YouTube Transcripts

Alex Finn
Latest Grok Bot is the best AI agent ever. Here's how to set it up
▶ Watch on YouTube
Grockbot is the absolute best AI agent
out there right now. If used correctly,
you will have an entire fleet of AI
agents working for you 247, 365. In this
video, I'm going to show you how to set
it up so you get the absolute most
productivity from this incredible AI
agent. If you stick with me until the
end of the video, you will have an army
of AI agents taking care of your entire
life for you all around the clock. Now,
let's lock in and get into it. We're
going to cover everything in this video.
We're going to cover setting it up,
setting up your first bots, first use
cases, how I'm using it, should you rip
out Hermes agent or Open Claw or not.
But first, I just want to give you a
quick update on why this is such an
incredible AI agent. Feel free to skip
around down below if you want to get to
specific parts of the video. I've been
using Grockbot literally non-stop for
the past week now, and I just can't get
enough of it. Every day I discover new
things, new ways to use it, better ways
to automate my life. And if this is your
first Grobot video you're watching, this
is what it looks like. By the way, it's
basically like this kind of iMessage
format where you have a list of bots
over on the lefth hand side. These are
completely custom bots. You can name
them, give them whatever picture you
want, give them a title, description,
and they all work together as a team.
They communicate with each other. They
message each other. They have their own
tools. Each one has their own computer.
So if you go to any of these agents, you
can actually see the computer they can
work on. These are fully working
computers. You can open up a browser,
use them any way you want. Each bot you
create has its own computer that lives
in the cloud. It doesn't need to live on
your computer, although it can control
your computer if you'd like. It's just a
really, really amazing simple
experience. And because of this simple
experience, I'm able to get so much more
productivity done. But I'll go over this
in a sec. I'll go over how to set up
each of these bots, what you should name
them, what should they do, what role
should they be, what use cases you
should do. We'll go over that in a
second. But here's the reason why it's
so good. It is by far the best out
ofthe-box experience for an AI agent
ever. You literally just open it up, you
download it, you install it, you have
your first bot, and you just get to
work. You don't have to choose models.
You don't have to choose anything. It
just out ofthe-box works. And I think
for a majority of people out there, this
is the ideal experience. This is how AI
agent should work. You shouldn't need to
configure a 100 different things. Give
it a 100 different permissions. Choose a
different model for every single thing
it does. Choose a 100 different tools.
It should just work. Grockbot is the
only AI agent I've used that just
straight up works. It has extremely
opinionated workflows, more so than
basically any other AI tool I've used in
the world. What does that mean,
extremely opinionated workflows?
Basically, it's gotten rid of most of
the customization that every other AI
tool gives you. For instance, with
Hermes, you can choose whatever model
you want. You can choose if it works in
the cloud or if it works locally. You
can choose if it uses sub aents. You can
choose the thinking level. You can
choose the context window. you right.
Everything you do has a hundred
different choices you can make on how it
does it. Not with Grockbot. With
Grockbot, it has an opinion on how you
should do things. It doesn't let you
edit all those things. It decides what's
the best way to do things and does it
that way. I think this is a supremely
better way to use AI tools. 95% of
people that use AI tools don't want to
think about which model and context
window and thinking level to do
absolutely everything they do. They just
want to give it a task and then it does
the task the best way. That's how
Grockbot works. It's opinionated about
everything. Every other AI tool, oh,
should I do this in the cloud? Should I
do it locally? Should I do it in the
terminal? Should I do it in the desktop
app? Grockbot just decides, no, we are
going to do everything inside this
virtual machine. everything these bots
do will be on their own computer. And
that kind of opinionated workflow makes
using Grockbots so much simpler. It's
cloud agent first. Again, as I said,
everything happens in the cloud
computer. I'm going to be honest, at
first I wasn't a huge fan going into
this. I just want all the work done on
my computer. Why do I want everything
done on the cloud? That feels like
there's like separation between me and
what the bots are doing. But after using
Grockbot for the last week and after all
the work getting done in the cloud, I
love it. I'm cloudpilled. There's this
natural security separation. There's
this natural scoping that happens when
everything happens in the cloud. I don't
have to worry about my bots having
access to accounts they shouldn't have
access to. If I want a bot to manage a
certain account, I just go onto their
virtual computer and log in for them. I
don't have to worry about them getting
access to the wrong things. So I am 100%
cloud build now. Zero config. As I said,
you never decide which model to use. You
never decide reasoning levels, none of
that. Hermes, you have 100 drop downs
for every single thing you do. With with
codecs, everything has 100 different
drop downs to configure how to do
things. None of that exists here. It's
multi- aent first. Again, I didn't know
I needed this, but after using it, now I
know I need it. By default, every agent
messages the other agent when it has
work to do. What I mean by that is if
I'm talking to my building agent, my
coding agent, and I'm talking about, oh,
can you code this thing that I tweeted
about the other day? It'll be, okay,
yeah, let me check in with your content
agent and see what you tweeted about,
right? It's by default thinks, okay, who
can I talk to to get the best context
from them? And so, you'll just be using
your bots and you'll see, oh, this bot
message this bot, this bot messaged this
bot, and they all talk to each other
just kind of by default, which is really
amazing. This is like a thing you have
to turn on with your other agents with
Grockbot just default really at the end
of the day. Why it's so great is it just
works with very little setup with very
little config. It just works and that's
what makes Grockbot so good. All right,
so let's get into setting it up. So
we're going to cover first things you
should do in Grockbot. We're going to
cover which bots you should set up.
We'll cover a couple different use
cases. I'll cover everything I'm doing
inside of my Grockbot just maybe as a
little inspiration for you. Then we'll
cover should you rip out Hermes? What
should you leave for Hermes? What should
you leave for Open Claw? How do those
two live together? Let's get straight
into setup here. So, when you start your
Grockbot experience, you will have one
bot. I recommend making this your CEO,
your chief of staff. You rightclick it,
you pin it to the top so it's always at
the top. Then I would go into edit
profile and I would give it the title
chief of staff, CEO, whatever you want.
Here's why you want like a CEO, chief of
staff bot because makes the entire
Grockbot experience far simpler. You're
going to end up as you go adding more
and more bots on your lefthand side
here, right? And there's nothing wrong
with that. You're allowed to have I've
seen people like literally like a
hundred bots. You're allowed to have a
lot of bots. There's nothing wrong with
that. But it might get annoying if you
have to scroll through hundreds of bots
and choose which one to talk to. That's
why you want a CEO bot because the
agentto agent communication between the
different agents is so good. You can
just rely on talking to just one bot and
allowing them to distribute the work to
delegate the work. it'll figure out
which bots have the right tools, the
right plugins, the right context, the
right roles, and then delegate that work
to the correct bots. So, yes, sometimes
I still do talk directly to some of my
other bots, but most of the time I just
go to Slate and I say, "Hey, here's what
I want to get done. Determine who should
be doing this work." And be like, "Oh,
that's coding work. I'll give it to
Build. Oh, that's this work. Oh, that's
Oh, you're sending an email. I'm going
to give that to Cindy." And it just
makes it much easier for you to work
with. Also, by the way, a majority of
people I see using Grockbot just like
make the names of each their bot like
content, email, coding, right? That's
like the name of all their bots. I
highly recommend against that. I've been
saying this for a long time. I think fun
should be like a big part of using AI.
Like, I think you should be having fun
when using AI. It'll make you use it
more. It'll help you be more productive.
It'll just give you more joy as you use
these tools. I find it a lot more fun
when you give your bots like real names.
Now, there's a couple exceptions here.
Last 30 days is a skill for my friend
Matt Van Horn. We'll talk about that in
a second. That just that set the name by
itself. Build. Okay, that's not exactly
a human name, but I'm going to change
that soon. But everything else is real
names. It's a lot more fun when you're
talking to your bots and it feels like
you're talking to a human being. So, as
you set up these bots, I'd recommend
giving them like real human names. It's
just going to be a lot more fun. You're
going to enjoy the process. So, when
you're setting up your CEO bot, give it
whatever name you want. The title is
what shows up next, a name. So, that's
it's going to allow you to give it like
human names and then you can look at the
title to see what it is. If you name
your bot like coder, it's just boring.
But if you give it a real name and then
you make the title like your CTO, then
you can know what it is very easily by
looking at your side here. Then the
description is basically the context
that gets sent with every prompt, right?
This is like the system prompt to the
bot. You can start off with like a basic
description of CEO for whoever you are.
I'm going to go in a second and show you
the easiest way to set up these
descriptions, the bots, and all that.
But just want to make sure your first
bot you have, you pin it to the top,
give it a nice name, one you won't mind
seeing over and over again. Don't name
it after your ex-boyfriend, your
ex-girlfriend, and then title it your
chief of staff. Once that's set up,
we're going to go into a world famous
process I invented and talk about all
the time, which is the brain dump to
reverse prompt process. So, this is the
first thing I do with every single AI
tool I get my hands on, every single AI
agent I set up, the brain dump to
reverse prompt. This makes it so you
don't have to think about setup. The AI
does all of the thinking and setting up
for you. So, step one, brain dump, you
tell the agent all about yourself. We're
going to go into all our passions,
goals, ambitions, what we want to get
rid of in our lives, what we want to do
more of in our lives. You just brain
dump it all in. Right after this, we're
going to do the reverse prompt, which
is, "Hey, what can you do for me? How
would you set yourself up? What bots
would you create? What routines would
you create?" Basically say, based on
everything you know about me, what's the
best way you can help me out? So, first,
put in your brain dump. I'm Alex Finn.
I'm the founder and CEO of Creator Buddy
and Henry Intelligent Machines. I'm an
entrepreneur and content creator. I have
a YouTube channel and XC account that
talks about AI. I also have the Vibe
Coding Academy, number one AI community
on the internet. By the way, join that
full boot camp on Grockbot coming up.
Join, be a part of that. You'll learn a
ton. Link down below. You hit enter on
that. Your new CEO is going to learn all
about you. Then you do the reverse
prompt I talked about. Based on what you
know about me, how would you set up
Grockbot? Which bot should we set up?
What should be their roles,
responsibilities, and routines? How can
we maximize our productivity as much as
possible? Then you hit enter on that and
it's going to recommend a bunch of bots
to set up. The glorious thing about
Grockbot is it can control itself. So
you can say that's great. Set all those
up for me and then your CEO bot will go
and actually create the bots, give the
names, give the descriptions, the titles
and all that and you'll be good to go.
That will give you a base of where to
start with. There are many other things
you want to do here. So I'm about to go
through the plugins, the tools you need.
We'll go through routines a little bit
and then I'll show you everything I'm
doing so you can get inspired by some of
these use cases and set them up for
yourself. This plug-in section in the
bottom left here, this is basically like
your MCPs, your tools, all that. They
just put them all in one place, which is
really, really smart. Again,
simplification, just making things work
out of the box. There's some basic ones
you want to set up, kind of
self-explanatory, Gmail and Google
Calendar. I'd also highly recommend
setting up X so you can pull posts very
very easily from X. It is great with
social media content. But here's the one
I've actually liked the most and this is
one I highly recommend you set up
yourself and that is Agent Mail. We are
not sponsored by Agent Mail. I just
discovered them like 3 days ago, but it
has made my Grothbot so much more
powerful. And I'd highly recommend you
do the same thing here. Set up Agent
Mail. It's completely free. You get like
three free inboxes with the free tier.
But here's how I'd use it. Basically,
what agent mail does is give your AI
agents their own email address, like
their own inbox, their own email
address, everything really, really
easily. It's way better than setting up
Gmail because when you set up new Gmail
accounts, there's so much security and
backup and and all this like complexity.
Agent Mail, you literally just sign up
in like 10 seconds, you have an inbox
for all your agents. But we're going to
use this inbox that you create with
agent mail for all your agents so that
when you invite them to your different
accounts, they can just have their own
account. You don't need to share your
own accounts with them. So for instance,
Dusty is my community manager for the
Vibe Coding Academy. Dusty answers
people's posts, helps them out with
technical questions, sends them DMs when
they have questions, gives them
recommendations for things to build,
does a whole lot of amazing things, but
I don't have Dusty logged into my admin
account in the Vibe Coding Academy in
school. I invited Dusty to his own
account. So, as you can see here, this
is Dusty's computer. Inside of the
computer, I invited Dusty his agent mail
email to my community and I made him a
moderator. Now he can do anything in my
community he wants through his own
computer. So I don't give the bots my
accounts. I don't give the bots access
to anything. Everything is done through
their own email address. You just invite
them to your team on whatever apps
you're using and then you give them
whatever permissions they need. This is
a way more secure way of interacting
with your bots, giving them accounts,
things like that. It keeps things
separated and using Agent Mail made it
like a hundred times easier. Again,
we're not sponsored. I've never talked
to anyone from Agent Mail in my entire
life. I just discovered them over the
weekend. Just made working with my bot
so much easier because they have now
their own accounts. I can just say,
"Hey, I invited you to my community
through your agent mail account. Go
accept it and now you're a moderator and
go do things in there." And it just
works and it does it. So, Agent Mail,
very important plugin everyone should be
installing. If you do any sort of coding
work, which I assume you do if you're
watching this channel, Verscell plugin
is a mustave. Makes it super easy for
your bots to push code to Verscell,
update your projects, things like that.
So, also add Versel as a plugin right
here, too. One really cool thing about
plugins as well and the way Grockbot
handles things is plugins are basically
everything, right? As I said earlier,
they're MCPs. They're also skills. So, I
install skills and they go into your
plugins as well. So, this is a really
cool way to manage all your skills,
MCPs, plugins, everything in one place.
I really recommend the last 30day skill.
This is made by my friend Matt Van Horn.
It is like the best researching skill
out there. I'll leave a link to it down
below. Basically, the way it works is
you can give a topic or a subject and it
uses like the API for every single
social media site on planet Earth. like
reverse engineered all the APIs and
gives you a rundown of what people are
saying, how they're using it, the latest
news, all of that. And it's like the
best deep research when it comes to
trends I've ever done. So, last 30 days
skill, get that installed as well. It
just makes like the research from your
agent so much better. I'm having
research cursor origin now, which is
Cursor's GitHub competitor. If you want
a video on that, let me know down below
if you're interested. Also, let me know
down below what parts of Grockbot you're
really interested in. I'll do like deep
dive videos. Do you want like a super
duper use case video? I'll do that next
if you want. Let me know down below.
Also, leave a like, subscribe, and turn
on notifications. If you learned
anything at all so far, all I do is make
amazing videos about AI. Join the Finn
fam. Hit subscribe down below. Tons of
amazing videos coming out very, very
soon. So, we talked about initial setup,
reverse prompting, brain dumping. We
talked about the skills and plugins you
need installed. What I'm going to do now
is I'll go through my use cases. I'll go
through my bots right here. You can go
ahead and steal any of my use cases I'm
doing or maybe just inspires you on
different things you can be doing with
Grockbot as well. So, the first bot I
use a ton is Build. Build is like my
network administrator/coder.
Basically does all my technical work for
me. If you're anything like me, you have
a bunch of different devices. Maybe you
have your iPhone, your iPad, you have a
computer, maybe a Mac Mini, maybe you
have other computers on top of that.
Build is my network administrator. I
gave it access to my tail scale network.
All you need to do is say, "Hey, my
network's on tail scale. Please access
it." If you have multiple computers,
highly recommend setting up tails scale
as well. Basically lets your AI agents
go across all your devices and do
whatever needs across them. But for
instance, I had build my technical bot
go last night, install Quen 3827B
on my 5090 computer. Then it built a
game for me right here on my Mac Studio
using that model that was running
locally on my 5090. So it can go across
any of my devices, manage any of them,
load models up, build apps, do any sort
of coding. It's hooked into Verscell so
it can see all my code bases, all my
projects I'm working on. I have like a
personal operating system I maintain. I
need a little fix in it. Today I went to
Bill. I said, "Hey, in my personal
operating system, can you edit this?"
It's basically like my CTO. Anytime I
need to do any technical work at all, I
go to build and it works really well.
The reason why like this is better than
having like one mono agent that just
handles everything is build just has
context just around my technical work,
just around my computers, how things are
set up in my projects. That's it. Like
the description it has just describes my
projects and computers. So when we give
it prompts and we tell it to do things,
it only has to pull from a very small
description, from a very small context.
When you have one agent doing
everything, doing your programming, your
content, this and that, its skills, its
description, its system prompt is
massive. And the bigger the system
prompt gets for an agent, the slower it
becomes, the more expensive it becomes,
and the stupider it becomes. This is the
beauty of the architecture of Grockbot
is all your skills, all your description
system prompts are split up nicely
between bots. So when you message the
right bot, it's quick, it's cheap, it's
easy because it just knows what it needs
to know. Just the system prompt it needs
to know, just the role it needs to know,
the tools it needs to know. That's it.
Nothing else. That's why this
architecture of Grockbot is so great.
Then we have Barry. Barry is my content
engine. Barry keeps me up to date on
trends, breaking AI news. It's always
watching X for me, looking for breaking
news and things like that. Barry has
access to X API. Barry does a few things
for me. One is it helps me write all my
newsletters. Newsletters take a long
time to write. Barry helps me write
them. It takes my posts, my YouTube
videos, repurpos them into newsletters
for me. It also keeps an eye on any AI
major products. Right? So, this is
another routine here. We didn't really
cover this too much. Routines are
basically just your cron jobs. You can
reverse prompt your routines as well. Go
to an age say, "Hey, what are the best
routines we can set up?" It'll set up
the routines as well. Every 30 minutes
from 7:00 a.m. to 11:30 p.m. Barry goes
and checks the SpaceX Twitter account.
It checks Anthropic. It checks Open AAI
and it sees if they made any new major
releases recently and then lets me know
about it, which is amazing. As you can
see, SpaceX just shipped a connector
inside Grock live now. Boom. I get that
breaking news the moment it happens. So,
even if you're not into content, I still
recommend having a berry of your own
that just keeps an eye on X and lets you
know the moment new things happen. Then
we have Dusty, who we talked about a
little earlier. Dusty is my Vibe Coding
Academy moderator. Anytime someone posts
a technical question that'd be
appropriate for an AI to answer, Dusty
answers it. People aren't as engaged
after a few days after they join, Dusty
will message them and say, "Hey, here's
a few ideas based on my research about
you of things you can build." And Dusty
just keeps people engaged, keeps things
moving, post new AI news in the
community every day. It just takes so
much work off my plate. And because I
invited Dusty through that agent mail
account, it's a moderator. It can do
whatever it wants. It doesn't have admin
access cuz it's not using my account,
but Dusty has his own account. If you
need any sort of community management,
Grockbot is like the perfect candidate
for that. Cindy is my revenue ops.
Cindy's basic job is monitoring my email
and finding sponsorship opportunities
and replying to people. My number one
pet peeve in life. I don't know what it
is. I don't know if there's something
wrong with me. I absolutely hate email.
I hate email. When I get a popup on my
phone, it's an email for me. I get like
anxious. Like I hate it. I hate getting
emails. I hate reading emails. I hate
replying to emails. The downside of this
is I don't get many sponsorships because
of that. I get hundreds of emails a day
from companies trying to sponsor my
videos. I don't reply to any of them
just because I hate emails so much.
Cindy does that for me. So Cindy's
hooked into my business account. Cindy
goes through the hundreds of emails I
get a day, researches who's emailing me,
if they're legit, if they're just a
scam. By the way, if you get into
YouTube making, you're going to get
40,000 scam emails a day. Be careful out
there. But Cindy researches who scammer,
who's real, and at the end of the day
gives me a spreadsheet of the legit
opportunities that might be real
sponsorship opportunities. So Cindy is
amazing. If you're anything like me, you
hate email. Set up a bot to monitor
whatever email accounts important to you
when it comes to business and alert you
when important business emails come in.
Get yourself a Cindy. And then I have
Reed. Reed is my last bot right now.
Reed is my experimentter. I have Reed go
and experiment on the internet all day.
This is going to sound a little strange,
but I have Reed going running
experiments, finding what people are
saying online, build products based on
what they're saying, post those products
online to see if people click it, use
it, see where there's demand. I
basically have Reed going out, finding
demands, running experiments, seeing
what people are into on the internet.
Reed's kind of like my go-to guy for
just finding business opportunities
online. Slate manages Reed, so I don't
go to Reed and say, "Hey, do this, do
that." I just tell Slate, "Hey, have
Reed try different things out and see if
there's different business opportunities
online." Then Slate just watches Reed do
all those things all day. So, he's kind
of my experimentter. These are my main
use cases right now. Based on those bots
I showed you, I can manage like 90% of
my business. I'm coming up with new bots
all the time. So, if you stick with the
channel, you subscribe down below. I'm
This is the first of many Grockbot
videos I'm going to do. So, you'll see
how my setup evolves over time. I'm
going to be adding way more bots to
this. Hermes versus Grockbot. Which
should you use? Which should you rip
out? These are two different agents.
It's not one or the other. It isn't
Hermes this or Grockbot that. Hermes is
a completely customizable agent running
on any model that can do anything you
want. If I load up Quen 38 on my 5090
computer, it's now powering my Hermes,
right? Because Hermes is fully
customizable. I can't change models in
Grockbot. You could only use Grock for
Grockbot. So, there's still going to be
things you need to do that require
customization, that require cheaper
models, require your agent going and
doing things on your computer,
tinkering, changing things around.
That's where Hermes comes in. That's
what I would use Hermes for. But for
basic day-to-day knowledge work,
Grockbot's the goat right now. Grockbot
is the best way to do it. So, I do not
recommend ripping out Hermes. Keep
Hermes. Use it for cheaper models. Use
it for changing things on your computer,
doing admin things on your computer,
things that require deep customization,
and then use Grockbot for your general
knowledge work. I'm absolutely in love
with Grockbot. I'm sure you are too if
you started using it. I hope this was
helpful. Let me know down below what you
want my next video to be about. Any
aspects, Grockbot, use cases, things
like that. Let me know. Hope this was
helpful.
Prompt Engineering
Latest Don't Pick One Coding Agent—Combine Them
▶ Watch on YouTube
Okay, so what is the best coding agent
available today? Is it Codex, Cursor, or
Claude Code? I think that is the wrong
question to ask and a lot of people
focus on that.
The right question is how do you combine
them together to build powerful
software?
And in this video, I want to really show
you my own workflow that I personally
use where I combine the strength of
different coding agents within a modern
IDE to build some really amazing
features. For example, this is an
educational tool that I'm currently
building for my son, and it really lets
you explore different organs, different
systems in the body. So, my goal for
this video is to show you a workflow
which is really easy to follow, and but
it's going to help you build some really
great software
that you can actually ship to your users
all within a modern IDE. Now, in the
rest of the video, I'll show you my
workflow, how to use the strengths of
different coding agents in your IDE.
For this, we're going to be using
PyCharm from JetBrains, who are also the
kind sponsors of this video. However,
the principles are going to apply to any
framework or IDE that you are working
with.
But, JetBrains does offer some really
neat extra features. First, you can
bring any coding agent of your choice
either through your subscription or
through an API.
So, you're not locked into a specific
vendor.
Anything that supports agent
communication protocol can be integrated
in here.
Second, and I think this is a big one.
This enables you to bring the IDE
context into your coding agent, which is
not available in a CLI-based coding
agent. Okay, so what exactly we're going
to be building? We're going to look at a
very common scenario where you often
provide specs or requirements to your
coding agent and it is going to build
something based on those requirements.
But often you're not sure whether the
code implementation actually follows
your full specs or not. That's the
problem we're going to be tackling in
this video.
Okay, so the app I showed you in the
beginning was implemented based on these
requirements.
Now, it's a really big build, so I'm not
sure whether the coding agent that I
used actually followed the set of
requirements fully or not. The output
looks correct, but I think there are a
lot of missing pieces.
So, we're going to use different coding
agents with different capabilities in
order to ensure that the agents actually
follow
our set of requirements.
Now, as I said,
this new AI
in the IDE feature lets you connect
multiple different
agents in the same IDE interface.
So, right now, if you look at, we have
Quad Code, Client, CodeX, Cursive, even
I have connected Kimicy and I. Or you
can just bring in a coding agent of your
choice as long as it supports the agent
communication protocol.
So, here is a list of all the different
coding agents that you can connect
through ACP.
This is pretty comprehensive list.
This includes both proprietary and open
source coding agents.
Now, when you start, you will first need
to configure your agents as using either
an API or through their subscription.
Now, right now I'm selecting CodeX. Now,
if we send in our first message, it's
asking me to install and continue.
So, we are going to install CodeX on
this machine.
Okay, you basically get access to all of
the different features that a specific
agent actually provides you here. So,
for example, we can select which coding
model that we want to use. Let's say we
want to go with uh GPT-5.6-Soul,
which is great for code review. And then
I'm going to put it on high.
Now, in terms of the coding agent
itself, it gives you all of the
different capabilities. So, for example,
right now you can select the model, the
effort level. Since we're doing a code
review, I would select uh the reasoning
effort to be high. And then you can give
access to the agent, whether you want
full access
or read only, or you want the agent to
ask you for every action. Right?
Similarly, if you go to something like
cloud code, again uh you have different
modes that you can set here. You can
also use any skills that you have
already installed with the agent, which
is pretty neat. And then you'll be able
to
set up the model and reasoning level.
Now, I am not using Fable for this
because this is a bio-related uh
application, and usually that flags
the classifier.
But, what we're going to do is we're
going to use GPT-5.6-Soul
in order to find any discrepancies.
And then we're going to use the Kimi
code to actually
implement those.
Okay, so in here I'm asking it to uh
look at the uh build.md, which has the
specs and uh the rest of the code, to
make sure there's a feature parity.
Also, you can enable or disable
the IDE context. Right now, it's
enabled, so it will be able to see
whatever I am doing in the IDE.
All right, so we send this in.
So, right now codex is planning on doing
the full audit, reading the specific
files that it needs to. We're going to
come back when it's finished.
Okay, so while the GPT-5.6 is doing the
analysis, a couple of things that I know
are missing is that I asked the agent in
the initial spec to
include glowing particles in these
animations, but seems like it forgot to
implement those. But, it said that the
implementation is complete. Okay, so the
first phase of analysis is complete and
it's actually a much stronger
implementation than I thought. Uh but,
it did find some missing features that
are not implemented. And that's why you
usually want to use a completely
different agent in order to verify the
work.
And I even recommend a different class
of agent. So, if you do
uh implementation with, let's say, topic
models, it's good to use OpenAI models
for verification.
Now, overall, the implementation is
pretty strong. However, it did find
some missing pieces. So, for example,
that uh
particle flow piece is missing and it
was able to identify it, which is pretty
neat, right? Uh based on this analysis,
I asked it to come up with an
implementation plan only for P0 and P1
for the sake of this video to keep it
simple.
So, here's the implementation plan that
it came up with. It also gives you a
diff of every change that it's making.
So, this is really helpful, uh
especially in a full IDE. You can easily
look at what the agent is doing.
Then, you can just take uh this and use
another coding agents like Kim i CLI uh
to actually implement this. Kim i CLI
using KT is a lot less expensive, but
it's a pretty capable model. Now, you
have two different options. Either you
can just switch here. And this way it's
going to have full context of all of the
conversation or you can just go to a new
chat session. Uh so for example, I
created a new chat session and in here I
simply asked it to look at the feature
implementation plan and test and
validate and document the implementation
that is going to be doing.
Now, uh for the model we selected Gemini
say light. Uh I have already connected
this through my Gemini subscription. I
chose Gemini 1.5 thinking for this
and right now is the default
permissions.
So at the moment the agent is working
through the implementation.
Okay, the first pass is complete.
However, the verifier found that the
implementer still had
gaps in the implementation and that's
why it's very important to use a
secondary model to
check the implementation correctness
based on the specifications that you're
providing.
This is the biggest lever that you have
when it comes to agentic coding.
You don't want to just rely on the word
of a model
in regards to its work. Either you want
to verify it or if it's a substantial
work, you want a secondary agent to
verify the work
and then provide feedback.
Now in this case I took that second part
of the feedback and basically provided
it to the implementer and the
implementer has implemented fix for
those issues that were highlighted.
I'm using the same prompt with Codex
solved on high setting again just to
verify the implementation.
And then it basically independently runs
everything and verifies them again.
So this time with the second pass a lot
closer to what we want. There are some
minor things that we still need to fix,
but this is in a very good shape now. So
here I asked it to run the app for me to
test it out. Okay, so the major issue
was with the organs and now it has
implemented that pretty neat article
effect uh list uh functionality is
there. So, this is really good, right?
Now, personally, I like this iterative
refinement and improvement. Still want
to personally test out at the
implementation. I have found that this
works really well for me uh because even
though I am spending some time up front,
it actually saves me a lot of time down
the line.
And for this, especially having an IDE
where you can actually uh check what
exactly is being changed and worked on
is extremely critical.
So, I usually follow this
planner,
implementer, and verify loop.
But, do let me know if there is any
pattern that you have found
that works really well for you. I would
love to learn from the community.
So, the question is not which coding
agent is the best. I think it's how you
are able to build workflows for your
specific needs
that combines the strengths of multiple
different coding agents for you
in a modern IDE.
Do let me know
your thoughts and I hope you found this
video useful. Thanks for watching and as
always, see you in the next one.
Matthew Berman
Latest 6 Open-Source AI Projects Trending NOW
▶ Watch on YouTube
Here are six incredible open- source
projects that you can install right now.
Let's get into the first one, Unsloth.
It started as a way to make LLM
fine-tuning locally much easier for the
average person like me. Now, it has
evolved into a powerhouse of a tool for
everything from fine-tuning to running
inference, training models, and
everything in between. So, let me show
you a little bit about it. Again, it is
free. It is open source. It comes in
Windows, Mac, Linux, everything you
need. You can run and train all the
latest open- source models including
Kimmy K3, Miniax, Quen 38, Muse Glimmer,
which just launched by Meta, Deepseek
V4, Gemma by Google, everything. It is
also now a fullyfledged agent UI. So, it
looks just like Chat GBT. Of course, you
could type anything in here, but of
course, it is running a local model. It
has web search. You can plug in tools.
You can have MCP. It has memory.
Everything that you need all in one
place and all local. All of the features
that you've come to really depend on and
love from products like Codeex and Cloud
Code and Cursor are all built into
Unsloth, but again, everything just runs
locally. It supports all different types
of hardware. You can actually remote
access into it, which is really nice. So
you have it running on one computer, you
can open it up from another and control
that computer from anywhere in the
world, which is very similar to what
Codeex and Cloud Code and Cursor also
do. You can run and train image models,
video models, text models, basically
everything really just out of the box.
It works extremely well. And one thing
that I absolutely love is if you're
trying to dip your toe into training and
fine-tuning, it can be extremely
intimidating. There are so many
different settings, so many different
ways that things can go wrong. But
Unsloth makes it really easy and you
don't need to code at all. It's
basically just a point-and-click
interface that allows you to do
training, to do fine-tuning, and it
walks you through it and just makes it
all really easy. And by the way, I'm
going to drop all of the links to all of
these projects down below. Next is a
project called diagram design. Diagram
design gives your agent the ability to
create diagrams that are actually
good-looking and don't have arrows that
are all broken and overlapping and
really nice, really simple and easy to
use. And they have a ton of different
diagram options for you. Here's a
flowchart. Here's an architecture chart,
a state machine, a timeline. Basically,
everything you could possibly need.
Quadrant, all easily available. You just
give it to your agent. You can install
it as a plugin in Cloud Code, plugin in
Codeex, in Pi, and you can install it in
Hermes Agent, which is what I'm going to
show you how to do right now. And today
I'm hosting Hermes on Hostinger, who's
also sponsoring this video. So, check
this out. I already have Hermes set up
for myself in Hostinger, which means it
is hosted in the cloud. I don't need to
actually have my computer on and running
all day long. I simply type install
this, give it the GitHub URL, and then
hit enter. So, it installed it as a
skill. It only took 1 minute. And now I
can create a diagram. Let me show you
how to do that. Create a diagram for
Hermes agent to show the architecture.
And of course, it knew to use the
diagram design skill. And while that's
going, let me just explain why hosting
and hosting in the cloud is so
important. You don't have to keep your
computer on. And in fact, that was fast.
You can run background processes 24
hours a day. It's actually a one-click
install to get Hermes running. So, if
Hermes was a little bit intimidating to
you, this is a great solution. One click
and you're in. The diagram design skill
worked flawlessly. This looks beautiful.
And don't forget to use code Matthew B
in the checkout to get 10% off. If you
sign up for 24 months, you get the
biggest discount. So, go check it out.
Link down below. Next is Obsidian
Skills. And if you're not familiar with
Obsidian, you should definitely check it
out. It's basically a simple note-taking
app, but everything is done in markdown.
Obsidian kind of reminds me of this
image. What you see when you first get
into it and start using it is this
little bit on top. But with those simple
primitives, you can get so much
functionality out of it. I sync it
across all of my devices. It is really
cool. And now you can give your agent
access to your Obsidian and let it use
it in all these different kind of cool
ways. Obsidian skills uses the agent
skills specification, so it can be used
with any agent, including Hermes, which
I just showed you, Claude Code, Codeex,
Cursor, Grockbot, all of the above. I've
seen people do incredible things with
Obsidian, like using it as a wiki, using
it as a knowledge base for their agent,
basically throwing everything into it
and using it as kind of the brain of the
agent. And obviously agents work really
well with markdown and Obsidian is all
about markdown. So it works really
cleanly with your agent and your agent
can just use it. It can store all of
these things locally if you want. It is
very much local first. If you want to
sync across all your devices like I do,
you can also do that quite easily. And
now again with Obsidian Skills, your
agent has access to knowing how to do
all of it. All right. Next from Jack
Dorsey's company, the creator of
Twitter. He also created Square and now
it's called Block. They have an
open-source Slack alternative. This is
called Buzz and it's basically like
Slack but made to be agent native, made
to be AI first. And the nice thing is it
is completely open- source coming in at
27,000 stars after like a week of being
published. So, this is what it looks
like. If you're familiar with Slack,
you'll probably be very familiar with
this. But the thing is here is the key.
Your agents are first class citizens in
Buzz. So you have your team and you have
your agents working in the same channels
working on the same work and they are
both users first class users of Buzz.
And I think that's kind of the
distinctive feature of Buzz is it's very
much agents are working with you with
your human team. It is self-hostable
which is obviously quite nice. They are
very much privacy and security first,
which obviously if you're wanting to
host all of this stuff locally, you're
probably thinking quite deeply about.
And it has all the features you're used
to, emojis, replies, threads, channels,
all of the things you're used to. And
then also, you get the agents as first
class citizens. It also deeply embeds
workflows and automations directly into
it. It is not really just a chat app. It
very much lets you build out automations
directly into it. This is meant to be
the central hub for your work. They
track everything being done in Buzz,
which is really interesting. It's a
noster, which I've never heard of.
Nostaster relay. Every message,
reaction, workflow, step, review,
approval, and get event is a signed
event in one log. Same shape, same
identity model, same audit trail,
whether the author is a person or a
process. And when they say process, they
mean agent, workflow, etc. It's all the
same. I really like this structure. You
can bring any model you want, open
source, closed source, hosted, local,
whatever you want. And because it is
backed by a major company in block and
by Jack Dorsey, you know, they're going
to continue to iterate on it and make it
better over time. And it is a newish
product, so go test it out, see if you
like it. Next is Ego Light, an open
source project that is self-described as
the fastest browser to give to your
agents. so your agents can control your
browser. This is Egoite. It has 10,000
stars, so relatively under the radar
right now. It describes itself as the
fastest browser for AI agents to run
browser automation. Built for sharing
your loggedin browser state with your AI
agents like Codeex or Cloud Code without
disturbing you. Zero cost, zero
configuration. Wonderful. So this is
what it looks like. This is codeex. You
type / browser. So it gets invoked kind
of like a skill. Let's actually see it
run. So this is the ego browser right
here
and we can actually see it running. So
it has that nice blue glimmer. You can
actually see the arrow the cursor and
there it goes. Super fast and especially
if you're going to be using it with the
new cerebrous powered soul, it's going
to be even faster. So very cool, free,
open source. Go check it out. And the
last project is Modly. Coming in at just
under 6,000 stars. It is image to 3D
design. Local open- source AI powered
image to 3D mesh generation. So whether
you're doing uh asset generation or
you're into 3D printing like I am, you
can basically just take any image, load
it in, and it will create a 3D mesh that
you can then use for your 3D printing or
your game or whatever other use you're
using that asset for. It runs entirely
locally, which is very, very nice. And
it's not really compute inensive either.
So, you can use this with your desktop
GPU most likely. It has support for
Windows, Linux, and Mac and many
different GPUs. So, if you're into 3D
modeling, go try this out. Seems very
cool. I made another video that covers a
ton more. Go check it out right
Parker Prompts
Latest How to Make AI Videos With Claude Fable 5 (Higgsfield Supercomputer)
▶ Watch on YouTube
Claude Fable 5 just completely changed
how AI videos get made, and every video
you're watching right now was made by
it. Four ads and a week of vlogs, and I
didn't edit a single frame. I just
approved steps in one chat. I've been
testing this setup for the past week,
and nothing else I've tried even comes
close. And so, in this video, I'll show
you exactly how to make AI videos with
Fable 5 from one brief to a finished
batch, so you can make your first one
today. The idea is you stop making the
videos yourself and start directing an
agent that makes them for you. With a
normal video model, you write a prompt
and take what comes back. But, here you
hand over the whole job and approve the
plan, the cast, and each render as it
builds. Higgsfield supercomputer runs
all of Higgsfield from a single chat, so
you describe what you want once, and it
plans the steps out, matching each one
to the right model and rendering the
lot. And the reason it can take on a
brief this big is that it's running on
Claude Fable 5. The way it works is that
Fable 5 is the planner, and each step
gets handed to whatever specialist fits.
An image model for the frames, a video
one for the renders, and a separate
model for the voices. So, Fable 5's real
job is lining them all up and passing
the right thing to each one. Normally,
four ads like these would mean jumping
between a few different tools and an
editor. A week of vlogs on top of that
is days of work. Here, all of it happens
in one chat. The only piece I make
myself is the product, so that's where
we start. Every ad is going to reuse
that one product shot, so it's worth
getting right before anything else.
Everything today runs on Higgsfield, and
the link is in the description if you
want to follow along step-by-step. So,
before I even open supercomputer, I'll
go to the image tab, pick GPT image 2,
set it to a one-to-one square, and
describe the product I want. For this
video, the product is a simple mug, a
matte cream ceramic one with a little
line art sun on the front. And I've got
a plain product shot of the mug on a
white background, which every ad is
built around from here on. You make it
once and reuse it, because if you let
the agent regenerate the mug for every
ad, the glaze and the shape come out a
bit different each time, and your four
ads stop looking like the same product.
The lighting on this shot matters more
than it looks. Keep it flat and with no
harsh shadows or reflections, because
the more the lighting is locked into the
product shot, the harder it is for the
agent to relight it convincingly in a
sunny kitchen or a dim reading corner
later. I use GPT image two for it
because it holds fine product detail
well. The shape and the little sun come
out exactly right, which counts for a
lot when this one image is about to
appear in every ad. I keep it one-to-one
on a plain background, too, since a
simple product plate is the easiest
thing for the model to drop into any
scene later. Now, I'll open
Supercomputer, and there are only two
settings to change here. I'll pick
approve before generation, so after
every prompt it lays out its plan and
pauses until I hit run instead of
spinning on its own. I'll set it to
Claude Fable 5, since that's the part
that reads the brief and plans every
step. And then I hand it a single brief
that has to come back as four ads that
all feel different. I'll give it that
brief in one message. The mug goes in
first, so it's locked as the product
from line one. Then I tell it what I
want. UGC style ads, four of them, each
a different angle, morning coffee, cozy
reading, office desk, and a gift idea.
All casual and honest, like real
creators shot them. I want them
vertical, 9:16, so they're ready for
reels and shorts. And I ask it to walk
me through each step, so I can sign off
as it goes. This is why I made the image
first, instead of just describing the
mug in the brief. If you only describe
it, the agent pictures it a bit
differently every time. An attached
photo gives it one fixed thing to
[music] copy, so the mug stays identical
across all four ads. From that one
message, it comes back with a full plan.
Four ads and four different avatars,
each one worked out down to the hair and
the styling. Everyone gets its own
setting and its own script. The casting
is properly varied. The coffee ad gets a
late 20s woman with ash blonde bangs in
a bright kitchen, and the reading one a
softer brunette on a couch with a throw
and books. For the office ad, it picks a
30-something guy with light stubble at a
lived-in desk. And the gift one a
bearded guy boxing it up under string
lights. The scripts are just as
different. [music] One leans on the mug
making the morning feel calmer. Another
plays it as a break from a desk full of
monitors. And the gift version sells it
as the present that always works. So, no
two of them sound the same. It all comes
down to the brief. I tell it to start,
and it builds the first avatar, a late
20s woman in a bright kitchen. Once Once
it she's good, it puts that frame
together, her holding the mug with the
sun clearly on it, and starts rendering
that first video while it moves on to
the next creator. That start frame step
is what keeps each ad consistent. It
locks the creator's face and the mug
into one still image first. So, when the
video renders, it has a fixed target to
animate instead of reinventing her frame
by frame. [music] I make it check in at
each step, mostly to save money. Catch a
bad avatar right here and it's a
one-line fix. Miss it and let it run and
you've paid to render four ads you'll
just throw out. The second creator is a
woman with soft brown waves and a
reading mug. And once I approve her, it
goes straight into her start frame and
renders the video off that. Then I tell
it to build the last two at once instead
of stopping to check with me on each and
it works the same way through both. The
avatar first, then the start frame, then
the video straight off that frame. So,
now there are four creators who all look
different, two women and two men with a
finished ad each. Then I change
direction. I tell it I want them looking
at the camera and talking. The video's
at 10 seconds and the scripts tighten to
fit. [music] And it regenerates all four
as 10-second on-camera videos with
proper lip sync, rewriting each script
shorter to fit, and it never once starts
over. I push it to 10 seconds and
on-camera on purpose. 10 seconds is long
enough to make one point and short
enough that nobody scrolls off before
it's done. And having them talk to
camera is what makes it read as a real
person.
>> This is the mug I reach for first. It's
just a simple ceramic thing with a sun
on it,
but pouring my coffee into it makes the
whole morning feel calmer.
This mug is my perfect reading
companion.
Tea in one hand, book in the other.
The little sun on the side just fits the
vibe.
>> My desk used to be all screens. This mug
didn't change much,
but having it next to my keyboard makes
the work day feel less robotic.
Whenever I don't know what to get
someone, I get this mug.
Simple, beautiful.
Everyone deserves a cup that makes them
smile.
>> Watching the four back, it's clearly the
same mug in everyone, just re-lit for
each scene. Ads are one thing, but the
same agent will build you a whole
content series. Now, let's do something
completely different, a week of casual
vlogs. This time I want a brand new
creator instead of reusing any of the ad
avatars. So, I have it build her
character sheet first, full body and
face, and then each day's video comes
straight off that sheet, so there's no
start frame to make this time. It's
seven vlogs, one for each day of her
week, 10 seconds each and all on camera.
Then I'll paste the brief in. It plans
her out, too, a warm mid-20s woman with
dark hair and light freckles. Then it
maps out 7 days for her, a morning
routine, a gym session, errands around
town, cooking dinner, then a night out
with friends, a day outdoors, and a slow
reset day at home. I'll approve, and it
builds the character sheet, then
generates all seven vlogs off that one
sheet, so it's the same face all week.
Get the character sheet right, and
you've basically got a creator you own.
Once she exists, you can put her in any
video from here on. So, it's worth
nudging her face and build until they're
right before you generate a single day.
It's also why the vlogs come together
quicker than the ads did. The ads needed
a fresh start frame for every video,
since each one was a different creator
holding a product. Here it's one person
all week. So, the character sheet on its
own is enough for the model to hold her
face the whole way through. The sheet
keeps her face locked, but the clothes
and the settings can still change
day-to-day. If you want a tighter look,
put that in the brief, too. This vlog
style is a different job to the ads.
It's how you'd run a personal brand or a
faceless channel. I do want to fix the
voice, though, so I ask it to put a
single consistent voice across all seven
days. It offers me a few presets that
suit a young creator. I pick one called
Quinn, and it applies that same voice to
every vlog. Preset voices like that are
quick, and they fit, but they can sound
a little generic. So, for a real
channel, you'd probably clone your own
voice instead. This is just the fast way
to keep the whole week sounding like one
person.
>> Morning coffee.
Let's see if I can be a functioning
human today. Same walk every day, and I
still love it.
Three stops, and I've already lost the
list. Post office, groceries, and one
thing I definitely forgot.
Wish me luck.
Only got like 40 minutes today, so we're
keeping it short.
Legs,
then I'm out.
Honestly, the hardest part is just
showing up.
Attempting to cook something that isn't
pasta. Wish me luck. Said I'd be ready
at 8:00. It's 8:15.
Meeting the girls in 20 minutes.
This is 5:00 now.
Needed some sun.
Found a spot and honestly, might just
stay here forever. Doing absolutely
nothing today and I've never felt
better.
Laundry's going, tea's made. That's the
whole plan.
Reset day.
>> That's a full week of content built from
one character sheet and one voice, with
me doing nothing but describing it. Once
you've got a recipe that works, you
don't have to baby sit it forever. You
can take it off confirm and let it run a
bigger batch unattended, which is how a
week of vlogs turns into a month without
you approving every frame. Because it
remembers between sessions, the mug, the
creators and the voice all carry over to
your next chat. So you're not rebuilding
your product or your cast every time you
come back. For anyone making content at
any volume, this is the change that
matters. One person can now brief and
approve what used to take a small team
and the hard part stops being the
production and becomes coming up with
the ideas. The whole batch, all 11
videos, cost a fraction of what a single
shoot or an agency day would, because
you're paying in credits per render
rather than day rates and revisions. It
suits anyone who needs volume and
consistency more than one cinematic hero
video, a brand running UGC or a creator
posting daily. And if you only ever
wanted one perfect 30-second film, you'd
still art direct that yourself. But for
a steady stream of content, this is the
fastest route I've found. So you hand
Claude Fable 5 and the supercomputer
your own product or your own week and it
builds the batch while you approve each
step. And the part I can't do for you is
pointed at your own product. [music]
What it builds from a brief of your own
is the part I want you to see. Open
Higgsfield through the first link in the
description and hand it your first brief
today. Thank you for watching and I'll
see you in the next one.
IndyDevDan
Latest FIXING Opus 5: PROOF that Prompt Engineering IS NOT DEAD
▶ Watch on YouTube
What's up, engineers? Andy Devdan here.
Like me, you've probably gotten sick and
tired of Opus 5's insanely verbose
responses and its overuse of phrases
like loadbearing, worth stating plainly,
here's the honest truth, and a bunch of
others. Or maybe you're tired of the
Anthropic team trying to take credit in
your Git commit messages for
intelligence you paid for. Or maybe you
notice Opus 5 is burning your cash with
way more output tokens than any model
before it. You're not alone. Myself and
many engineers feel the exact same way.
Obus 5 is one of the best
state-of-the-art ultra smart models and
one of the worst state-of-the-art models
ever released because it talks like a
complete smartass. In this video, we
turn smartass Opus 5 into a precise
senior engineer that's enjoyable to work
with. How are we going to do that? We're
going to use one of the most important
skills any engineer using agents can
learn. You know what it is? It's prompt
engineering. The skill that was once a
complete joke is now the most important
skill for [music] any engineer to master
to scale their impact with agents. If
you hear someone say prompt engineering
is dead, completely ignore them. They
have no idea what they're talking about,
I've been engineering for 15 years now,
and one thing that all the best
engineers I've worked with know how to
do extraordinarily well is this. Above
all, they know how [music] to
communicate with their technology. And
guess how you communicate with agents?
Yes, prompts. There are two ways [music]
to prompt engineer your agents. Most
engineers only use one when the second
is most powerful by far. By the end of
[music] this video, you'll have the
prompt engineering expertise to make
your AI agent communicate like a precise
cracked senior engineer. No matter the
model that's running today, it's Opus 5.
Tomorrow [music] it'll be something
else. Let's fix Opus 5 with system
prompt engineering.
What are the two ways to prompt your
agents? There's, of course, the user
prompt, which you're very familiar with,
the single task at hand. But then
there's the system prompt, which is the
law for every task you hand your agent.
Most engineers fixate on skills, skills,
skills, skills slash this slash that
slash plan/grill/re.
Don't get me wrong, skills are powerful.
The user prompt is powerful, but the
system prompt is vastly more useful and
most engineers never even touch it.
Because the system prompt is where every
single word you write is multiplied over
every single user prompt. The system
prompt is essential for setting up great
communication patterns with your agents
and for reducing those expensive Opus 5
output token costs dramatically. The
system prompt is where the real leverage
is. Let me show you exactly what I mean.
in VS Code. Here we have the simple
three file structure. I've clone down
Zuck's the future is for everyone. We're
going to do what all of us degenerates
do nowadays. We just summarize a long
article like this. We're not going to
read this whole thing. We don't have the
time. We're too busy with 10 million
agents open in our terminal window. To
do that, we're going to use a standard
powerful clawed opus 5 agent. As we work
through this, we're going to constantly
improve the system prompt we're passing
in to clawed code. I have this simple
command here called compare. So, let's
go ahead and fire it off inside of a
terminal. We're running herder. This is
my terminal multiplexer of choice.
You'll see exactly why in a second. If I
type J, you can see all the commands in
this directory. And what I'll do is just
type J compare. And we'll start with
zero. I now have a brand new workspace
open. Let's go and hop into it and see
what's going on. You can see here I'm
running two cloud code instances side by
side. On the left we have smartass opus
5 and on the right we have senior opus
5. Right now these are both in smartass
mode. This is just the default system
prompt and the default clawed opus 5
model with no tweaks, no changes, no
prompt engineering to adjust and improve
the system. As you can see all of the
ticks and problems and verbosity of this
model are coming alive right now. Zuck's
blog is very very long, but you can see
here these models spin for a long time.
The one on the left is really really
going off here with a lot of
information. Okay, so 53 seconds, 35
seconds. Most of that was just output
text. Look at how many output tokens my
agent just torched responding to me.
Most of us truly were not reading this
whole thing. Just like we're not reading
Zuck's blog, we need concise, compressed
information so we can move on or decide
to invest more. A lot of engineering
with agents is about figuring out where
to spend your time. So let's improve the
system by prompt engineering the system
prompt.
So in VS code I have senior opus 5
system prompt and this is getting passed
directly into that herder pane. You can
see here we have one version that's just
claude code opus out of the box and then
we have the other where we're pending
the system prompt file using this
argument flag. So that's what we're
going to do here. And so if we open up
this file right now we have nothing in
there. So we just got the default
response out of it. Let's improve it.
Let's write a great system prompt to
steer the behavior of every single run,
every single input and every single
output from our model. What we're
building here is a clear concise
communication document. I am typing this
by hand. When I'm working with pieces of
autonomous software that's going to be
multiplied over many, many, many runs, I
do it by hand. If you don't understand
your technology, if you're just vibes
slopping everything, you're not going to
get the results who someone paying
attention is going to get. We're going
to start really simple here. We're
appending this to the system prompt.
Clear, concise, actionable
communication. And then we're going to
talk through the purpose. First things
first, you and I maintain a no BS,
clear, concise, actionable relationship.
Notice how I'm talking to my agent. I'm
not giving it a role. I'm just talking
to my agent as almost if I'm talking to
like another engineer, okay? And and
kind of setting the standards. We
maintain a no BS, clear, concise,
actionable relationship. Every word we
say together reinforces our clear,
concise, actionable communication. I'll
jump through some of this. We're here to
solve problems and create value. And our
communication reflects that. And then
I'm going to start the next section,
instructions. And this is going to be
where we dial into several ideas. We'll
have four sections and then one kind of
recap section. Um, this will be the
examples and then we'll work through
these sections one by one. But I'm
writing this out to be able to reference
these sections so that I can clearly
communicate inside the system prompt
what I want my agent to focus on. Okay?
So, pay close attention to the details
throughout instructions to maintain our
great communication patterns. And then
I'm going to do something here that I
think is really important when you're
prompt engineering. I'm going to explain
why. Why? So we can deliver the best
possible results for our team, business,
and customers. Just walking through
things here really, really cleanly,
really concisely with my agent. Let's
pause for a second here. Let me just
actually delete this and just kick this
off again. So let me save that text
there. Let's go back into our terminal
here. Let's close this down. Let's fire
off our J compare uh with purpose. So
it's going to just rerun that exact same
thing. Already we have one kind of clear
change in our system. And you can see
here at the top we do have that
reference append system prompt. We're
attaching that. But already we have made
a small change in our purpose. This is a
good start. As you can see here, it's
not going to change a lot of the ticks
and the problems inside the model. As
you can see, there's loadbearing right
there. That's one of the classic ones
with Opus 5. It's using a lot of dashes,
but it is communicating a little bit
better. You can see that one came in at
35 seconds. Not a huge difference at all
from previous runs, but you can see here
if we looked through these, we would
probably see some glimpses of clear
communication, but we need more than
this. This is all very high level. Let's
get more midlevel, more lowlevel with
our system prompt. So, I'm going to copy
this in. And now we have instructions.
Okay, so purpose, instructions. Let's
start giving some real [music] clear
communication patterns.
Now we get into like concrete prompt
engineering patterns that are repeatable
across all of your work. For instance,
here's one clear pattern. Positive
patterns and negative patterns. And so
what does this section do? Replicate the
and I'm going to just write out a H4
here. Positive patterns. And of course
we'll have that and then we'll have
negative patterns. And so we're just
going to walk through this. And you
know, as you can imagine, I'm going to
clearly reference this section as
behavioral references. So, I'm saying
replicate the positive patterns. And
then I'm going to, of course, say avoid
the negative patterns. So, we want to
replicate these and we want to ignore
and not do these. Okay? So, we're doing
both positive, do this, and negative.
Don't do this. And now, we can just
start writing out these bullets. What do
you want your agent to do? What do you
want them to not do? I'm just going to
start from the top. And this is one of
the important things that is missed a
lot. So, I'll say I always see the last
thing you write first. Place the most
important information there. As you're
working with these agents, you have
likely noticed this, too. You see the
last thing. So, it's important that the
last thing you see is the most
important. So, you can, if you want to,
skip all the other stuff. What else do
we want the agent to keep doing?
Positive patterns. Use plain specific
language. State each fact once. There's
no need to repeat anything. Match the
level of detail to the level of task and
request. Okay, so this keeps the agent
aligned with the amount of detail and
effort you're putting in. Challenge
incorrect assumptions directly and
explain why. We do not want any sickle
fancy here in our responses. We want a
useful engineering partner. Optimize for
clarity and engineering value, not
quotability. I don't want to have a good
conversation. I want to deliver
engineering value with clear, concise
communication patterns. Okay. One more
note here on domain terminology. Use the
simplest domain terminology that
compresses information. So just some
positive patterns here. You want to
write what you want to see. And then of
course the negative patterns. You know
exactly what we want to add to this.
Avoid words and phrases in this list.
Okay. And then here we can just go crazy
with all the obnoxious things that this
model writes. We can do things like
loadbearing. That's going to be at the
top here. We can do things like worth
stating plainly. Here's the honest
truth, the [snorts] real tension.
[laughter]
And then there's carry the argument.
There are a bunch of these. Put
whichever ones here you want. A couple
things I don't like. Personally, I don't
like avoid analogies. Discuss what's
right in front of us. Not a huge fan of
analogies when working. I just want to,
you know, focus on the thing right in
our face. Of course, do not overuse em
dashes or dash chaining. I think some of
these are fine. My writing style
personally has a decent amount of EM
dashes. I have had to personally dial
these back because I don't want people
to think I'm just spamming out using
models when I'm replying. I think it's
really rude if it looks like you're just
responding with AI with zero thought in
specific circumstances. Do not flatter,
praise, validate, or agree without
reason. Again, no sick fancy. Just kind
of re-emphasizing the idea. Uh, do not
use decorative headings, emoji, or
motivative language. Again, I just want
this model to talk like a cracked senior
engineer. Let's just solve the problem
at hand. Avoid semicolons, fragments,
non-standard punctuation. I just want a
normal conversation. This is one really
important section in your prompt
engineering. You can use this when
you're building plans. For the system
prompt specifically, this has a lot of
weight to it. We're explicitly talking
with our model saying, "Replicate these
things. Avoid these patterns." So, how
else can we prompt engineer our system
prompt for consistent results across
every single execution? Before we do
that, let's actually run this. Let's see
how this is developing. Okay, so I'm
going to go ahead and just flip this,
place this here, go ahead and just
rerun, trash this session, and let's run
it again. So, now we have pause
negative. All right, so let's rerun a
new pair of cloud code again. On the
left we have our smartass opus. No
changes. And on the right we're going to
have our smart senior concise
engineering opus. Okay. So let's see how
it does. Now right away you can see
something very very cool. Very few
dashes here. We do have a couple dashes
showing up. I don't see loadbearing
anywhere [laughter]
which is good. I'll do a quick search on
that in a second. But there's a bunch of
these obnoxious keywords that are not
getting referenced. 31 second execution.
So we are speeding things up a bit. And
by speed I mean it's outputting fewer
tokens. So, we still have some dashes.
We didn't say never use them. We said
use fewer. The language is looking a bit
clearer. We have clear headers. You
know, no analogies. We do have a dash
chain here, private by default. Maybe
you're okay with that. Maybe you don't
want to see that. That's up to you. You
can prompt engineer that however you
like. You can see here we're making some
improvement. And we're starting to save
on the output tokens. Even in your
subscription, your output tokens still
churn up the most of your usage. So,
it's important to keep that in mind. We
can do this at the system prompt level.
Why? because we want to affect every
single input and output from the agent.
Your prompts in and its responses out.
If you have something that you want to
apply globally, that is the time to
reach for the system prompt. So, let's
keep moving. How else can we improve our
communication patterns with our agents?
So, number two, this is a really
powerful one that I'm a huge fan of and
I think you will be too. Check this out.
We can use reference points. What is
this? we use. And again, I'm talking to
my agent, right? I'm talking directly to
them. I'm not giving them a weird funky
role. I'm just talking to them. These
models are getting smarter. And even the
lower class workhorse models, the A and
B tier models that aren't your S tier
state-of-the-art models, they respond
really, really well to this additional
direction. We use reference points to
communicate quickly with each other. And
now, what do I mean by that? What is a
reference point? Use numbered lists and
markdown headings when they improve
navigation when presenting three or more
findings. And this is stuff like uh
decisions, options, risks, questions, or
actions. Assign everyone a short code.
Okay, so what does that mean? So use D1
for decisions. And what I'm really
looking for here is D1, D2, and then DN.
You can imagine this going on, right? So
I'll do this and do a dot dot dot. And
I'll skip over this in the video so you
don't have to wait for me to type all
this out. And then we do dot dot dot for
continuation. When we don't have
something here, I want it to invent new
references for sections we don't have.
Preserve same codes throughout the
conversation. Do not create codes for
short, simple answers. This is really
cool. So you can see this in action
right away, right? Let's go ahead and
fire off our system and see how the
system prompt materially impacts every
single prompt. Because again, that's the
scale of what we're dealing with here.
The system prompt is the law for your
agents. It affects every single task.
Let's delete the previous session. Bam.
And let's rerun. This is going to be our
ref points. So now we have that new
workspace. We're comparing side by side.
And of course, smartass on the left and
we have our senior opus slowly improving
with each change we make. So you can see
here we have the P's, we have the Rs
coming up. And so we have risks there.
And now our agent is even referencing
one of the Rs R six. Okay, so now we can
just jump to the reference. Our agent
isn't repeating itself. It's not wasting
time in tokens and it's repeating the
risks here. And then if we scroll back
up to the PS, I assume these are the
promises that Meta says it's going to
actually make. What is Meta saying it's
going to ship? It doesn't really matter
what the P stands for. The fact is that
we can reference them now instantly.
Okay, so for instance, like I can
continue the prompt here, right? So, if
I said talk more about R6, and of
course, my agent knows exactly what I'm
talking about when I say R six because
we've created this quick language
together. This is risk number six in
this section, existential RSI,
self-improvement risk that Zuckerberg is
talking about here. And our agent is
just going to kind of continue breaking
things down here, right? What R six
actually says, the section claims and
sequence. There it goes. And it's going
to break down that section more. We can
jump into it. And you can see here we
have another reference here. We have FS.
These are findings. That's exactly what
we've encoded here. Fs for findings. And
so our agent is just using our reference
system. And again, like the key I really
want to communicate here with you for
your engineering work is like um great
communication is great engineering and
vice versa. Knowing how to properly
communicate with your technology now
more than ever is a massive advantage
you can use. It's so underutilized and
it's so powerful. You can see here we're
starting to improve things. We're adding
layer by layer. Let's add some more
really important [music] layers.
What is the third improvement to our
system? Let's go ahead and dial into
this. Hard operational boundaries. Okay,
one of the very annoying things with
some of these state-of-the-art models is
that they will do things that they have
not been requested. This is part of
their reinforcement learning. This is
part of the loops that they're going
through when these models are being
trained where they're just taught to
find the answer at all costs, no matter
what it takes. Opus 5 is like really
guilty of this. Comment down below. Let
me know if you've experienced this. Opus
5 will just find and call out and
reference problems you did not even ask
for remotely and it's trying to pull
everything together. It's trying to do
as much as possible. But in that um it
actually loses focus pretty quickly. So
hard operational boundaries solve that
problem. Let's write that out. So in
addition to clearly communicating, it's
important that we clearly communicate
our work operational boundaries. What do
I mean by that? So deliver only what was
requested at the intended scope. That
single sentence really is this right?
But we can add more detail. Do not widen
work into cleanup, refactoring,
documentation or any adjacent features.
Just stay focused. That's what I'm
saying here. Stay focused. Do not
speculate on abstractions for future
requirements. Do not claim completion
without evidence. Very, very important.
These models are really good at that.
Now, this is not a great line to add,
but going to add it anyway. And then
here's another one. So annoying when
these models do this. Never add a
co-author to a commit message. Like nice
marketing trick for Enthropic, but they
got to stop doing that for completed
work. Concisely restate, but do not
overload with response detail. There's a
better way to say that, but you kind of
get what I'm saying. Opus 5 specifically
will respond with everything it's done
and kind of like do a big recap. Don't
really want that.
And let's go ahead and stack another one
on here. This is a really powerful one
that I think is very useful and it
really starts veering into the fact that
the system prompt can help you control
your inputs and your outputs. This one
is aliases. So, let me break this down.
Aliases are reminders of great
communication and patterns we want to
uphold. When you see these exact
aliases, expand them and act as if their
expansions were given to you directly.
And then we want to clear this up to
make sure we're being very clear. If
these are referenced in a longer string,
they are not aliases. Do not expand.
Okay, so these are exactly what you
think they are. They're short codes.
They're bash aliases. They're
expansions. They're commands inside your
system prompt. And if you want to, these
can reference your commands, your
skills, and other things you have. But
you can also just write them in line
like this. STR equals simplify,
compress, and repeat your response.
Let's write a couple more and then we'll
demo these. Okay, so Eli, you already
know what this is going to be. Explain
this like I'm We don't want to do five.
Explain this like I'm 18. Simplify your
language. Shorten your response. What
else? Uh, one of my favorites. Focus on
what matters most here. What's the true
signal? What's the true value? Boil your
response down into the most important
thing we need to focus on. And then ref.
Let's just do one more here to reference
our previous work. Cuz once you start
building up your system prompt, you can
reference other sections. So I'm going
to say I'll rewrite your responses with
reference points, which is exactly what
we dialed in above. So now let's try our
improved enhanced system prompt. Same
deal. Open the terminal. kill this
session with our aliases. So, we're
going to compare, boot these up, and now
our agent has a little bit more guidance
into exactly what it's going to do.
Okay, so you can see there we have C123.
That's our core thesis. We have our
risks addressed. This is a much better
generation really using all of our
reference points. Very nice. And done in
30 seconds, done in 30 seconds on the
left. So, you know, sometimes the
generations are just fast. These are
nondeterministic systems, so they're not
always going to be generated faster. And
in fact, we can improve that, which
we'll do in a second here. But remember,
we just added those aliases. So, let's
run scr simplify, compress, and repeat.
This is effectively a micro scale you're
giving your agent. So, there we go. It's
compressing that. It's simplifying it,
and it's repeating it. So much more
simplified. I can act on this
information a lot faster. Even if our
agent responds with a lot, which again
we'll tweak in a second, we can improve
on that. So, let's run another right
focus. This is focus on what matters the
most here. here. So our agent sees that
and it's going to do it. Now this should
be a lot more concise. There we go. The
signal alignment is being redefined from
the model holds the lab's value to the
agent holds yours. Everything else is
downstream of that. This is a speaking
pattern from Opus that you might want to
get rid of. Totally up to you. We have
why it matters and then pulling some
memory stuff here. And then we have our
final paragraph there. Right. A couple
things I want to do here. And once you
like set up this foundation inside of
your system prompt, you can really start
to control it and fine-tune it. These
patterns and these sections are really
important because we can now collapse
and we can jump into anyone we need to
very very quickly by hand. We are
controlling the experience of our agent.
Again, I really want to stress this
point. The more multiplicative the thing
you're working on is going to be for the
rest of your work, the more you should
stop, slow down. Don't just throw a
prompt at this. Even if it's a
state-of-the-art model, as you can see,
the state-of-the-art models have
problems. They might accomplish your
goal, but they'll add or do a bunch of
other We're being really, really
clear here. We're doing this by hand.
We're slowing down. I know a lot of
people are using Whisper Flow. They're
just talking in to their device. There
is going to be a level of detail that is
missed, and you're going to be very
exposed and very leaning on the normal
distribution of what the models can do,
and you're going to be inside that zone.
If you're just viro prompting things,
there's a time and place for that, but
the high leverage places is when you
want to slow down and step in to the
loop. And I mean old school, hands-on.
So anyway, let's look at this. So I want
to do some positive. Where do I want to
put this? This is definitely going to be
positive. And I want to say if you can
communicate the idea in one paragraph
instead of two, do so. And I want to say
without losing valuable information, do
so. Same idea for one sentence versus
two sentences. So I'm just saying be a
little bit more compact, be a little bit
more concise. This is important. Do not
repeat yourself. State every idea once.
Only repeat if it's relevant to
subsequent queries. Yeah, something like
that. Because we really don't want
repetition. And then I want to add
another positive pattern here. Don't use
overloaded terms that could mean more
than one thing. Use the simplest word
that satisfies the idea you're trying to
communicate or I should say simplest
words and maybe PN that. You might be
like, "Wow, you're getting really
detailed here." And yeah, that's exactly
right. I am getting really detailed
here. And we can see the results of
that. So, let's boot this up again. And
this is going to be uh fine-tuned. And
now let's go ahead and see what this
execution looks like. And to be clear
here, we are using append system prompt,
not overwrite system prompt. But let's
see how uh this execution runs now.
Fewer dashes, no loadbearing. We have
our reference points policy ask again
with reference points. And our sentences
are simpler. They're more concise.
They're not getting overloaded where the
argument is the weakest. Still
outputting a lot here. This is a pretty
big one. 35 seconds. So, not great, but
it is going to be better than this
generation on the left. Again, these are
nondeterministic systems, so it's not
going to be consistent. 43 versus 35.
Not bad. We could probably do better in
our compression. If this happened
though, in this specific case, we can
just throw an STR at this to get a
simplify, compress, repeat, or we can do
an ELI. Throw an Eli at it. explain this
like I'm 18 and there you can see the
language is really going to get boiled
down and sometimes this is just useful
right it doesn't mean you're an idiot if
you want simpler concise language it
means you're effective but you can see
here even still the agent is saying a
bunch of stuff so I might run a str
after this simplify compress and repeat
so as we start stacking up the user
prompts in between every execution of a
user prompt the system prompt is still
going to be ever present so you can see
argument really broken down here AI is
safest when everyone has it not when a
Few model labs control it. Other labs
say AI is dangerous. Lock it down. He
says when just a few people have it with
enormous power, that's never gone well.
One safe AI fails because people want
different things. Any single AI has to
pick whose values win. He's got a lot of
really great arguments in here. Highly
recommend you read and then compress if
you need to uh the kind of core ideas
coming out of Zuck's argument here. With
these like uh CEO posts, it's always
like understand the incentives. when
you're not leading the AI race, of
course, you would have a narrative
something like this that you would want
to share. I agree with him, frankly, on
a lot of this stuff, but it is very,
let's say, positional of him to put out
something like this. That's not the
focus of this video, though. I'll link
that in the description for you if you
haven't seen that yet. One more thing I
want to show you that can really, really
dial in the performance of your models,
and it's a prompt engineering technique
as old as time. Again, if someone tells
you prompt engineering is dead, don't
listen to them. The PI coding agent has
a small system prompt. Cloud Code just
got rid of a lot of their system prompt,
that is not the signal to not use a
system prompt. That is the wrong signal.
They're doing that because the models
can just do a bunch of stuff without
needing it. If you want the model to do
specific things, to perform in specific
ways, to not torch your output tokens,
to have cool unique capability like
aliases or reference points or hard
operational boundaries or patterns that
you want to see and words and concepts
you don't want to see, you must system
prompt engineer. The system prompt is
the most effective place to prompt
engineer because it's applied across
everyone of your prompts going into your
agent and the responses coming out of
your agents. Okay, so one more section
here for you. Check this out.
Examples to kind of put it all together.
Here are concrete examples of how we do
and do not communicate together.
Replicate how we do communicate and
avoid how we do not communicate. a lot
like our positive and negative prompting
section except the main differentiator
here is we're giving real examples. So,
it's just like training data. User is
legacy JSON still referenced. To-do, no,
the only match is the file itself. There
are no imports, blah blah blah blah blah
or documentation links. Very good. Um,
we could also, you know, make this
tighter. The only match of the file is
itself. We could make this the to-do.
And then here's what not to do. Great
question. I will research the repository
and blah blah blah blah blah blah. Just
do the work. Respond concisely. Right.
And then here's another one. Engineering
recommendation. And it looks like I'm
missing my user prompt here. So, I'm
just going to add that here. User.
Should we add reddus to this system?
Write something just like this to-do. Do
not add reddus here. This process has
one writer, stores from SQL, and has no
cross host coordination requirement.
Reddus adds a failure domain without
solving the correct constraint, the
current constraint. What not to do? You
are absolutely right. Blah blah blah
blah blah. Again, just a really simple
prompt engineering technique that is
older than time at this point. Um, I was
writing these examples back when GPT 3.5
was the model, GPT4, when there was
really no anthropic. This is still
relevant today. And that's how you know
something is a valuable skill to learn
when years ago it was still relevant.
These concepts, these patterns, these
principles, great engineering is great
communicating. These things just kind of
don't go away. That's how you know you
have something valuable to hold on to.
Now, we can continue doing this. And
kind of one big idea, one Easter egg for
whoever's still watching, engineers that
are really trying to absorb this value.
One thing I want to mention here is that
you can boot up a model. Let's say you
want to boot up a cloud fable model and
you want to run that same prompt that
we're running. So let me just go ahead
and pull that explain this. Copy paste.
And so Fable 5 I think is a great model.
I think it responds well. It doesn't
have as many ticks and problems and
repeated patterns and verbosity as Opus
5 has. So what I like to do is in
context distillation. In context
distillation is exactly what you think.
That's what examples are, right? It's in
context distillation. Do this, don't do
that. And you can do this by pulling out
specific context, specific responses
from models that you like. For instance,
like this oneline core thesis here.
Maybe we like this. We just copy this
out and we create another one here.
Summarizing a blog. And then user is
summarize the blog. and then XYZ. We
just template this out. And then we have
our to-do text. And then we can paste
the response from the model that we
like. And we can go even further. We can
tweak the response from the model that
we like. Get rid of a couple of lines.
Make sure it's formatted properly. So we
can do a dash search to get rid of all
of our dashes. All of our framing. Maybe
we like that one. We keep that one in.
Super intelligence. Super intelligence.
Super intelligence. This should be more
concise and compact like this. So maybe
we like this version. Then we say not to
do text. Text. We hop back over to the
previous version, right? We go to our
smartass opus and we copy all this. This
is not what we want to see. So, just a
bunch of extra information and we want
to see a concise, simple response.
Really just get to it. And really, I
would want to see something like this, a
little bit more broken up, easier to
read. Even Fable has some long kind of
run-on sentences here. Might break this
up a little bit more into a couple
section. And then we would fire this
off. We have a new version here. Let's
go ahead and run this. It's super
correct senior opus. And then we can run
our comparison here. So now our agent
has examples of how we like to
communicate. It has reference points. We
have cleaned up the language. We taught
it to respond concisely with fewer em
dashes. It's not loadbearing anymore. We
have reference points. This one does not
have reference points. So thankfully
we've encoded a way to add those. So we
could just type ref. And now our short
aliases are going to expand the
response. There you go. You can see we
now have references. You'll notice this
response from Smartass took 41 seconds.
Our original response over here took
just 22 seconds. So, we are saving our
output tokens. And again, if you need to
spend output tokens, spend them. But in
this case, you know, when we're uh
summarizing things where we're just
getting concise responses, we don't want
all of that. You know, you can see here
we have a nice response. We use our ref
alias, which we have encoded into the
system prompt. If we just search for ref
equals, you can see exactly what this
expands into. You could point this to a
scale. You could point this to a
command. Whatever you want to do here,
you can do. My whole point here is to
communicate with you that if you really
want to control your agent and get the
best results, you must know how to use
your tool. A lot of engineers are
overleveraged on the user prompt and you
get literally less leverage. You get
less output by fixating on the user
prompt. Opus 5 and all models coming
out, they are going to be smart asses.
Quite literally smart. What we want to
do is keep the smart but drop the ass.
[laughter] With system prompts, you can
do that across every user prompt you
write. You can use these exact prompt
engineering patterns to guide Opus 5 and
whichever state of the art models come
next. If we don't like this response, we
can shoot up another alias at it. Eli,
explain it like I'm 18 or whatever age
you want to set your uh response
reasoning level to. You can add an Eli
5, you can add an Eli 10, whatever you
want to do. Great engineering is about
great communicating, right? great
communication with your team of
engineers, with your agents, and most
importantly with yourself. And this is
how you can tap into that. You've
noticed this if you're using agents on a
daily basis for many tasks. You and I,
the developer, are the bottleneck. It's
not the model. It's not the tools. It's
not anything else. The hard part now is
communicating quickly, concisely, and in
the most value accreative way. I'll
leave this cracked senior engineer
prompt linked in the description for
you. It's super super simple. You can do
the comparison side by side and see
yourself improving every time you add
something by running the just compare
command. I'll also add my classic readme
and just install and all the quick start
so that you don't have to waste your
time doing any setup and configuration.
I want this to be as simple and quick
and easy for you to use. On the channel,
we've been focused on much higher levels
of abstraction on top of agents like
software factories, agent sandboxes. Uh
check out last week's video if you want
to dive into some really advanced
agentic engineering concepts. I'll link
that in the description as well. But
every once in a while, it's going to be
really important for us to step back
into the fundamentals. You don't want to
miss the atomic units, the primitives
that make up the larger powerful
composition pieces because if you stack
a bunch of agents together and you burn
a bunch of tokens doing in half the time
by just putting together the right
system prompt, you want to be doing that
instead. Don't waste time using really,
really clever solutions when there's a
much simpler solution. And engineering
is about getting the job done, not
looking cool, not doing what everyone
else is [music] doing, not listening to
the idiots that say prompt engineering
is dead. That returns false. Your agent
is just another tool. If you understand
your tools, [music] you'll understand
the results you can get from your tools.
Use your system prompt as the law for
every [music] task you send to your
agents and use your user prompt for
individual tasks. If you got value out
of this video and you made it to the
end, like, subscribe, drop a comment,
and let me [music] know how you're
leveraging your system prompt to get
outsized results and to clean up some of
the insane verbosity from models like
Opus 5. You know where to find me every
single Monday. Stay focused and keep
building.
Matt Wolfe
Latest AI Rants & Rabbit Holes: Does AI Make Our Lives Better?
▶ Watch on YouTube
Could not fetch transcript (it might be disabled or unavailable). Error: 
Could not retrieve a transcript for the video https://www.youtube.com/watch?v=IeT5Jn7d1aM! This is most likely caused by:

The video is unplayable for the following reason: This live event will begin in 7 hours.

If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
AI Master
Latest How to Create an AI Study Agent with Claude AI (No Code)
▶ Watch on YouTube
Most people spend hours watching
tutorials and taking notes, only to
forget most of what they learned by the
next day. That's because consuming
information isn't the same as actually
learning it. But Claude can dramatically
speed up the way you learn by turning
any learning material into an
interactive learning system that helps
you understand, remember, and apply what
you learn. By the end of this video,
you'll have a complete workflow you can
use to learn any skill. We begin by
creating a separate Claude project for
learning. I also add a short description
so the purpose of the project stays
clear. A normal chat can answer one
question. This project keeps the
teaching rules and course materials
together. The first mistake is starting
with a vague request. It doesn't know my
level, my goal, or my deadline. It can't
tell whether I need exam prep or a
working skill, or whether I learn faster
from examples than from questions. So, I
open the project instructions and
explain how I want Claude to teach. This
prompt can controls the whole lesson
instead of one answer. First, Claude
checks what I already understand. Then
it chooses an explanation that matches
my level. The next rule stops the
finished solution from appearing
immediately. That matters because
copying an answer isn't the same as
learning the method. The rule about
mistakes is just as useful. A good tutor
explains where your thinking went wrong.
It doesn't simply delete your answer and
replace it. The final rule gives us
something useful for the next session.
At the end, the project records what
worked and what still needs practice.
These instructions can also change for
different subjects. In language
practice, the tutor corrects mistakes
and brings new vocabulary back. On a
coding lesson, it asks for my
explanation first, then shows the fix.
The goal is deciding which parts Claude
handles. The learner should still do the
thinking that builds the skill. I test
the setup with one simple topic. The
first reply asks what I already know
about prices and demand. The next
question covers why I need the topic. I
explain that I want to understand the
idea and use it inside Excel. One final
question checks how much study time is
available. I give it 7 days with 30
minutes each day. Claude confirms that
budget before it builds a single lesson.
That small check keeps the plan
realistic because the schedule starts
with I actually have. That totals 3 and
1/2 hours focus study across the week.
Instead of writing a huge course, Claude
builds a short path around that goal.
The first lesson uses one simple
example. Day one covers the basic price
and demand relationship with round
numbers. Day three moves into the
elasticity idea behind that
relationship. Day six ties everything
back into the spreadsheet work I
actually need. Later lessons at practice
and a final check. Day seven ends with a
short revenue exercise inside my own
spreadsheet. The schedule isn't locked
after day one. A rough first exercise
can push the spreadsheet task back while
an easy session moves the practical work
forward. Claude adjust the order around
what actually happens instead of
pretending the first plan was perfect.
That's the first real difference between
a chatbot and a tutor. A chatbot answers
the question you type. A tutor changes
the lesson around what you need. All
right, let's park the study agent for a
minute because we're staying inside
Claude anyway. I do all my thinking in
Claude, but every serious deck hit the
same wall. The charts and tables never
looked presentable. Quick heads-up, this
part is sponsored by Perceptus. It's a
deck platform that plugs right into
Claude. And honestly, I use it almost
every day. Claude is great for the
outline and the research, but it falls
short on real charts and financial
breakdowns. So, you export it and
rebuild everything by hand in
PowerPoint. That easily eats an hour of
your day. Perceptus kills that step with
a connector right inside Claude. It
works from any chat, so whatever you
research turns into business grade
slides. So, I asked it for something I
actually needed. I typed, "Build me a
12-slide deck on my content roadmap."
Seconds later, I had clean slides with
my messy notes sorted into real sections
and titles. It even added a chart for
the growth numbers, and the whole
PowerPoint file stays fully editable.
That's the part one really love.
Perceptus doesn't just design the
slides, it structures the whole deck for
you. So, if you're still fixing Claude's
decks by hand, this is the fix. And my
timing here is kind of lucky because
they just shipped a new model called
Perceptus 2.0. Design Arena already
ranked it number one in its category,
ahead of tools like Gamma, and even the
Claude models. There's a link in the
description if you want to try it. Okay,
let's jump back into Claude and keep
building our study agent. The next step
is giving Claude something real to work
with. For this example, I've got a short
marketing guide, a page of audience
notes, and a sample campaign task. All
three files cover the same product, but
they don't tell the same story.
The guide explains how to choose an
audience and build one clear message. My
notes focus on freelancers managing
several client projects. The campaign
task targets small creative teams
launching a shared workspace. Most
people would ask Claude to summarize
everything. That would only create
another document I still need [music] to
study. A better first step is asking
Claude to organize the material around
the task. The source map pulls out only
the sections I need. Out of nine guide
sections, only four end up on the map.
It links audience segmentation to the
brief [music] and value proposition to
the core message. It also shows that the
call to action must support a free
trial.
Claude even quotes the exact line from
the brief that mentions the trial. Then
Claude catches the conflict between the
files. My notes describe freelancers
while the brief targets small creative
teams. That's a real conflict, not just
two different writing styles for the
same audience. That matters because the
strongest message changes with the
audience.
For freelancers, the guide suggests
protect and focus time. For small teams,
it suggests keeping shared work visible.
Claude also flags what the files cannot
answer yet. The final offer, launch
date, and proof points are still
missing. Claude lists all three gaps as
open questions instead of guessing a
launch date or a price point. The useful
part is how quickly I can check the map
later. I don't need to re-open every
file whenever a question comes up. That
gives me a clear steady order before I
make any campaign decisions. It also
keeps the lesson tied to a real task
instead of another generic summary.
Next, I decide which source Claude
should trust for each answer. Now Claude
knows which source should guide each
decision. The brief defines the audience
and the campaign goal. The guide
provides the method while my notes add
useful context. A personal note can
quietly override the actual task. Once
that source order is clear, every later
answer becomes easier to check. I can
trace a recommendation back to the
brief, the guide, or my notes. When
Claude has no source, the gap stays
visible instead of turning into a
confident guess. This is more useful
than a general summary. Claude shows
what matters and where it comes from. I
can scan the map in seconds and know
exactly which file supports each
decision. That same setup works with a
professional course or internal
documentation.
Start with the files for one skill or
one project. When two sources disagree,
decide which one Claude should trust.
When a detail is missing, make Claude
flag it instead of filling the gap. I
pinned that rule into the project
instructions, so every future chat
follows it. That's what makes practical
learning more useful than collecting
random information. Inside AI Master,
the academy turns that approach into
structured lessons. It includes over 200
lessons across roughly 30 hours of
practical training. You can test each
workflow immediately on your own
project. AI Master also puts leading
language models inside one workspace.
That makes model comparison easier and
keeps the workflow in one place.
Generation also uses lower token costs
than many separate subscriptions. Text
models are only one part of the
platform. You can generate images and
voiceovers inside the same workspace.
The platform also supports video
generation at lower token costs. The
character tools keep one hero visually
consistent across different generations.
You can publish that character directly
on the platform. You earn whenever other
people use your published character.
Generated content can also be shared
with the wider community. Sharing helps
that content reach more people. The
platform also runs a full content
factory. You build your own channel
inside it. And AI agents plan the
schedule, write ready to record scripts,
fill in the descriptions, and generate
the thumbnails. More than 12,000 people
already use AI Master. The community
helps users find feedback and new
connections. It also helps creators meet
collaborators and build an audience. The
landing page shows real cases from
current users.
The demo video includes real reactions
from people using the platform. Wow,
thank you so much AI Master Pro. Big
shout out, it helped me so much. I love
this service and I'm enjoying using it.
Getting access takes only a few clear
steps. Use the link below and open the
selling landing page. Click buy and
choose the annual subscription. The
annual option includes the strongest
available discount. The purchase
includes
money-back guarantee. After checkout,
open the confirmation email and you're
in. Okay, back to the learning agent.
Now, I want Claude to teach one idea
without giving me another long lecture.
A shop sells 100 items for $20 each. The
owner raises the price to $25.
Monthly sales then fall to 70 items.
Price multiplied by units sold gives
$2,000. The next question moves to the
new revenue. My first instinct is to
focus only on the higher price. Each
sale is now worth $5 more, so earning
more sounds reasonable. Instead of
giving the answer, Claude points out the
missing sales volume. Using both new
values gives $1,750.
[music] Claude makes me type the
multiplication myself before it confirms
the number. Revenue actually fell by
$250.
Follow-up question asks why that
happened. The extra $5 didn't cover the
30 lost sales. That explanation is more
useful than memorizing a textbook
definition. The lesson changes only one
number. Instead of 70 sales, the shop
now sells 90 items. That produces
$2,250.
This time, the higher price does
increase revenue because the price stays
at $25, the comparison is clean. Claude
isn't giving me another random example.
It changes one variable so I can see
exactly what caused the result. That
checks whether the idea transfers to new
situation. It's not just another
explanation with different wording. The
final question asks whether lower
revenue always means lower profit. We
still need information about costs
before making that claim. There is no
need to turn this into another finance
lesson. The question simply checks
whether I notice missing information.
This teaching style works far beyond
economics, SQL, a new language, a
certification you're cramming for. The
same basic pattern stays in place. You
get a question built in what you already
know and you have to attempt it before
any hint shows up. You can also change
the difficulty during the lesson. When
the questions feel easy, ask Claude to
remove the hints. I try that here and
the next question arrives with no hint
at all. When the lesson moves too
quickly, ask for a smaller example. You
don't need to restart the whole
conversation.
The next step can change while the
learning goal stays the same. The
learner still has to reach the answer.
That also makes the pace feel more
natural. Claude can slow down when I'm
missing a step then move faster once the
pattern is clear. I'm not locked into
one teaching speed for the whole topic.
Now I want to use the same idea inside a
real spreadsheet. The first task is
calculating revenue for every order. I
could ask Claude to fill the whole
column immediately. That would finish
the spreadsheet but it wouldn't teach me
much. Instead, Claude should guide my
first attempt. Claude first asks which
values create revenue. Each row already
contains the price and number of units.
I enter a simple formula in the first
revenue cell. The formula multiplies
price by units for the current row. When
I copy it down, both row references move
correctly. I spot check row 11 and the
formula correctly shows E11 * F11. The
next task [music] compares each revenue
value with the target in K2.
I write an IF formula using the revenue
cell and K2. The first result looks
correct, but the problem appears after I
copy the formula.
The revenue reference should move
because every row uses a different
value, while the target should stay in
K2 because every row uses the same rule.
Claude doesn't replace the formula
immediately. It asks which reference
should remain fixed. That question
points me toward an absolute reference.
Now the revenue reference moves, while
the target stays locked. I copy that
corrected formula down the entire order
list. Every row still points back to the
same target cell in K2. I checked 11
rows and the reference held steady each
time. The important part isn't
memorizing where the dollar signs go. I
need to understand which value changes
and which one stays fixed. Next, I ask
for a task that combines several
conditions. Claude begins by asking
which column contains the values I want
to add.
>> [music]
>> The revenue column becomes the sum
range. Then I choose the region column
and the selected region. The final
conditions define the start and end of
the month. The first day condition
includes the first day, while the second
stops before the next month begins.
Using the next month as the boundary
avoids guessing the final calendar day.
>> [music]
>> It also works when different months have
different lengths. The same boundary
trick still works if I check quarterly
totals instead of monthly ones. Claude
then changes the region and month. I
have to update the inputs without
rebuilding the formula. It also asks me
to predict the new total before the
sheet shows it. I switch the criteria to
the West region and the month to June.
The total updates immediately without
touching a single formula reference. The
spreadsheet is still doing the
calculation. Claude is helping me
understand the decisions behind it. This
is where the difference between help and
automation becomes clear. Claude can
explain why a formula breaks, but I
still choose the reference and test it.
The spreadsheet stays useful because I
understand what changed instead of
trusting a finished result I never
built. That is the balance I want from a
learning assistant. It should support
the work without removing the part I
need to practice. After a lesson, most
people create notes that repeat the
original material. That wastes time
because the easy parts get just as much
attention. A useful review should focus
on the moments where your thinking broke
down.
Claude can see which answers I changed,
where I needed a hint, and which ideas I
still couldn't explain clearly.
>> [music]
>> It can use that history to build a much
more focused revision guide. The guide
spends most of its space on the areas
where I struggled. In this session,
[music]
one section returns to the conflict
between two audience sources. It reminds
me which file should guide the final
decision. Then it gives me a new example
where two sources disagree. This time
the conflict sits between a pricing
sheet and an old sales deck. But the
same review system changes completely
with another skill. For language
practice, Claude can reuse vocabulary
inside a short dialogue. It can ask me
to complete the next line before showing
a correction. A later card can bring
back the same phrase in a different
situation. For coding, Claude can show
me a small function. Then it can ask
what that function returns. Another card
can hide one mistake inside the code.
Instead of giving the fix, Claude asks
me to explain which line caused the
problem. That's much more useful than
reading the same explanation again. The
format changes, but Claude still follows
the skills I need to practice. This only
works when I make a real attempt first.
If I copy every answer, Claude has no
useful mistakes to work with.
They questions give you generic notes,
and passive reading does the same. I
still need to explain the idea and test
my own answer. The active recall
questions now become interactive
flashcards. One card asks me to explain
an idea in my own words. Another asks me
to predict what happens next. A
different card asks me to find and
correct a mistake.
>> [music]
>> One useful detail is that Claude doesn't
need my answer to match its wording.
I can explain the idea in plain English,
and it checks the reasoning instead.
That makes the cards feel more like
practice and less like memorizing a
definition. Cards answered correctly
with high confidence disappear from the
session. The card about the audience
conflict stays in rotation for two more
rounds, and I only stop seeing it once I
answer with high confidence. Weaker
cards return later with a different
example. The goal isn't memorizing the
sentence Claude wrote. It's building a
review loop around mistakes I actually
make. Reviewing those mistakes helps,
but it doesn't prove I can use the
skill. A short quiz puts that to the
test in a new situation. The goal isn't
collecting more questions inside the
chat. I want to see whether the same
ideas still work in a new situation. The
Claude artifact gives that test its own
interactive space, but the artifact is
only the format. What matters is how
Claude arranges the questions. I used
the same three marketing files and the
work from this session, giving Claude
real material for every question. The
first questions check ideas I should
already remember. Then the quiz moves
into new campaign situations. One
question asks whether a message pulled
from audience notes should be used for
the current campaign. I choose it
because the message sounds well-formed.
Claude marks the answer wrong and
explains that the notes are exploratory
and target freelancers, while the
approved brief targets small creative
teams. The feedback stays short enough
to keep the quiz moving, and then it
goes straight to the next question. The
next campaign example asks me to choose
the strongest message for a launch team
whose work is scattered across chats,
documents, and trackers. I have to
connect that specific problem to the
message instead of choosing generic
productivity copy. Another question
shows a colleague drafting subject lines
and choosing a send time before the
audience, message, and offer are locked.
I need to identify why that sequence
creates polished work aimed at the wrong
target. The next question asks how many
days the free trial lasts. None of the
uploaded files includes that detail, so
the correct answer is that the available
sources can answer it. That checks
whether I can spot a real gap instead of
guessing. Another scenario asks whether
I should invent a customer testimonial.
I choose the wrong answer and Claude
reminds me that testimonials need
approved evidence. The same setup also
works for certification prep, where
Claude can turn course rules into short
workplace questions. After the final
question, the quiz shows my score. I
finish with six correct answers out of
eight questions. The two misses are the
source authority question and the
testimonial scenario. Six correct
answers out of eight puts the score at
75%
and it also shows which two skills need
more practice. Those results give Claude
a clear starting point for the next
session. A new quiz can keep the scale
but change the situation. Claude keeps
my two weak skills, but writes a
completely different campaign scenario.
The score itself isn't the useful part.
What matters is where I still get stuck,
because that shapes the next round of
practice. I ask Claude to build that
practice around my weak areas. The PDF
doesn't repeat the whole lesson. It
starts with the source mistake and the
proof we still need, and each section
gives me one correction and one quick
check. The Doc X plan spreads those weak
areas across seven short sessions.
Each session still stays under 30
minutes, matching the original weekly
plan, and later days bring them back
inside new campaign situations. The
final practice file depends on the
skill. For language practice, it could
include short speaking prompts. For
coding, Claude could create debugging
exercises. Here, it creates an Excel
workbook, because Excel was the
practical skill. One exercise checks a
copied formula with a fixed reference.
That's the same absolute reference
formula I built earlier in this session.
Another asks me to build a monthly total
from a new table. The new table uses
different column positions, so I can't
just reuse my references. The workbook
holds two Excel exercises, one built
around absolute references and one
around a fresh SUMIFS table. Claude
leaves both exercise sheets unfinished.
The answers stay on a separate sheet
called instructor check. That gives me
another real attempt before I see the
solution. Finally, I save a short
learning log for the next session. The
log records what I can now do without
help and tells Claude what should come
back next time. When I return, Claude
can start with the next useful question.
It doesn't need to repeat the entire
first lesson. Claude becomes much more
useful when it understands your goal. It
also needs to know which materials
matter. Clear rules help it guide you
without taking over the work. Start with
the smallest version that solves a real
problem. One project, the files you
already have, one lesson.
Every prompt from this video stays on
screen so you can copy the setup. And AI
Master Pro is still sitting there if you
want the models, the agents, and the
courses in one place.
Your call. Like and subscribe to the
channel.
I'll see you in the next one.
Ryan Doser
Latest Stop Using Claude Code Like a Chatbot
▶ Watch on YouTube
What it actually becomes once you treat
it as more of an operator, right? Not
just a coding assistant or a chatbot.
It's called operating systems, that's
the vernacular of the AI space. I call
them command centers. So, we have like
20 command centers across the whole
business. Each one is like a subject
matter expert domain for that specific
thing. So, if I were to open my other
tab here, I have a tax [music] operating
system, I have a finance operating
system, I have a YouTube operating
system, I have a school operating
system. Every single part of the domain,
the business, even personal. I literally
have a an entire domain that helps me
get out of jail free card. And I have a
thing called {slash} date night that
will go through every single the top 10
most recent posts, go through all of
them, watch, score them on ambiance and
novelty, and send me a top three every
Friday. And it's really a matter of just
like digging and deciding what is the
goal of this operating system, how will
I maintain it, and how will I get
infinite leverage from it.
The AI Advantage
Latest ChatGPT Plugins Finally Work!
▶ Watch on YouTube
Chat GPT just looked at my schedule, my
calendar, found the best opening, and
added a 30-minute break so I can relax
and recharge. All right, now we're
talking. I can go out there. But
seriously though, I didn't even open
Google Calendar in this entire
transaction, and these things were quite
hard to achieve just a few weeks ago.
You needed custom setups and da da da da
da da. Now, Chat GPT is all you need, or
Claude that is, okay? So, this is what
plugins can essentially do inside of the
new Chat GPT work, also in Claude
connectors, but today we're looking at
Chat GPT work as most people are using
Chat GPT, and I want to open your third
eye to what's possible with these AI
tools because it's unbelievable. So,
here's the deal at its core. A plugin
can teach Chat GPT a specialized
workflow and connect it to an outside
service. So, you can basically be like,
"Hey, you can see my Google Calendar now
just like my assistant or or my phone or
my own eyes when I open it, and you need
to be inside of Chat GPT work to do
this, okay?" So, once you have that
plugin set up, you can well, schedule.
It can help me make changes. It can hop
between other applications. You just
need to be inside of Chat GPT work for
this, and I know many of you will notice
already, but these capabilities, they've
been getting better, and I wasn't a big
fan of plugins, and now I kind of am.
You tab over to Chat GPT work, then you
go to plugins down here. You say connect
plugins. You can see already Google
Calendar here cuz I selected it, but if
you go to connect plugins, if it's not
up here, you should be able to find it.
Yeah, worst case you go to browse all
plugins, and then in here, well, very
worst case I have to search for it. So,
you know, calendar. There you go, Google
Calendar. Click that.
Click the three dots. Okay, mine is
already installed, but you know what?
For you, just cuz I love you so much, I
will de-install this. Now, I will
install again, and you should see
something similar to this. Continue to
Google Calendar. I select account with
my main calendar, and then the beautiful
thing is these permissions are only to
access the calendar. It's not going to
be looking at everything across Gmail,
etc. That's a different plugin, but if
you want to see it, you could click here
and see what services you're giving to
it. These are the six. All right, I'm
good with that. I'm going to click
continue. And there it is. Now we're
connected. I could try it in chat, and
the way you do that, well, now it's
already preset, but the way you would
get to this from a naked chat, is you
need to use the add button. This is a
bit different from Claude. So, just say
add calendar.
There it is. This is how you would
access it, okay? And hey, if you're
wondering what else you could do with
ChatGPT your work, well, I put together
a PDF with the top 10 prompts that you
could try today within there. And you
can have it completely for free right
now. First link in the video
description. All right, back to the
video. So, now I could ask, "What does
my schedule look like?" And for anybody
who's been watching the channel for a
while, you'll know I've been an avid
hater of these connectors, these
plugins, because they just weren't
reliable. You ask this type of question,
and it was like, "Well, you actually
don't have anything going on today." Or
well, there's two events today, when in
reality there's three. And getting two
out of three things is not good enough.
They were not reliable, okay? Now,
several things changed. The models
changed. They reworked normal ChatGPT to
ChatGPT work that thinks a bit more, is
a bit more thoughtful. The MCP standards
underneath also changed, although I'm
not even sure if this one is updated to
the new one. The point is there has been
several improvements over time, where
ChatGPT with uh Google Calendar login a
year ago was bad, trash, unusable. I
told you this straight up. I didn't used
to recommend it. Now, it just gets it.
It knows exactly where everything is. It
can call multiple things. It used to be
in a way that you could call one thing
and get that reliably, but if you wanted
multiple, mm, messy.
Now, you could do something like using
voice input. Okay, every time I'm in
meetings for more than 2 hours, put in a
30-minute recharge block. Now, you'll
notice that today's day has passed, but
this is smart. This is ChatGPT work, and
it thinks through the problem through a
wider lens than you might have
specified. That's how it is. It kind of
plans for the problem and looks to
resolve the entire category of work, not
just this one instance. That's the
product, okay? So, as you see, it first
didn't just look at my day that has
passed, but it set up a schedule task
where it's like, "Okay, I'll review your
Google Calendar for the upcoming 7 days,
and whenever there's more than two
consecutive busy meeting hours, I'll
schedule a 30-minute busy calendar event
titled to recharge, okay?" And it set
the frequency to check every hour, which
I think is totally overkill, but, you
know, good starting point. And then, it
just prompted itself, essentially, and
continued and said, "Okay, recording
block mastery da da da da da weekly ah."
So, now it didn't change anything
because I'm already beyond all of these
meetings, but it set up a schedule task
that should do this for my calendar. And
this is the power of plugins, but with
together with schedule tasking work,
it's just unbelievable. Like, you could
set up a few rules like this that just
run your calendar. And let me show you,
if I go here into scheduled, and then I
would take this post meeting recharge
one that just got created, and I say run
now here under the three dots, there it
is. It's going to do that for the entire
week, and then I could change the
frequency, right? I could say, "Hey, I
want this to run once a week or once a
day, maybe." There it is. Added them.
And right there in my Thursday, we have
the recharge block. These plugins work
so well now that you should be thinking
about your life in a new way now.
Together with this ChatGPT work product
or Claude co-work or whatever you might
be using, you can set up a few rules
like this for reoccurring problems that
just solve this category of problem for
$20 a month. And considering the fact
that paid plans are roughly used by 1%
of the working population on planet
Earth, it's ridiculous to me that the
edge you can get in terms of
productivity and ease by eliminating all
these annoying admin tasks, if you use
this correctly, it's just an amazing
opportunity, and I wanted to show you in
this video. Nice and easy. And finally,
I'll say, if you have a task yourself
that you set up some schedule task for a
plugin that you use all all time, leave
a comment below. I really want to use
which ones you use. Let's learn from
each other. Let's share the love and
share the knowledge. That's pretty much
it. My name is Igor Pogany. I hope this
was helpful to you and I will see you
very soon.
Higgsfield AI
Latest KÖK BÖRÜ | AI Animated Short Film | Higgsfield Originals (2026)
▶ Watch on YouTube
قر
س مختسن
ب
قاسر قزم
قازر قازر
Монда
шар
‏Miss
ش
Где лежат?
اجي
جم
شرم
Жи
بج
‏G
ب
А
س
Лдажал
бетай
ا
برن
قرب
بز
زم
م
او ماي قزم
ماي
قم
قم
‏Oh
The Zinny Studio
Latest This Workflow Makes AI Animation Look Real (Seedance 2.5)
▶ Watch on YouTube
If you've ever made an animated video
with AI, you already know the problem.
It looks like AI, faces change between
shots, hair moves on its own, the eyes
are empty, and viewers click off the
second they see it. Everything on your
screen right now is AI. Same scene, four
completely different looks, and none of
them have that problem because the
workflow I'm about to show you catches
everything that gives AI away before you
ever generate a frame. So, in this video
I'm walking you through the seven steps
that take you from a blank page to
scenes like these in minutes. Let's get
into it. Now, the most important step is
to generate the story, and here's the
one I am using for this build. It's
about a man called Victor Lustig who
conned scrap metal dealers and tried to
sell them the Eiffel Tower. All right,
so now the next step is to choose a
style based on the story we selected.
I'm going with faceted 3D because this
is a story about a man lying to people's
faces, so I need a face I can actually
watch him do it with. Now, picking the
right style took me the longest, so I
put everything I found into one place.
Let me walk you through it because we'll
be pulling from it for the rest of this
video. So, this is the style library I
built for you, 28 styles, and each one
is tagged with the niche it actually
works in. So, if you know your niche,
start there and it'll show you what
fits. And if you already know the look
you want, go straight to it. Now, open
any style and you get three prompts. One
makes your character sheet, one makes
your storyboard, one makes your video.
They run in order, so whatever comes out
of the first image gets used to generate
the next. Inside each of them, there's
one section marked as the only part you
write. That's your story, a paragraph or
two. Everything around it stays exactly
as it is. And if you want to take it
further, copy the prompt into Claude or
ChatGPT with your story attached to it
and ask it to merge the two. It will
generate refined prompts for you. Then
it is just copy and paste. That's what I
will be doing for storyboard and video
prompt. Real quick, if you want this
style library with all the prompts,
comment the word styles and I'll send it
over. So, let's go ahead and start with
the character sheet. I'll build it in
this first style, faceted 3D painterly,
then show you the same sheet in a
completely different style. So, you can
see this work with whatever you choose.
For this tutorial, I'll be using Design
AI and you can use the link to sign up
and follow along. Once you come into the
platform, the left side gives you home,
projects, asset and result and your AI
tools and the middle is just a chat box.
Go ahead and copy the character sheet
prompt, drop it in here, then scroll
down and paste your story into that one
section we just talked about. Now, the
settings. I'm switching over to GPT
image 2. 0 {comma} 16 by 9, 2K and
quality on medium and watch this. The
second I switch the model, the character
limit drops and my story is too long to
fit. So, I'm cutting it back to the
three sentences that carry the plot and
that fits. Let's go ahead and generate
and we'll wait for it to finish and
there it is. Now, look at what you
actually got because this is a lot more
than a character. Every person in the
story has their own block with different
angles and expressions and your
locations are built out along the
bottom. So, that's one generation and
the whole world of your video is already
decided. And now, before we move on, let
me show you one more thing. So, I'm
going back into the pack, opening
claymation and copying its character
sheet prompt. Same story, same settings.
The only thing that changes is the style
block. Drop your story in and hit
generate and there you go. Same story,
same characters, completely different
world. That's the whole point of the
library. Every style has its own version
of this prompt. So, whichever one you
pick, this still works. I'm going to
delete this one though because we're
staying with faceted 3D for the rest of
this build. So, the storyboard is next
and this is where that sheet starts
doing its actual job. Click on chat
editor on the character sheet you just
made and drop in the second prompt. Now,
this is where people mess it up, so stay
with me. You'll see image one written in
a few spots inside the prompt. That's a
live [music] link to your image and you
have to point every one of them at your
character sheet. Miss even one and the
model stops looking at your sheet and
starts making the face up, which is
exactly how you end up with a different
guy in panel four. So, let me just
quickly tag all of them and generate and
look at [music] that. Six panels, three
across and two down and every face in
there is the same face of the sheet.
Your whole scene is mapped out and you
haven't made a single second of video
yet, but you've now got two references
sitting in your project and only one of
them is going to give you the better
video. All right, the next step is to
generate our video. I will be using the
latest Students 2.5 model. Now, Students
2.5 can generate your video from either
the character sheet or the storyboard.
You can attach both, but it does
sometimes mix them up. So, pick one and
I'm going to run it both ways so you can
see which one actually comes out better.
Now, go ahead and click into AI video
and set it to reference. Model is
Students 2.5, 16 by 9, 720p, but can go
up to 1080p
and pull the duration all the way up to
30 seconds. Then, paste in prompt three,
the clip prompt, and the same rule
carries over. So, tag your image
everywhere it's required. One thing
you'll notice inside this prompt is that
it blocks speech completely. Ambience
and sound effects only. Zero human voice
for the full 30 seconds. That's
deliberate and I'll come back to it in
the voice step. For this first run, I'm
attaching the storyboard. Let's see what
happens.
>> [snorts]
>> 30 seconds, one generation, one
continuous shot. The camera moves, the
light holds, the faces hold. That's six
clips of work coming out in a single
take. Now I'm running the same prompt
with the character sheet attached.
And here's the difference. Both came out
strong, but the character sheet version
gave me lip movement I never asked for
and the storyboard came back cleaner,
which makes sense because the storyboard
had already told it what happens in
every beat, so there was less left to
guess at. So, the storyboard is what you
animate from and the character sheet
stays as the thing that built it. Go
ahead and click download. That's your
video, but it's still silent. And
silence is where most of these builds
quietly fall apart. So, let's generate
the voiceover. I'm using 11 Labs for
this. So, once you are into the platform
at the left side, click into
text-to-speech and paste your script
into the box. I'm using Roger's voice
for this one. Now, remember the clip
prompt blocking speech? This is why. You
already saw it happen back in the
character sheet version. The mouth moved
on its own. Students can do lip sync.
And it does it well when you actually
ask for it. But, one mouth landing
slightly off is the single thing anyone
remembers out of a 10-minute video. And
the animated history channels that
actually work are narrated anyway. Night
Shift, Bible Project, Yarnhub, all of
them run narration as the spine and let
the visuals carry the acting. Hit
generate and it hands you two takes. I
like the first one, so let's go ahead
and download that. All right. The last
step is to stitch everything together
and it's the easy one. Open CapCut,
click import, and bring in both files,
your video and your voiceover. Drop the
video onto the timeline, drop the audio
on the track underneath it, and line the
two of them up against each other. And
that is done. So, that's every step
stacked into one one piece. Let's let's
play it back.
>> In 1925, Victor Lustig walked into a
Paris hotel and told a room full of
scrap metal dealers something
unbelievable.
The government was secretly selling the
Eiffel Tower.
He convinced them he was an official,
convinced one dealer to pay him, and
walked away with the money.
Then, Lustig did something even crazier.
He came back and ran the exact same scam
again.
>> So, that's the whole thing, story,
style, character sheet, storyboard,
animation, voice, assembly. And the
reason it's seven steps and not one
prompt is that these problems don't all
come from the same place. I spent a long
time trying to fix all of it inside the
prompt, and it just doesn't work that
way. So, everything you just watched was
built inside Design AI, who sponsored
this video. Sedans 2.5 sits in there
alongside everything else I used. So,
the link is in the description if you
want to build this one alongside me. And
if you want the style library, every
style I've tested with all the prompts,
comment the word styles and I'll send it
over. The thing I do next, if I were
you, is pick your story before you pick
anything else. It's the least exciting
step, and it's the one that decides
everything after it. I'll see you in the
next one.
Planet AI
Latest World Leaders as Rockstars | Ai Generated
▶ Watch on YouTube
Could not fetch transcript (it might be disabled or unavailable). Error: 
Could not retrieve a transcript for the video https://www.youtube.com/watch?v=9Ky5QTpVjAw! This is most likely caused by:

Subtitles are disabled for this video

If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
AI with TechZnap
Latest Suno Just Changed The Rules — Is Mozart AI The Answer?
▶ Watch on YouTube
If you use Suno, you got an email this
week. And if you skimmed it, I'd go back
and read it properly because that
completely [music] changes what you're
actually allowed to do with your Suno
songs.
Starting September 3rd, Suno is putting
a monthly limit on how many songs you
can download. Here's how it breaks down.
On the free plan of Suno, you get seven
downloads. No, not seven a month, seven
total for lifetime. And they're for
personal [music] use only. On Pro, it's
20 a month. On Premiere, 60 a month. And
if you're working inside Suno Studio,
those don't count against you. And no,
unused downloads don't roll over to the
month. [music] Yeah, that sounds pretty
strict. And here's the part I'd
underline. Under the new terms,
>> [music]
>> your commercial rights are tied to those
downloads. You can only use a song
commercially if you've downloaded it
properly under your plan's allowance.
Now, if you're someone who's making
music just for fun, you might not need
to touch that ceiling. But if you're
making music for scoring your YouTube
videos, taking client work, or building
a catalog, 20 finished tracks per month
can be pretty tight. There's more in the
update. Suno's now watermarking and
fingerprinting [music] generated audio,
so tracks can be identified once they
leave the platform.
>> [music]
>> There's automatic scanning of uploads
and lyrics for copyright, and the terms
of service got rewritten around
commercial rights and how disputes get
handled.
Suno's reasoning for all this is that
people were mass exporting AI tracks
onto streaming platforms at scale.
That's a real problem, and I'm not going
to pretend it isn't. But this fix landed
on paying customers, and the reaction
from creators has been rough.
If that's where you're at, and you're
looking around at what else is out
there, I've been using something else
for a few months now, and I want to walk
[music] you through it. I'm not saying
it's better than Suno, but this could
genuinely be one of the best
alternatives that navigates around the
current AI music landscape. It's called
Mozart AI, and it's built on a different
idea. And right now, that difference
matters. Mozart is built for musicians,
by musicians. It started from one
number, 85% of music never gets
finished. They wanted those songs
finished. So, they didn't build a song
generator. They built a workstation that
takes the tedious [music] parts and
leaves the creative decisions with you.
No training on copyrighted music,
either. Only licensed data sets. Because
the last thing a producer wants is a
one-click [music]
song.
Enough talk. Let's get on with it. This
is the main create screen. They call it
vibe sessions. You describe what you
want, and it generates. You can type in
your lyrics, or use the AI assistant to
generate some lyrics right here, as
well. There's an instrumental toggle if
you don't want vocals at all. Your
library sits [music] on the right.
Everything you've made, ready to pull
back up. Same idea you already know.
Their latest music model is called Hans
2.0, and it's [music] their best one
yet.
Some people have claimed it to be even
superior to Suno's V5.5 model. And in my
previous [music]
video, I actually tested it against Suno
on identical prompts, and it
surprisingly held its own. Let's hear
something. I'll open the lyrics
co-writer with [music] my idea, generate
the lyrics, and hit create.
>> So, go on without [music] me.
Count stars for two.
Our whole [music and singing] is raining
till the ocean's through.
Don't you look back.
Don't [music] you say my name.
Just remember Tuesday.
>> Yeah, that is as good as it gets. The
vocals are very realistic. The lyrics
are exactly built on the mood [music]
that we asked for and it delivers.
Now, here's where it gets more
interesting. This is personas.
Similar to Suno's voice feature, a
persona is a saved sound identity. Think
of it as your voice and style captured
so you can reuse it across every song
you make.
You can give it the source audio, upload
a file, pull something from your
library, or just record directly in the
browser.
>> [music]
>> Once it's built in vibe sessions, you
can select that persona for your next
song. There's a strength slider so you
control how much it colors the track
[music] and you can publish a persona
for other people to use which earns you
credits. There's a whole gallery of them
if you want to browse what others have
made. That's the difference between
generating a song and building a sound
that's actually yours across everything
[music] you release.
Next, we have layers. This is something
that Suno doesn't directly have as a
separate feature. In this workflow,
Mozart AI lets you use text prompts to
generate musical loops and stems that
automatically complement your track's
existing key, BPM, and rhythm. Let's
create a reggae groove.
>> [music]
[music]
>> Both variants sound great. This is a
very useful feature, especially [music]
for generating ideas for your beats.
Next, we have the big one, studio. This
is a proper multi-track timeline, tempo,
time signature, metronome, key, record
and play controls.
Separate tracks stacked up and you can
see the waveforms. Every [music] track
has its own track agent. You click into
a track, type what you want in plain
English, and it generates onto that
track alone. Nothing else moves.
There's a synth in here, too, with real
controls, envelope, filter, [music]
reverb, delay, and presets for pads,
leads, plucks, bass. There's a piano
roll for MIDI. So, if you want to get in
there and move notes around by hand, you
can.
There's an agent side to it, as well.
Once a [music] song exists, click three
dots and go to edit with agent, and you
can just talk to it. "Brighten the
chorus. Add a saxophone solo. Extend
this with a 30-second guitar section."
It tells you what it did and lets you
review the changes before you keep them.
Mastering is the same. Ask it to master
for streaming, and it handles loudness
and clarity targeting specifically
[music]
for streaming platforms. And what's
more, it also makes music videos. It'll
either generate a character from the
song or take an image you upload, build
a storyboard, and then build a cinematic
sequence to match. Here's one it made
for our initial country song on the
Titanic sinking.
>> Count the stars for two.
I'll hold this railing
till the ocean's through.
>> [music]
>> That's not exactly on par yet with those
AI video giants, but it does a good job
at bringing life to your idea instead of
a static image. It has understood the
context of the scene well, and the
imagery is honestly beautiful. However,
the lip sync needs to be improved, and
it would be nice to have more dynamic.
And with that, we come to downloads.
On Mozart's paid plans, downloads aren't
rationed monthly. There's no counter
sitting in the corner telling you you've
got four left until the 3rd.
Commercial rights come with creator,
pro, and premium plans. And here's the
bit I like. Once a song is generated
while you're subscribed, [music] those
rights are yours permanently. You keep
them even if you downgrade later or
cancel completely. [music]
They don't get taken back. I'll be
straight with you though. Generating
still costs [music] credits, so it's not
unlimited in every sense of the word,
but there's no monthly ceiling on how
many of your finished songs you're
allowed to earn from. Right now, that's
the difference that actually matters.
Look, I'm not telling you to delete your
Suno account.
Suno is still excellent at what it does,
and if 20 downloads a month covers how
you work, you're fine.
But if that number is a problem for you,
if you're doing client work or scoring
or building out a catalog, it's [music]
worth having a second option open before
September 3rd rather than after. And
Mozart AI is a genuinely strong
contender.
Go make a few tracks and judge it
yourself. That's the only way you'll
know.
Also, there's a link in the description
that gets you 15% off the annual plan if
you want to try it properly. And tell me
in the comments what you want next.
[music] A full deep dive on studio, a
different alternative, or a video on
getting the most out of Suno's new
limits. That's it for me. See you in the
next one.
Peter Yang
Latest Grok Bot: 5 Must-Try Use Cases for Work and Life (Full Tutorial)
▶ Watch on YouTube
Hey everyone, I think Grockbot is the
future of personal agents and I'm going
to show you how to build five of my
favorite bots. In this video, we're
gonna build an advisor to create and
manage your other bots. A YouTube
researcher to find out videos in your
interest. An ex Scout to surface the
most insightful and funny tweets from
your timeline. A digital Marie Condo to
clean up your email and find paid
subscriptions to cancel. A personal
concurge to help you get good prices on
travel and vacations. And as a bonus,
we're going to build a gamer to see if
Grockbot can play classic games like
Doom and Red Alert. We'll finish with a
discussion about privacy and whether
Grockbot should be your new AI daily
driver. All right, but first, let me
show you what makes Grockbot different.
A few weeks ago, I tweeted that chat GBT
was one feature away from building the
AI product that I really want, which is
a dedicated personal computer in the
cloud for running AI agents. Lee from
the cursor team DM'd me shortly after to
test Grockbot.
What exactly is a dedicated cloud
computer? You know the meme of people
walking around with their laptop half
open to keep their agents running? You
don't have to do that anymore if you're
using Grockbot. Basically, Grockbot
lives in a computer sitting somewhere in
SpaceX AI servers instead of in your
home. It has its own browser and
operating system that you can use and
logging to your favorite apps. Now,
let's do a quick comparison of Hermes,
Chat GBT, and Grockbot. Hermes I run on
my Mac Mini at home. It's on 24/7 and it
gives it a persistent computer, but you
have to go buy the machine and set
everything up for yourself. You can
technically also run Hermes on a virtual
private server, but that again requires
more setup work. Now, chat GPT work uses
plugins and a cloud browser, but the
problem is that browser at is right now
cannot stay signed in to your favorite
apps. The UX for chat GBD work also
feels scattered across chat work and
codecs. Let's take a quick look here.
It's just kind of like a mess of chat
threads here. There's chat GPT codeex.
There's like another tab up here for
work and chat. And it's just all really
confusing even to someone like me. In
contrast, Scrotbot gives you a
persistent cloud computer right out of
the box. And I think it's the easiest of
the three to set up and get going. and
his interface, as you can see here, also
feel a lot more focused and delightful.
I love the little bots and the
animations that they make when they
start working. Right, let's say look up
the weather. And you can see here that
there's some pretty cool subtle
animations. Each bot has a different
personality and style. And this is the
work of Kurser's great design team,
including Jenny Wen and other folks on
her team that I've interviewed in the
past. Overall, I think this UX feels
more like talking to a coworker than
getting lost in a 100 chat threads. So,
can Grockbot actually replace Chat GBT
and Codeex, which is my current daily
driver? Let's run through some of my
favorite bots right now. So, the first
bot that you should create after you
install Grockbot is an advisor. Here I
have my advisor ped to the very top of
my Grockbot list. With your adviser, you
want to start by telling it about your
work and life and then actually ask it
to give you some ideas for boss to
create. Here's what I told my adviser.
I'm a creator and founder of Behind the
Craft. I have two daughters and I live
in the Bay Area. I publish newsletter
posts and YouTube videos focus on
practical AI tutorials and podcast
interviews. And I also have a membership
portal over here. So based on all this,
suggest five bots that can help me save
me time or money. And for each explain
what it would do in a few sentences. As
you can see here, it created five bot
ideas. A daily AI news scout, YouTube
comment miner, membership concurge,
social media poster, and podcast
prepper. Right now, what many people
don't realize is that you can actually
get one bot, in this case, our advisor
bot, to help us create the rest of our
bots. Here I asked my adviser to build a
single bot that both researches other
channels and mine's comments and let's
call it YouTube researcher. Right? So
now let's go ahead and check out the
YouTube researcher use case. You can see
that the advisor kicked off the YouTube
researcher bot with this instruction to
create a morning brief to send me
YouTube intel every morning. But I want
to give it more specific instructions.
Right? First, I asked it to monitor
specific channels in my niche. And let's
scroll down here. And it's starting to
monitor channels. And here is its first
attempt at a brief. You can see here
that the brief is like just like very
verbose and hard to follow. It's like a
lot of stuff here. It did end up pulling
my YouTube comments and finding themes
there. So, that's great. But, this is
just like too much to process, right?
So, I gave it some more instructions. I
told it to give me a report in the
specific format which is first list your
top three content ideas. Then give me
the top five outliers for other channels
and then give me the top performing
overall. Limit your research to the last
14 days because YouTube with AI stuff is
very topical and also add the top common
themes back to the report. And by the
way, this is how you should be working
with AI to refine its output instead of
trying to oneshot something. Now you see
here that the report is much more
concise and easier to follow. Let's take
a quick look. Top three content ideas.
It wants me to make a video on my spec
scale that I plan to do soon. It tells
me to make a video on Grockbot that
we're making right now. And here are the
top outliers from the other channels.
Outlier meaning that it beat that
channel's performance for the last 14
days. Yep, pretty interesting topics
here that we can follow up on. and top
performing overall and also the top
common themes and it did more here and
then I told you to send a full report
again just to make sure it looks good
and now what I can do is basically
there's a daily job that sends me this
report every morning right just this
morning for example it sent me a new
report with the top angles the watch
list and comments and I basically have a
YouTube researcher that proactively
finds interesting topics for me to make
videos on and gives me feedback back
based on my user's comments. Okay, so
this is the process that we're going to
follow for pretty much the rest of the
boss that we're going to make, which is
we're going to kick it off with an
initial prompt. We're going to iterate
back and forth with it to make his
output good, and then we're going to
schedule a routine or job that runs
daily, weekly, or monthly to have it
proactively do work for us. Now, let's
actually see how we can follow the same
process with our X Scoutbot.
So, Grockbot is part of Space X AI,
right? So, it should have firstparty
access to X and Twitter data. And even
though I have an unhealthy addiction to
X, I still miss plenty of great content
or bookmark tweets that I never look at
again. Here's the prompt that I gave my
exbot. Find the most viral ex posts from
the last seven days in my niche and
create a weekly report with the top 10
tweets from people I follow, engaged
with, or bookmarked grouped into a few
categories. Include a full tweet copy
link and your analysis and end with
three content ideas for me to tweet
about. Next, you can see here that it
found my ex account and it automatically
created a routine to do this every
morning. And here is the initial output,
right?
So it found themes around playbooks that
people save not just like from Greg. 23
ways I use AI agents to grow my startup.
Chinese developer from loop to graph
engineering and more tweets from Greg.
Another theme that I found is just about
Grockbot codeex and chat GPT work
compared and actually used right a tweet
from Riley and from Ben who works on the
cursor team on the most loved internal
Grockbot use cases and more. And the
third theme is around creative workflows
that I actually am interested in tasing.
And they ended this report with the
three things to tweet about next. Some
ideas here. And just for fun, I asked
you to also give me the top five
funniest tweets from my timeline. I
think digging through it. And the number
one funniest tweet is as usual, OpenAI
and Anthropic kind of sniping at each
other and kind of joking,
right? And then there's some more funny
tweets down here. all kind of in my AI
knee niche. You can see here that it set
up a routine automatically and it sent
me a new report this morning. Now, let's
actually tell it to do something else.
I'm going to turn on voice here. Can you
send me this report to
peterbehindthecraft.com
and also make sure you include a full
tweet copy in your report. Send to my
email right now. All right. So, let's
see if Ashi is able to send this report
over email because I don't want to have
to open Grockbot each time to see it. I
want it just in my inbox every morning.
Awesome. Here's a report from Grothbot.
You can see here that it has this
themes. It's included the full tweet
copy. You can scroll down here and yeah,
I I think this looks pretty good, right?
And then it has down here, I'm sure, the
top three things to tweet about next.
Again, the process here is to give
Grockbot initial prompt, have it pull up
the information, iterate with it to get
the output and the report right, and
then either tell it to send you a daily
job through Grockbot itself or through
your email. In this case, I prefer email
because I like to wake up with the top
tweets to consume directly in my inbox
instead of having to open a separate
app. And by the way, it's able to send
me an email because I connected to a
bunch of plugins, including the Gmail
plug-in. Now, let's move on to a fun bot
that I call Marie Condo that cleans up
your digital clutter. There are three
places where clutter piles up, right?
Your email, Google Drive, and your pay
subscriptions. Now, I've connected both
Gmail and Drive in plugins to Grubbot.
And then I gave it this prompt. Audit my
Gmail, Google Drive, and recurring email
receipts and create a cleanup plan. Find
newsletters I already open. Find large
or abandoned drive files. Try to
identify pay subscriptions from email
receipts. And also you can hook up
Mercury MCP to get this data. Mercury is
the bank that I use for my business. And
then group everything into these
different categories. Use a number list.
And crucially, I told it to not move,
delete, or unsubscribe or cancel
anything without my approval. Right? The
last line is really important for a bot
that cleans up your files. You always
want to review what it plans to delete
before letting it remove anything. And
you can see here that first Grockbot
confirmed that it's connected to Google
Drive and Gmail. And then it actually
connected to Mercury MCP. I had to sign
in on the remote cloud computer to make
this work. And then it gave me this
initial report. This is a massive list
of stuff to clean up in my email
subscriptions to review and other things
that it should not touch. Right?
Honestly, this list is like pretty
overwhelming. It's incredibly long. And
I gave it some feedback of like, hey,
list it properly, categorize it
properly. It's still an incredibly long
list. Right? Again, this stuff is not
going to get it right in one shot. And
eventually I told it to just show me max
10 items in each list and get rid of all
the random labels that it has. So now
this is much more digestible. It has a
bunch of emails here, Google Drive and
pay subscriptions. And I gave it one
final piece of feedback to only show me
emails to unsubscribe to, Google Drive
files to delete and pay subscriptions
that I want to cancel. And then here's
the update list that I came up with. Now
it has this list as a number list and my
feedback is to actually get it to take
action which is email unsubscribe to
three and four 5 6 7 and 8 Google Drive
delete the extra tax return delete some
of these large files and pay
subscriptions cancel 11 and 15 right and
just as a general tip it's always good
to ask Grothbot or AI to give you its
response in a number list like this to
make it super easy for you to just tell
it to do things by just referring to the
number in the list instead of have to
type everything over again. I use this
pattern all the time. All right, so now
Crockbot is actually taking action,
right? This is where the magic happens.
You can see here is already unsubscribed
to a bunch of senders. It's trashed
three Google Drive files that we asked
it to and now it's cancelling Lovable
and Equipped Foods, which is a protein
company. Now to cancel lovable I first
have to sign to the cloud computer with
my lovable credentials which I did
manually and we found out that actually
lovable is already scheduled to be
cancelled. So lovable is done. Then it
asked me to sign to equip foods and I
did that and then it cancelled uh the
protein subscription for me. By the way,
Equip Foods is a great protein company.
I'm only canceling because I have too
many protein powders at home already.
Uh so it basically did all this work,
right? trashed a bunch of stuff. It's
unsubscribed to a bunch of emails and
it's canceled a bunch of subscriptions
to save me money. And you can see that
it did all this in around five minutes
when it would have taken me probably 30
minutes to an hour to do all this
manually. Marie Condo is very very
useful. Of course, and there's something
missing with this bot. I think we should
ask it to can you talk like Marie Condo
from now on? Give me an example because
right now it sounds too much like a
robot and not so much like Marie Condo
and let's see what it comes up with. All
right. So now uh Marie Condo is actually
going to talk like Marie Condo. You can
see Hermis and the singing have
completed their work. We thank them and
place them in the trash where they may
rest. Equip foods no longer sparks joy
and release prime protein subscription
with gratitude. [laughter]
Yeah. So now we can set up Marie Condo
to maybe talk to us every week or every
month to clean up our digital files and
spark joy, right? So this is definitely
a a bot that I recommend you setting up.
Just remember to ask it to give it
output in a number list instead of just
doing the cleanup for you so that you
can review it first, right? You don't
want it to accidentally delete some
important file. Now let's cover the
personal concurge bot. This is a bot
that I want to use for all my vacation
and travel plans. It has access to my
vacation document where I've listed my
December trip to Japan. Here's a quick
preview of the document. It has the full
iterity for Japan that I created with AI
as well. Let's go back to Grockbot. And
what I want Grockbot to do is I haven't
quite booked my flights to Japan yet.
So, I wanted to monitor for price
changes and alert me on what routes are
the best deal. Here's a prompt that I
give it. Read my vacation document and
monitor the exact flight legs for my
family trip. Get the dates, the best
options for each flight leg. Check
regularly and let me know when the price
improves. You see here that I found my
flight legs and is using Google flights
to check the prices. Now, you may be
wondering what's the advantage of using
Grockbot for this instead of just
setting up Google flight price alerts.
And the value here is that Grockbot can
understand my whole trip based on my
document and decide what's a better
option for my family. If you scroll down
here, the big finding is that a Tokyo
round trip is about $2,700 cheaper than
the open jaw flights that my dog
prefers. Right? My dog prefers it to fly
from SFO to Tokyo and then from Tokyo to
Fukoka, which we're going to, and then
from Fukuoka back to SFO. But actually,
Grockbot found a better itinerary. It's
about $2,700 cheaper to just fly round
trip to Tokyo instead of doing this
itinerary that I laid out, right? A
simple Google price alert would not have
found this. And now I ask it to check
every morning at 9:00 a.m. to see if the
price improves. And you can see here
that the price is still around the same.
The Tokyo round trip is still much
cheaper. Eventually, I can ask Rockbot
to also just go ahead and book the
flight for me or check into the flight
when the time comes.
And generally speaking, I think it's
always a good idea to have a travel
thread or travel bot to both help you
plan travel ahead of time and also when
you're at a location to help you book
amusement parks, to help you figure out
what to do every single day. So, I
imagine I'm going to be using this
personal travel concurge a lot. All
right, let's test one more use case with
Grockbot. Because Grockbot gives me a
dedicated cloud computer, I thought it
would be fun to ask it to install and
let me play some retro games. I asked it
to install Red Alert, Doom, and
Commander King. And you can see here
that it found the files and installed
it. So, why don't we just ask it now?
Let's say open Doom for us to play. And
let's see how things work out. All
right, looks like Doom is up. Let's open
it in our virtual cloud computer. And
here it is. And here we can start a new
game. Let's pick this episode.
And yeah, let's say Hurt Me Plant Penty.
Here's Doom. And here's kind of where
Grockbot falls apart a little bit.
Unfortunately, because it's on a virtual
cloud computer, the mouse isn't quite
configured right to actually play Doom.
For some reason, it's looking at the
floor all the time, and I can't seem to
adjust the mouse to actually look up.
All right. So, Doom doesn't quite work
here right now.
Now, let's actually try playing some
other game. Let's try asking to play
Commander King. I'm not sure if you guys
know this, but Commander King is an
awesome platform game that I played in
my youth. And now it's loading Commander
King. So, let's see what it comes up
with. Awesome. This is Commander King.
Let's see if it plays well or not. Let's
start a new game. One player normal
difficulty. And here we go. Here we go.
It's a very basic platformer, but I
remember really enjoy it in my youth.
How do I jump? I forgot how to jump. Oh,
so control is jump. And yeah, it plays
better than Doom for sure, but there is
some lag in the keyboard and mouse that
makes it difficult to actually control
the character. Okay, so um yeah, that's
just for fun. But basically, Grockbot is
not replacing your gaming PC or GeForce
now yet. But the fact that the agent can
install and launch these games on its
own computer gives you a sense of how
open-ended this could become and how
much potential there is here. Right? So
maybe Space X AI can use all the data
centers and GPUs in space to actually
deliver AAA games through this virtual
cloud computer. That would be a dream,
right? And while we're at it, I open
this file manager here. And here's all
the stuff that we installed into our
virtual cloud computer. It kind of looks
like Windows 3.1 a little bit. I'm not
sure what kind of thing it is. There's
Japan flights. There's other stuff here,
right? And then there's games that we've
installed. And also we have Chrome and a
terminal. So that's kind of our virtual
cloud computer. All right. Now,
All right. Now before we wrap, I want to
talk about probably the biggest hurdle
for Grobbox adoption, which is trust.
When I see a Google sign screen like
this on my laptop, I don't really think
twice before signing in. But because
this appeared on the virtual cloud
computer, I hesitated a bit because how
do I know that nobody else is seeing
this screen on the virtual cloud
computer? Now, if you go to the Grogbot
website, if you scroll all the way down
here, there is a note about privacy
here, which is how does Grockbot handle
my privacy? Grabbot uses the same cursor
SSO off and privacy mode you already
trust. Your cloud computers encrypted in
transit and at rest and there's no AI
training on top of it. Right? So, you
know, I'm willing to give Grockbot and
the virtual cloud computer access to all
this stuff because I'm an early AI
adopter, but I can see normal people
struggling to understand what this cloud
computer thing even is and kind of
hesitate to sign into their favorite
apps on this device. So, I think the AI
Asian platform that figures out trust
will be the first to get mass adoption.
But overall, I think Grockbot is a clear
sign of the future. We're moving away
from manually using our keyboard and
mouse to do work on our laptops to using
our voice to orchestrate a bunch of
agents that live in a dedicated cloud
computer. And Grockbot is the first
product to actually enable this. Now,
it's not quite my daily driver yet
because I think Chat GPT still offers
more for $20 a month while Grockbot
requires paying $200 a month to use on a
regular basis. But I think Grockbot has
much cleaner UI than chat right now and
is also very very capable. And overall,
it's just great to be in a world where
cursor and SpaceX AI are just as viable
a competitor as OpenAI and Anthropic. I
think cursor may even have the edge if
it can continue to support multiple
models from all providers.
So, Grockbot is available for free and I
definitely recommend downloading and
trying it to see a glimpse of the
future. I'll include some of the prompts
from this video in the pin comment
below. And I also have an exclusive
interview with the Cursor team on how
they built Grockbot coming up in the
next few weeks. Overall, I'm really
impressed by Grockbot. I think the team
really cooked here and I can't wait to
hear the story behind how they built
this. So, please like and subscribe if
you enjoy this video and I'll see you
next time.
The AI Grid
Latest The Most Advanced AI Email Agents Are Now Here
▶ Watch on YouTube
Most people using AI in the slowest way
possible. They open Chat Gypt, they copy
an email, they paste it in, they ask for
a reply, then they copy the reply back
into Gmail, and somehow after all of
that, they call it an automation. But
the real future of AI is not you
babysitting a chatbot. It is AI sitting
inside the tools you already use,
quietly doing the admin work before you
even ask. And that is exactly what
today's sponsor, Fixer, is trying to do.
Fixer is an AI assistant for your inbox.
It works with Gmail and Outlook,
organizes your emails, draft replies in
your voice, and can even take meeting
notes for you. So, in this video, I'm
going to show you what Fixer actually
does, how the setup actually works, how
it fits into a real workday, and why
tools like this are probably where the
AI productivity is going next. Because
the goal is not to have more tools, the
goal is to have less admin. And because
Fixer is sponsoring today's video,
they're actually giving me an exclusive
link in which you can get 25% off your
first month of subscription and a 7-day
free trial so you can try out the
software risk-free. So, if you run a
business, freelance, manage clients,
work in sales, or just have a job where
people constantly email you, your inbox
eventually becomes a second job. You
wake up, open Gmail, and immediately get
hit with a wall of unread messages. Some
need a reply. Some are just FYI. Some
are newsletters you forgot you signed up
for in 2021. And some are actually
important, but they are buried under 40
random notifications. The problem is not
just email. The problem is the tiny
decisions emails force you to make all
day. Should I reply to this now? Is this
urgent? Do I need to schedule something?
What did we agree in this meeting? That
is where Fixer comes in. Fixer is
basically an AI executive assistant that
lives inside your existing email
workflow and helps surface important
messages and tasks. Not another
dashboard, not another tab you need to
remember, not another AI tool where you
have to engineer the perfect prompt. It
plugs it into Gmail or Outlook and
starts helping with the repetitive part
of communication. And according to
Fixer, it is already managing over
350,000
inboxes for mostly client-f facing
professionals, people who live in their
email all day. So, let's start at the
beginning. The setup. Setup. Connecting
your email. This part is genuinely
simple, which matters because most
productivity tools die at the setup
screen. You go to Fixer's site and you
click start with Gmail or start with
Outlook. And here's where you just grant
permission. And once you do that, that's
basically it. You just allow Fixer
access. And Fixer describes this as
one-click setup with no configuration
needed. No rules to build, no folders to
design, no settings menu you have to
figure out before it does anything. Now,
once you're connected, Fixes starts
learning how you write. It looks at your
past emails. It picks up your tone, your
phrasing, and how long your replies
usually are. That is what powers the
reply drafting later. So, the more email
history it can see, the more the draft
sounds like you. One practical tip is to
give it a little time after connecting.
The first hour or so, it is getting
oriented, and after that, the features
start kicking in. There's a 7-day free
trial, so you can test it out on your
real inbox and back out if it's not for
you. Feature one, automatic inbox
organization. The first thing that
you'll notice is that your inbox looks
different. Instead of treating every
email like it deserves the same
attention, Fixers sorts incoming mail
into categories automatically. Emails
that need a reply, emails that are just
an FYI, marketing and newsletter style
messages, notifications. And this
matters because most inboxes fail for
one simple reason. Everything looks
equally important. A client email, a
calendar update, a receipt, a
newsletter, and a random automated
notification all sit in the same pal. So
your brain has to sort everything
manually. That's not deep work. That is
digital housekeeping. With Fixer, you
open your inbox and immediately see what
actually needs your attention, and
you're not locked into its default
either. You can adjust the categories,
turn ones that you want and turn off the
ones that you don't. You can also set
email rules. So if you want messages
from your boss or a key client to always
land in your torespond list, you can
tell fix it once and it handles it from
then on. So just come on over to
categorization custom rules and then in
this box I can essentially say any email
that comes in from contactigrid.com
I can put this in FYI to respond any
section that I wish and this is one of
the biggest shifts with AI productivity.
The most useful AI tools are not always
the ones that generate the flashiest
content. Sometimes the useful ones are
tools that remove 100 small decisions
from your day. You're no longer digging,
you're deciding. Feature two, draft
replies in your voice. The second big
feature, and this is the one that most
people come for, is reply drafting. For
emails that need a response, Fixer
writes a draft and has it waiting for
you. You see, most people use AI to
write their emails, but they do it
manually. They copy the email to the
chatbot, then they write, write a
professional response. The AI gives them
something that sounds like a LinkedIn
comment from a corporate bot, and nobody
wants that. What Fix It is doing is
different. Because it studied your past
emails, the drafts are meant to match
your tone and style. So, if you're short
and direct, the replies are short and
direct. If you're more polished and
formal [music] with clients, it should
adapt to that. It also uses the context.
If you've had prior conversations with
that person, it pulls that in. So, the
reply actually makes sense instead of
being polite but empty. So, here's the
workflow in practice. You open the
email, the draft is immediately already
there, and you do one of three things.
You either send it if it's good, you
mark a quick edit if it's a one line
off, or you delete it and write your own
if the AI missed its mark. And that's
the key point here. AI doesn't send
emails for you. It drafts them. You
review, edit, and approve everything
before it goes out. And that matters
because for most people, the scary part
of AI email tools is not whether they
can write, it's whether they might send
something stupid without permission.
Instead of starting every reply from a
blank page, you're starting from a draft
that is most of the way there. That
might only save a few minutes per email,
but if you reply to 20 emails a day, it
compounds very quickly. Feature three,
meeting notes and follow-ups. The next
part is meetings. Because email is only
one half of the admin problem. The other
half is what happens after calls. You
join a meeting, you talk for 30 minutes,
someone says, "Great, let's follow up on
that." And then nobody writes the
follow-up properly. Or someone writes
notes in a random document, or the
action items disappear into the void.
Fixer has a noteaker that can join your
calls, transcribe the conversation, and
create summaries with key decisions and
action item. It supports Google Meet and
Microsoft Teams, which makes sense
because those are the meeting tools a
lot of Gmail and Outlook users already
live in. And this is useful for a very
specific reason. Most people do not need
more meeting recordings. They need the
useful part of the meeting extracted.
What was decided? Who needs to do what?
What should the follow-up email say? And
because everything is transcribed, it's
searchable. You can find exactly what
someone said by searching the text
instead of scrubbing through recording.
But here's the part that ties it all
together. Because Fixer also lives in
your inbox, it can draft the follow-up
email based on the meeting. The great
talking to you. Here's what we agreed.
Here are the next steps. Email. Meeting
happens. Notes are created. Action items
are pulled out. Follow-up draft is
ready. And again, you review it before
anything goes out. That is the
difference between AI as a novelty and
AI as infrastructure. One gives you a
call output. The other removes a step
from your actual workflow. Feature four
is fixer chat and scheduling. Now if you
go beyond the basic plan, fixer has more
advanced features. On the professional
plan, you get fixer chat which gives you
instance answers from your inbox and
meeting notes. And this is the part that
becomes really powerful if you deal with
a lot of context. Because your inbox is
basically a private chat knowledge
database, client history, project
updates, meeting decisions, feedback,
but normally all of that is trapped
inside search. You have to remember the
right keyword, who sent it, and roughly
when it happened. And if you don't
remember those things, good luck. With a
chat layer over communication history,
you can just ask a person, "What did we
do with this client? Which emails still
need a reply? What are my important
emails?" The professional plan also
helps scheduling across teams and time
zones. So, instead of endless just
Tuesday work, no, how about Thursday
back and forth? It helps coordinate the
time for you. You can also upload files
to train it further. And there's a
HubSpot integration if you run client
work through a CRM. This is where AI
assistants clearly are heading. Not just
answering general questions, sitting on
top of your work, and helping you
operate faster. That is a much more
useful category than write me a poem
about productivity. At least unless your
inbox is somehow full of poems, in which
case you have a different problem.
Security and control. Now, because this
is email, we have to talk about security
and control. You're not connecting an AI
tool to a random notes app. You're
connecting it to your inbox. So, this
part matters. Fixer is Google verified
and Microsoft verified. And you must
understand that Fixer has passed the
security reviews required for those
platforms. It is also SOCK 2 type2
certified, ISO/IEC27001
certified, GDPR compliant, and that data
is encrypted at every stage and that
your email data is not used to train
external AI models. But the practical
point for viewers is this. You stay in
control. Fixer drafts emails, it doesn't
send them for you. You can review the
drafts, you can correct the categories,
and if you delete your Fixer account,
Fixer will revert your inbox to how it
was before. And that doesn't mean you
should ignore privacy. You should always
read the permissions before connecting
any AI tool to your email. But for a
tool in this category, these are the
kind of controls you want to see. So,
who is Fixer for? I wouldn't say this is
for someone who gets three emails a day.
If your inbox is already clean, you
probably don't need an AI assistant
living. This is for people whose inbox
has become a part of their job.
Consultants, recruiters, sales teams,
creators working with sponsors,
founders, agency owners. Basically,
anyone who spends a meaningful part of
their day sorting messages, writing
replies, scheduling meetings, and trying
to remember what happened on calls. If
your email is just communication, Fixer
might be useful. But if your email is
your operating system, this is where it
makes the most sense because a product
is not really selling AI email. It's
selling a cleaner workday, fewer open
loops, fewer blank replies, fewer
forgotten follow-ups. Here's a quick
breakdown of what it costs because that
matters. Pixer has a 7-day free trial on
every plan, so you can test it before
paying. The starter plan is listed at
$22 per user per month when build
annually or $30 per user per month
monthly. that gives you your
organization for one inbox and calendar,
draft replies in your invoice and
meeting notes. That's the core
experience and for most individuals that
is enough. The professional plan is
listed at $37.50
per month when build annually or $50 per
user per month monthly. That adds
multiple inboxes and calendars,
scheduling across teams, time zones,
fixer chat, the HubSpot integration,
uploaded files to train fixer, and
onboarding with a specialist. If you're
juggling multiple inboxes or running
client work through a CRM, this is the
one. And then there's an enterprise plan
for larger teams with bespoke pricing
and features like SSO, dedicated account
management, and custom security tools.
So if you're testing it as an
individual, start with a free trial and
connect one inbox. Don't over complicate
it. Just use it for a few days and ask
one question. Did this reduce the amount
of time I spent in email? Because that
is the only metric that matters here. So
that is fixer, an AI assistant that
organizes your inbox, drafts replies in
your voice, and takes meeting notes
inside the workflow you already use. And
I think this category is important
because it shows where AI is going next.
Now, if you want to try Fixer, you can
try them for free with a 7-day free
trial and 25% off your first month with
the link in the description.
Thomas Frank
Latest I never dread this workout
▶ Watch on YouTube
Could not fetch transcript (it might be disabled or unavailable). Error: 
Could not retrieve a transcript for the video https://www.youtube.com/watch?v=vhqq4U3mnNQ! This is most likely caused by:

Subtitles are disabled for this video

If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
Kurzgesagt – In a Nutshell
Latest What the Flu Does to Your Body
▶ Watch on YouTube
Could not fetch transcript (it might be disabled or unavailable). Error: 
Could not retrieve a transcript for the video https://www.youtube.com/watch?v=U5-dE2nCzjM! This is most likely caused by:

Subtitles are disabled for this video

If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
Veritasium
Latest Total Solar Eclipse From Space
▶ Watch on YouTube
In just a couple of seconds
Spain is about to get its first
total solar eclipse in over 100 years.
And we're here with NASA
to film it from space.
But we almost didn't do this video
because it turns out
there are a great couple videos
on solar eclipses on YouTube,
and we just thought
we couldn't add anything.
But then we started looking.
[laughs]
Oh my God, that's insane!
And it turns out...
Oh, I can't, I just gotta go watch this.
That's insane.
So, Henry, Emilia and I still had
some unanswered questions like,
why is the Northern Hemisphere
getting more total eclipses
than the southern hemisphere?
Why does the sun suddenly turn white
only during pure totality?
And how come
we still haven't figured
out what causes
these elusive shadow bands
on the ground?
Oh, that's so weird.
Okay, so we are on the
way to Burgos, Spain,
because it's going to be right
in the center of the moon's shadow.
So perfect for totality.
And we're also meeting up with NASA
and a bunch of other science groups
because they're launching a couple
of balloons up to the edge of space.
So they will actually be
filming the eclipse,
and hopefully they can
answer some of our questions.
Vamos!
Oh my God!
It's getting occluded, oh no!
Angela?
- Yes.
Oh, lovely to meet you.
I’m Emilia.
Oh, so nice to meet you.
This is Gregor.
Nice to meet you,
thanks for having us.
So, can I ask you?
Yeah.
Are those total
solar eclipses on your toes?
Yes. Totally geeking it up.
Yeah,
I guess, first question right off the bat.
How many eclipses have you seen?
So this will be
my fourth eclipse
and the third
total eclipse.
There's lots of folks here
who have seen many more than me.
Eight or even as many as 12. That's crazy.
So how rare is it
that Spain is getting two
total solar eclipses in
less than 12 months?
If you picked any random point
on any place on the earth,
it's about every 300 years
that the eclipse will repeat there.
But then there are some places, you know,
where there happens to be occurrences,
because there is kind of a pattern
to the eclipses.
You can see one of those patterns if you map out the paths of all total solar eclipses
between 2000 BCE and 3000 CE.
Notice that the Northern Hemisphere
consistently gets more eclipses,
about 15% more than the
Southern Hemisphere.
Now for there to be a
total solar eclipse
where the moon completely
covers the sun,
you want it to be
bigger than the sun in the sky.
Now, the biggest factor for
that is the moon's elliptical orbit.
Sometimes it's further from the Earth,
sometimes closer.
So over the course of the month,
its apparent area increases by up to 30%.
But the Earth's orbit around
the sun is also elliptical.
So from our point of view,
the sun's area is about 7% smaller in
July than it is in January.
Which means there is a higher chance
of a total solar eclipse around July.
But that's also when it's mostly
the north half of Earth
that is tilted towards the sun,
the Northern Hemisphere summer.
So that's where the total eclipses
tend to fall.
Now the Southern Hemisphere
is actually going to get
more annular eclipses
where the moon is slightly
smaller than the sun
so you have this ring of sunlight
around it.
And that's because during the southern
summer the sun is actually closer.
All of this is actually going to flip,
because the Earth precesses around
its axis and its elliptical orbit
also slowly shifts around the sun.
In about 9500 years,
it will be the Southern Hemisphere
that gets more total solar eclipses.
Now, there is another interesting pattern
that shows up when you look at all
types of solar eclipses,
including annular and partial ones too.
So I crunched some numbers,
like 5000 years of eclipses.
And if you look at how many happened
per year, there's never a year
without some sort of solar eclipse,
which kind of feels mind boggling.
You at least have to get two
and sometimes you can get five.
How do you get five?
- Yeah, yeah I know.
Check this out, so:
The moon's orbit is tilted about
five degrees out of the Earth sun plane.
So most times the moon passes
in front of the sun
It's at the wrong height
to cast a shadow on Earth.
And that's why we don't get
a solar eclipse every month.
But there are these two nodes
along the moon's orbit where it does
actually cross the Earth-Sun plane.
And throughout the year,
both of these nodes are going to land
nicely in between the Earth and the Sun.
This happens about six months apart,
and it opens a window
of around 34 days
where solar eclipses become possible.
And these are called eclipse seasons.
This is huge window to get eclipses.
- Yeah.
So as long as you get a new moon
somewhere inside of this eclipse window,
you get an eclipse.
But it turns out
you have to get at least one.
The reason is
the eclipse season is 34 days,
but a new moon happens every 29.5 days.
Of course.
So you can't go through an eclipse season
without the new moon happening
at least once.
However you roll the dice,
it always has to land
inside the eclipse season at least once.
So you always get two a year.
Great.
Yeah, right?
Now, if the new moon falls
towards the middle of an eclipse season,
it usually produces
a total or an annular eclipse,
and you can only fit one of those types
of eclipses into a season.
So you can get about two per year.
Yeah, this doesn't seem rare,
definitely not once in a lifetime.
Yeah, but it gets worse
because if you also count
partial eclipses,
those happen more towards the edges.
And because they happen to the edges.
You can get two
in a single eclipse season.
Or you can get four a year.
But wait, then how do you get five?
Yeah. The thing is,
the eclipse seasons themselves
also drift around the Earth's orbit.
So if you start on January 1st
and you get an eclipse
really, really soon into the year,
you can hit like a partial
and then another partial.
And then as you're moving,
these eclipse seasons are shifting.
So you get another one and another one
and then you think you're done.
But the eclipse season has actually
already shifted into December.
And then you get a final one.
So you can get up to five.
Woah.
Okay, wait, I realized
I forgot to tell you
my favorite fact out of this, out
of this whole discussion, and that's that
if you look at the eclipse season,
when the eclipse happens,
the solar eclipse here,
and you rewind by about two weeks,
you'll realize that now the Earth is in
between the sun and the moon.
So you get a lunar eclipse.
So you get a lunar eclipse, which means
every time there is a solar eclipse,
it is accompanied by a lunar eclipse
either two weeks before or after.
Oh, that's so cool. Yeah.
So there's one happening in two weeks.
Oh, amazing. Yeah, yeah, it's really cool.
I actually tried
to find an answer to this question,
but if you look at timelapses
taken during the eclipse.
You'll see that the partial parts
are usually like yellowish in color.
And then the totality is white.
If you Google on Google you get like a little animation and it turns white.
Yellow, and then it's white
and then it's yellow again.
I could not find a satisfying scientific
explanation for why that happens.
It's so bright, right?
In order to take those timelapse pictures,
you have to put a filter in front of it.
And so the most common with those filters,
it's that kind of reddish orange color.
So it's just the most common
filter makes it that color
because the sun is white.
But then it's dim enough that we can look
right at it during totality.
And so you see white.
- Yeah
It's just a filter, okay.
It's just the filter.
Maybe I can show you where we're at. Sure.
You can kind of get set up. Hey.
We're a NASA funded lab in Bozeman, Montana,
and this is an extension of what's called
the
Nationwide Eclipse Ballooning Project.
So we studied the 2023 and 2024 eclipse.
So we have two teams
in Reykjavik, Iceland, and we have three
teams here in Spain.
The Iceland teams are launching
essentially weather balloons, radiosondes.
And then we're launching what we call
larger engineering balloons here.
Look, cowboy engineering here.
What are you doing?
We attach this to our basically
the balloon.
Like right below where we're connecting
the balloon to the payload.
And usually we have a little bit
more sophisticated weights,
but we didn't want to just ship
weights to Spain with us.
So we took water bottles
and filled them with rocks.
I love it, I love it.
We have a little bag of like, souvenir
kind of things, like signed NASA stickers
and things like that
that we're going
to send up on the balloon.
This is going up so you can sign it.
- Hey!
You're signing it ‘Eclipse’?
I mean, Eclipse 2026,
I want to show it off
I thought you were going to say
like ‘Henry’ or something.
So, “Henry was here”?
We're not putting you
in the balloon, though, so...
We're gonna set up kind of in the middle
just because of the wind direction
at the moment.
Okay.
So right now the wind's going
the wrong way from the forecast.
Right. Blowing it that way.
It’s just finding a way to be far enough
that we don't hit the church.
How strong does the wind
have to be for you to hit the church?
Like it's pretty far away, no?
It is. We just have to outclimb it
before we get there.
Okay.
It's kind of weird.
I feel like
it was one of those things
where I'd seen pictures of them,
but I didn't realize how much it
just looked like a regular balloon,
but just super sized.
On the count of three,
one, two, three.
Amazing.
That launch went absolutely perfectly,
and right now our cameras are on their way
up to the edge of space.
But about 3000km away
NASA's launching
another eclipse mission, one mounted
on a 70 year old Cold War
era bomber plane.
They're trying to chase the totality
for as long as they can,
but surprisingly,
this is actually already been attempted.
In 1973, the Concorde went Mach 2
and stayed in the totality for 74 minutes.
Here on the ground, the longest
anyone's going to be able to see
the totality is two minutes
and 18 seconds.
And with NASA's jet going at 460mph,
they're going to be able
to stay in the totality for
just around three minutes.
It doesn't seem worth it,
but they're not just trying to stay in it longer.
Compared to a balloon,
which is a bit shaky.
A jet is much more stable, so it lets them
mount much more sensitive equipment
so they can study
a very specific part of the sun,
part of the sun, that once led to the
discovery of a fake element.
On the total eclipse of
the 18th of August, 1868,
French astronomer
Jules Janssen went to India
to look at a part of the sun
that's normally outshone from observation,
the prominences.
Arcs of glowing gas at the very edge.
He put a slit in front of a prism
and broke their light into its constituent wavelengths.
Out came five bright bands.
One of them matched no element
known on Earth.
Two months later, the English astronomer
Norman Lockyer saw the same line.
He called it helium, after Helios,
the Greek god of the sun.
The next year, other scientists
used the same method on the hotter corona
and found another unexplained band.
For 70 years it was thought to be
an element called coronium,
but in 1939 it turned out to be iron,
so hot that 13 of its electrons
had been ripped away.
So an eclipse led to discovering
both a real element and a fake one.
Now, to study the eclipse, Janssen used a slit,
but you actually don't have to.
So here I've cut out a triangle shape
in this cardboard,
and we're going to see the shadow
that it casts.
You can see there's
this projection of light
that makes a triangle, not too surprising,
but watch what happens as I slowly
move it away from the ground.
You'll see that triangle slowly morphs
into a circle.
I can even take this eclipse shape,
which looks like an eclipse
when it's close to the ground, but
I pull it again and it's always a circle.
Gregor, what are you seeing brother?
Wait, let me switch the glasses.
Oh, we're getting, like,
a nice croissant shape.
You know, I think it's called a crescent.
No no no no no no no.
A croissant, it's like very thick.
You check it out. Check it.
Yeah. No.
It's cool.
Here we take those same cardboard
cutouts
Here's the one with the triangle.
And again, you're seeing a triangle.
But look what happens when I move it away.
Look, it's a crescent.
How cool is that? Okay, and
I can do it with another shape.
I'm gonna try with this star shape,
which again, close up.
But if I move away? This star...
Also a crescent.
How sick is that?
You can even do it.
With something like this,
with all these little circular holes.
Circle. That's a circle. But even quicker.
A bunch of 
little crescents.
That's amazing. Right? Come on, come on.
See, when the cardboard is high enough,
you're no longer looking at the hole,
but a projection of the light source
itself.
Light travels in straight lines,
so a ray leaving the top of the sun has to
angle downward to get through the hole.
So it lands at the bottom here.
The top goes to the bottom
and the left flips to the right.
Everything crosses at the hole and
it comes out on the other side reversed,
which means that during an eclipse,
the crescents on the ground
point the opposite
way to the crescents in the sky.
All cameras actually work like this,
flipping your image.
But to get this effect,
you need the point to be rather small,
it's like a pinhole camera.
That's why all the spaces between leaves
and a tree are perfect.
During an eclipse,
they all become crescents.
And there's another crazy effect.
Watch my hand.
The shadows around my fingers
are quite fuzzy, but if I turn my hand
90 degrees, those same shadows
suddenly become much sharper.
As we approach totality,
every object has a sharp direction
and a soft one, 90 degrees apart.
Take one point source of light, on its own
it casts a perfectly sharp shadow,
but add a second point, also perfectly
sharp, but arriving at a different angle.
Now, where the two shadows overlap,
you get full darkness,
but out at the edges,
where only one of them lands,
you get half the light.
If you do it for more and more points,
well, those edges
start to stack up into a gradient,
and that's what makes the blur.
And during an eclipse, you're left with
this sliver shape. In the tall direction,
it stays roughly the same width as normal.
That's why in this direction
you still get blurry shadows.
But change your orientation 90 degrees
and now, it's much more narrow.
So now it's like a point source again.
That's why the shadows in this direction
are perfectly sharp.
So we're a few minutes from totality.
And, what we're going to do
is we're going to hold up this sheet,
and what we're looking for
are these long moving shadows.
They're sometimes called snake shadows
because of how they move.
So you're talking about seconds
before totality
that you'll see these shadow bands occur
so maybe 20s, 10s before totality.
And so you really have to be, you know,
looking for them and aware
that they're going to exist.
Gosh, I'm a bit stressed about us
filming these now.
- Oh,
you'll be fine.
Like they're
either going to be pronounced,
you're going to be able to capture them,
or they're going to be so subtle
that nobody's going to see them anyway.
Is that it?
No, no, I think there is a little
something, like it's faint.
Oh, that's so weird.
See? Did you see it? Like it's
definitely there.
I don't know, I felt like
you were just moving the sheet.
No. Come on.
It was there, like, they are real.
We saw it, we saw it.
It was faint.
There is still some mystery
around shadow bands,
but exactly what causes them,
there is still some discovery there.
And so, it's absolutely an open question.
Right as you get to the final seconds of light,
before the eclipse
and you just get it's a Bailey's beads,
where you just get a single point of light,
or just a few points of light
along the limb.
The leading theory
is that these narrow beams of light
pass through many layers of air,
of varying temperature and densities.
Each of the boundaries
between those layers of air
refract the light, bending it
this way and that.
This is actually the same effect
that makes starlight twinkle.
But if that were the whole story,
as you went up through the atmosphere,
surely you wouldn't
see these shadow bands.
Did you hear?
Because there was an experiment
done in Burgos by an army engineer
in 1905 in the 1905 eclipse,
going up in a hydrogen balloon?
Oh, wow.
And he was actually looking for shadow
bands.
Oh, really?
So they launched these balloons.
They've put these big white sheets
under the balloons, to try and spot it,
And then they weren't
seeing anything, there standing at it,
and then all of a sudden,
someone pointed out
they were just everywhere, like,
on their hands, on the basket.
And they saw them everywhere,
but not on the white sheet.
They just they were just surrounded by it.
But again, someone was already testing it,
trying to test it,
at altitude back in 1905, right here.
That's amazing.
But what's going on there?
Is there another explanation?
So there are, the other thing
that is coming into effect
there, was the actual
topology of the moon itself.
So those different point sources, as they
line up and interfere with each other,
that's where you get
some of those sort of secondary effects
and make them more pronounced,
because you have
you can imagine it like in a theater.
If you could picture yourself in a theater
and you have one spotlight shining,
you have one point source.
If you had three spotlights,
you're going to get some interesting
sort of overlapping shadows,
and that's one of those effects that
amplify that, the shadow bands.
Oh that's cool.
15 seconds.
Okay, quick whip the glasses on.
Oh my gosh.
Wow. That's so strange.
Man. This is, this is so crazy.
Unbelievable.
I don't know what I was
imagining,
but this is so much more insane.
Everyone is going wild.
I love this guy.
Oh my God, you can see
the corona. And look at that, like,
really bright spot on the left.
Like sticking out. It's like ethereal.
Is that like a solar flare?
I don't, I don't know
I don't know what the words are.
I'm just like, this is cool.
Yeah, right,
- Its the 360!
- Oh, the 360 yeah, you're right.
There's a 360 sunset.
Oh my goodness!.
- I love that you're have a good time here, right.
It's incred... I know you?
Where are you from?
- Veritasium is the channel.
Jesus Christ, this guy's from Veritasium.
This is amazing. 
It’s coming back, it’s coming back!
Look at the sky. It's unbelievable.
It's made my... my life.
It's got me unreasonably fired up
and then like, whoa!
And then that ring in the sky.
This is sweet.
Whoooaa, I have never been this 
pumped up
on a natural event before, my god.
Bro how was that?
How do you feel, man?
It was insane.
It was insane.
Really, I don't know,
I just can't explain it.
Looking at it in person,
it is kind of cataclysmic.
You gotta cry.
We're up at the crack of dawn.
Chasing this balloon.
It did this crazy path, and then ended up
falling much closer than they thought.
Still probably going to be in some farmers
field somewhere.
Might have to do a little hike to get it.
Do we go for it? I mean, I don't know.
He's going for it.
And then
we have our cameras, so that's good.
So this is the one. Yeah.
Okay.
Cool.
Something going on.
That's crazy.
That's so cool.
Favorite shots of the shadow
are where you can just see part of it,
like receding.
Yeah.
So, we had about a seven hour drive
to get here, so we had plenty of time
to catch up on podcasts.
And we ended up listening to one
from today's sponsor, 80,000 hours.
It was about the Fermi Paradox
and some potential solutions.
It was pretty intriguing.
If you're into long conversations
that dig into big ideas,
their podcast is worth checking out.
80,000 hours is a nonprofit
that's built around a simple idea.
Your career is about 80,000 hours long.
That's a huge chunk of your life,
and it's probably your best shot
at making a real impact on the world.
What I like is they don't just tell you
to follow your passion.
Their advice is actually evidence based.
It's supported by over
a decade of research.
The site's got career guides, deep dives
into different paths, and a job board
that's full of high impact roles
and all of it,
the podcast, the research, the guides.
It's all completely free
because they're a nonprofit.
They're not trying to sell you anything.
So if you're trying to figure out
your next step or just curious
how to have more impact with your work,
head to 80000hours.org/Veritasium.
It'll get you their
free career guide, which walks you through
what makes for a high impact career,
it might give you some ideas
you might not have considered,
and it'll help you
turn them into an actual plan.
It'll also sign you up for their newsletter,
so you'll get updates on new research
and job opportunities
a couple times a month.
Thanks to 80,000 hours
for sponsoring this one and as always,
thanks to you for watching.
One last thing.
I wanted
to thank Astrum, and if you want to learn
more about The Sun's corona,
you can check out their video
in the description.
Also, thank you to Exploratorium for
letting us link into their live stream.
And of course, these guys, the team
from NASA borealis for all of their help.
Yeahhhh.
So we've taken a bunch of stickers
and we flew them to the edge of space.
You can see them right
here. We're now all signing them.
It went to what, 92,000ft?
Yeah like 30km? Insane.
If you guys want your hands on one,
we're going to give them out
to our Patreon members. It's totally free.
So if you want to check that out,
the link is also in the description.