualizer. So I went into

v0 because I originally did this in

Eraser.io are one of these platforms

where you literally draw. And I felt it

didn't really

do it justice. So I just went into vZero

and just created the app. And it's just

much easier to have

a real time app that you can actually do

and just deploy it to production. And so

that way,

you're able to play around with this and

then I could share the link as well. So

I'm actually

going to turn this Vercel app into a real

website that you can actually go interact

with and start playing with and actually

use it to maybe teach other people about

context windows,

So hopefully, I think the best form of

learning is teaching other people. So if

you can take what I just explained to you

and explain it to someone else, then

you've really mastered it.

It's part of the reason why I to just

turn on the camera to try to explain these

things, because then it helps me solidify

these thoughts. And that's just so

awesome.

So yeah, appreciate y'all for first

chilling on the growth here. So I really

need to watch. Okay, yeah. If you design

in Claude Code, it will build it. Yeah,

yeah, yeah, for sure.

All the Claude Code vids are popular.

Yeah, I think a lot of people are trying

to figure out

kind of how to use it, how to best

utilize it. And right now, for me, I'm

taking the less is

more approach. You can see why this is

really valuable for the less is more

approach, because

any type of tool that you add, it's going

to a tool call will start to eat up the

window,

and then you add some thinking tokens,

and then you add some more code output,

and you add more

agent thinking. And yet, look at that,

we're just filling this thing up fast, you

know, and

and let's just say it spawns off all

these sub-agents.

Do you want to be that guy who's the 5%?

Who's actually eating up all the windows

and everything that and you're gonna get

that bill

from OpenAI or ChatGPT or someone,

or even Claude Code and be , yo, give it

up.

Pay up or we're not gonna give you any

more tokens.

We're , yo, yo, yo, yo, yo, man, chill,

chill, chill.

Chill, chill, chill.

It's not that in here, you know what I

mean?

But if you are one of those people, we

are here for you.

This is the AI Anonymous group and we're

here to support other people you. We think

that

you're misunderstood just the way we are

misunderstood. Yeah, I might have a little

bit of addiction to AI products, but I

think that's natural, right? This is the

new age that

we're in. So if you're in here, smash

that thumbs up, definitely drop the

comments because it's

super helpful for everyone in here to

support each other. And I really, I'm just

in love with

the community that's kind of showing up

right now. And this is , I this vibe. This

This is such a good vibe because we're

all trying to learn and share best

practices.

And I think this is a really great way to

kind of start on that track, if you know

what I mean.

So, yeah, it's been crazy exponential.

Yeah, I mean, you got something good

going.

I'm fixing to join. Yeah, for sure, man.

You'll see that AI updates itself paper,

AlphaGo.

I've seen something this where it

rewrites its own weights.

And I'm really curious to read that

paper.

I wish you grow this channel fast and get

some fun spin off more contents.

That the goal right now I think it to

kind of focus in a little bit and start to

produce some content And so far everyone

been doing a great job kind of

participating and joining in so I really

appreciate that A new member here went to

spam so I going to just go ahead and

double check my spam folder so I just

reached out to the recent people who

reached out to me and they were in that

spam folder so please forgive me for that

as well

Can you go more in depth on how to make

Grok streaming work?

I think many people are looking for this

info.

Nobody's talking about Grok, even though

it's the most insane tech for AI.

Yes. OK, cool.

You know what I'm going to do? I'm going

to do a couple of things.

I'm going to reach out to them because

I'm very close to the Grok team

because they're literally in Mountain

View, down the street from my house.

I'm in Los Altos and they're very nearby.

I'm going to see if they can hook it up

or do something where I can have a session

with one of their engineers and we can

kind of talk a little bit more.

Last time the QEMI model got released on

Groq, they hit me up.

They're , yo, this thing is ready.

Can we get an engineer on your stream?

I was , hell yeah, let's go.

And so I want to do more about that

because I feel ...

Let me just... man.

OK, if I get some time at the end of the

stream, my people who are here

to clean my room are going to kick me out

in a couple of minutes.

Let me just go through the rest of the

comments

and let me just show you how I get

started real quick,

because you could just literally throw

the documentation into v0 and have it do a

quick prototype for you.

And you're , you'll be you'll be halfway

there.

And that's how I've coded my app.

Yeah, that's what's up, man.

Yo, ChordsMaze, oh my God, with the super

chat, bro.

Yesterday, here's some updates.

OK, this OK, for those who don't know,

ChordsMaze yesterday dropped in the chat.

I call ChordsMaze Mr. Three Comma Club.

If you don't know about the three commas,

that means you're a billionaire.

This guy is a billionaire but from a

tokens perspective. This guy literally

yesterday dropped 5,703,144,183 total

tokens used in Claude Code, my friends.

That's 10K, 10 bands.

The Boy Cooks. And I also want to say as

a disclaimer, he's not proxying tokens.

He's

not running this stuff overnight. He's

not doing all the crazy stuff that

Anthropix says.

He's literally locked in 1000% and he's

just up, no touching grass, just

constantly

on this thing. I don't know what type of

app you're shipping, but I told you

yesterday,

Make sure you link that below because

you're on it and you have such a small

window.

You have less than 30 days right now.

The clock is counting down, ChordsMades.

This is why I started the AI Anonymous

group yesterday because I feel you're

misunderstood.

Anthropic is misunderstanding us.

They think we're using AI to do all this

stuff.

It's , no, we're just locked in, bro.

We're locked in this.

Let's go.

Need an app engineer to talk about in

five chips.

Hey, they're not going to talk.

PDF on the right.

CleanYourRoom, NiceFlexRay, love the

content bro. I'm just living my life. I

don't know if you if you live in the Bay

Area

Housing and everything is expensive, but

engineers are paid very well, right?

if you I worked at Apple for 12 and a

half years and I also invested a lot and

You know did pretty well at Apple

But at the same time now I'm kind of ,

you know, they let me go which was insane

You know, I was dealing with long COVID

and all that crazy stuff. They're just

bye. I'm , bro

I saved you half a billion dollars how

are you gonna do that to me they're just

let

us know when you get better we'll be

happy to have hire you back I'm all right

I see you

and then I see this AI wave and it's

literally that meme it's I'm checking out

with my chick and it's wait there's this

AI wave so I'm now looking this way I'm

let's go let's go so that's that's why

I'm here I started this channel and y'all

are

showing up and this is such an amazing

moment okay I love Grok Kimmy K2 is on

fire yeah Ruben

just became a member. Thank you so much.

Y'all are the real MVPs right now in July.

Congrats on your success. I'm going to

follow for that four hour stream after

this. Yeah,

check out that four hour stream. There's

so much sauce in there. Also check out

that members only

post to get into the discord, ask some

questions. I have some sections in there.

A lot of people are

starting to hop in right now and

participate. I thought my CC usage of 1700

was an API estimation

last month was impressive. No, bro. My

guy is locked in, locked in back there. I

think

I think KordzMaze just became a member.

Yo, KordzMaze, bro, I'm so excited to see

you in the Discord right now.

, you gotta pop in and drop that comment

and just say, yo, Mr.

And I'm gonna make a tag just for you in

my Discord saying, , you're Mr. 3, Claude.

So, , three commas is gonna be, , an

actual Discord tag for anyone who's ever

broke the billion token barrier, which

looks this dude.

But then he only, he didn't do one

billion. He did two billion. He did three.

Bro, bro, , my guy is locked in, has no

chill, bro. No chill. Five billions, bro.

Five billions.

How do I get into the Discord? There's a

post that I have. Just go to my YouTube

channel. You'll see members only.

That's where you'll see the members only

content as well on the YouTube channel.

There's a members only post.

And then there's just an email that you

send with your information. YouTube

doesn't provide emails or anything that.

They just show me your name and then I

think your channel.

And so if you just provide that in the

email, that I will be able to link you up

and stuff.

Orange hands go right.

So yeah, yeah.

AI is doing something to us for sure.

Day by day, raised vibes getting darker.

I used to go over 38 million tokens alone

last night, but only 100%.

Also I think there's somebody else

asking, I think if you remember, I might

just do

and so on.

So, I'm going to go ahead and show you

how I prototype because I got to roll out

in a

sec with v0 and stuff that.

Hacking Claude Code to get that Opus LLM

download for sure, man.

Yo, man, let me see if, let me just kind

of catch up with the other chats as well.

Just make sure I didn't miss anything.

And a lot of people kind of popping in as

well and making sure we're good because we

cover quite a bit today and I just wanted

to give you that quick AI context

visualizer

type of thing.

And I think the app is technically

deployed live right now.

I'm just going to go through some

changes, but you can kind of play around

with it right now if you're just cooking.

So, yeah, it may get taken down when I

redeploy the project, so just FYI.

But this is a great way to kind of

visualize and teach your friends.

I feel if you can teach this to other

engineering friends just the way I'm

teaching you,

it's going to help you go a long way in

terms of visualizing these things.

So the biggest takeaway I'd say is, with

your contacts, try to think about,

if you don't know and it's not really

clear what you're trying to solve,

I'm coining this thing called a discovery

prompt, right? literally just say, I don't

know where to go with this. Can you help

me? , I need to break this down so I don't

eat up the token context window. Just tell

it what your problem is, right? literally

it's the confession booth. I think AI is

our confession booth. So you sit down, you

tell it, and it's going to be , okay, let

me use my AI smart brain to kind of break

this apart into different pieces.

Once it comes back to you with that, then

start the new chat window, start the new

Claude Code window, start the new cursor

window.

Take one of those concerns and just pop

it in and just cook.

Just let it cook.

Say, in your first request, before you

pop it into these new windows, you should

also tell it the tooling you have

available.

So say you have access to subagents which

can take some of these requests or these

concerns,

researching, access to the web, searching

through codebases.

They can do these tasks in parallel. So

the more you describe what that scenario

looks , and then you can just transcribe

over, you know, what your problem is or

what feature you want to ship, the better

results you're going to get.

So Opus will then take that whole thing

and then write basically what it's going

to write to the subagents for you,

basically. And so that is kind of what

you'll take.

and you can just save that to a markdown

file so that you can either hand that off

to Claude Code.

You can see how this kind of keeps

compounding, but this is going to, in a

year from now,

this is not going to be that. In a year

from now, you'll just say whatever you

want. It's going to figure all that stuff

out for you.

But this is where we're at right now. We

have to do these little things piecemeal

because the tooling is currently just not

there.

And everyone's just learning as we go.

these studies that were done by Chroma was

recently just published.

and so on.

So, everyone had a feeling that this was

the vibe around AI, but they couldn't

really pinpoint

why.

And so, now that we have this good data,

the best practices, the engineering

efforts, everything,

every single company is going to add

tooling around this.

It only gets better for us.

So that's what's up.

Let's see.

I'm just kind of catching up some more

here.

So I'm struggling with Claude Code in the

Windows terminal.

Can't paste screenshots.

Does anyone know how to do it?

I'm on the Mac.

I can do it.

Are you using it in Cursor? Because when

I use Cursor, as soon as I drag it, I hold

shift.

But another person said you can copy and

then paste it in. And basically, as long

as you have

the directory, it'll go to that directory

and scan it. And it'll even ask saying,

hey, I'm

checking out your desktop. Is that okay?

So it will put it in single quotes. It'll

say this,

this, this, and then it should be okay. I

think when you drag it in the terminal,

let me even put those quotes in there for

you. And so play around with that. But as

long as you

You put whatever the directory path for

your screenshot in quotes, it should be

able to

know what that means and do something

about it.

Claude Code needs an interface to uncheck

context you don't need.

That'd be cool.

But at that point, I think the whole goal

of it is to be more hands off, which is

kind

of why they exposed subagents to us, so

that they can try to intelligently pull

that out

from whatever we're saying.

But I'd be interested.

You got to paste it into Claude Code by

dragging from the file from the directory

view.

Really hope Claude ends up in the IDE.

Yeah, that'd be so fire. How many years

you code?

Let's see. I mean, I've been in the

industry, shipping software, probably a

decade plus.

And as far as coding, a lot of that was

more early on. So I think initially I

started with,

what was it? C? Yeah, I did C code and

then C++ and then got into some Swifty

things, Python,

I did a lot of shell scripting and then I

got into Ruby on Rails, I got into PHP

because I needed tooling and dashboards

for internal tools.

And then less of that was... I started

getting more responsibility around

shipping software and then quality and

stuff.

So then there became less coding, more

understanding the business of apps and

quality and output and how that affects

users.

So that was kind of where I started to

spend more time in that arena.

and then later on was because I knew so

much of the stack all over the company

it was just fighting fires I would be the

person that would go to to figure out

what the heck happened so debugging was

my biggest skill it was okay I know what's

going on at this part of the stack you

know you know pulling all the people in

the room to figure

this stuff out as well so yeah so just a

wide range of experiences there's some

really

nice terminal windows that are better

than the standard Windows terminal but I'm

old school I

I just the standard ZSH shell. Maybe

throw some PowerShell or something on

there, but I'm good at that.

How do I check my token usage in Claude?

You just run bun x cc usage if you have

the bun or mpx.

So cc usage is the command and then bun x

cc usage is the one I do as well.

And that will do that. Or you can just do

mpx instead of bun x if you're a thing.

I gotta head on out and I will see you on

tomorrow's stream. Thank you very much.

Peace out.

Support.