July 1, 2025 • 30 mins
👉 Fill out the listener survey - https://services.multiplai.ai/lai-survey
👉 Learn more about the AI Business Transformation Course starting August 11 — spots are limited - http://multiplai.ai/ai-course/
Can you build custom software without hiring a dev team or learning to code?
The short answer? Yes — if you understand Vibe Coding, the new AI-powered way to develop applications just by describing what you want in plain English.
In this jaw-dropping session, Mark Kashef — AI innovator and data scientist — shows business leaders how tools like Lovable, Replit, and Bolt are changing the rules of software development. Whether you're a tech-savvy exec or a total non-coder, this episode will unlock new possibilities for product creation, operational efficiency, and creative freedom.
Pro tip from this episode: Don't just prompt — strategize. Mark reveals how to pair AI tools with platforms like Perplexity to generate high-performing, production-ready prompts that launch serious MVPs.
In this session, you'll discover:
- What Vibe Coding is and why it's blowing up in 2025
- The key differences between Bolt, Lovable, and Replit — and which one is best for your needs
- How non-developers can use AI to design, build, and deploy real apps
- Mark's favorite prompt engineering hacks using Perplexity AI
- How to avoid “spaghetti code” and failed builds by getting the prompt right
- Real-world examples of fast, functional MVPs created with vibe coding
- Insider tricks for editing UI visually and debugging smarter — not harder
- Why these tools are the “nuclear bomb” upgrade to basic AI agents
Mark Kashef is a data scientist, software engineer, and one of the most compelling voices in the world of AI-powered app creation. With a background in software infrastructure and a gift for translating complex tech into simple workflows, Mark’s tutorials and strategies are empowering leaders to build their own tools, faster than ever.
đź“ş Watch his experiments on YouTube: Mark Kashef
Don’t forget — this is part two of the Ultimate AI Showdown. Catch part one in Episode 200, and stay tuned for part three, featuring AI image generation tools, next week!
About Leveraging AI
- The Ultimate AI Course for Business People: https://multiplai.ai/ai-course/
- YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
- Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/
- Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events
If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Welcome to part two of theUltimate AI Showdown.
(00:04):
If for some reason you missedpart one, go back to episode 200
of this podcast.
We hosted Lisa Adams and she wasdoing the ultimate AI showdown
comparing tools for doing dataanalysis for businesses.
It's absolutely brilliant.
So go back and check that outand now two part two, comparing
vibe coding tools with theamazing Mark Kashef.
Isar Meitis (01:08):
our next topic today is maybe the most exciting
aspect of AI development thatexploded in 2025, and that is
Vibe Coding.
Now, those of you who don't knowwhat Vibe coding is, and you
never heard the term before,first of all, the term was
coined by Andres Carpathy, who'sa big name in the AI world,
(01:29):
who's been part of the, one ofthe founders in OpenAI.
He left OpenAI a while back, uh,but he's still, uh, followed by
a lot of people.
But what it actually means, itmeans using AI development
platform and explain to it insimple English or other
languages if you prefer.
What is it that you want todevelop and the platform will
develop the thing for you.
(01:51):
Now, when I say develop fornon-developers like me, that's a
pretty broad term.
Well, depending on the tools youpick, it does the design, it
does the development, it doesthe QA and solving bugs.
It does the deployment.
It creates an application thatyou can actually use.
And again, not all the tools doall these things, but just like
(02:13):
we spoke before, you can mix andmatch these tools to create
applications that you, that youcan use.
Now, the cool thing about it,people like me who are not
coders, who are not computerengineers, can develop
applications for anything thatyou want.
If you are a software engineer,you now have rocket fuel in your
back pocket to do a lot morestuff, a lot faster now.
(02:36):
The interesting thing here isthat if you know what you're
doing, you can do magic If youdon't know what you're doing,
this could be really, reallyscary.
But the cool thing is if youknow English, or like I said,
many other languages as wellwould work.
You can write applications thatwill solve either basic or
complex problems in yourbusiness.
(02:57):
But there's a wide range oftools right now that do a huge
variety of things, and they eachhave a specific target audience
in mind, and they still have.
Pros and cons.
And so our guest today, uh, markKashif comes from a background
of data science.
He's a data scientist, asoftware engineer, but he became
(03:19):
a vibe coding ninja.
He shares priceless videos onYouTube, on how to use AI tools
to do really incredible stuff.
A lot of it around vibe coding,but not only that, but he's
definitely somebody to follow ifyou want to go that next step in
your AI development.
And he brings together bothworlds one and a deep
understanding of how softwareactually works and how data
(03:42):
needs to be analyzed.
And on the other hand, he'sreally, really good in
explaining it to monkeys like meso I can follow and do the same
stuff and develop my ownapplications and do some cool
things with it.
And so from my perspective, vibecoding is the next version of
superpowers that AI gives us.
After we learn how to chat andlearn how to build some basic
(04:04):
agents, now you can use it toactually build applications that
you need.
And why would you wanna buildapplications?
Because every application youuse today has between 20,000 to
172,000 functions when you needfour, and yet you use the really
complex application becausethat's what you have.
Well, that's not the caseanymore.
You can create your ownapplications that do exactly
(04:26):
what you do and connect exactlyto the data that you need and
display it in exactly the waythat you want.
Um.
Consistently.
And so you can start creatingapplications as of today.
And we're gonna do this with theamazing help of the amazing Mark
Cashew.
And so Mark, welcome toleveraging ai.
Mark Kashef (04:46):
Thank you so much.
That's a very humbling intro.
The bar's way too high for whereI axia, so I appreciate it.
Um, I,
Isar Meitis (04:53):
I, that was underplaying it.
Like I, I, uh, what you do isreally amazing, so I appreciate
you.
I appreciate it.
Right on.
Should I share my screen?
Yeah, let's, let's get started.
That's why we're here.
Mark Kashef (05:03):
Alright, cool.
So, hey everyone, I'm gonna walkthrough three main tools right
now when it comes to vibecoding.
There are, if you look it up,probably now a hundred copycats
of these tools, but these arethree of the most powerful ones.
And this basically builds on topof what Lisa just showed in the
last segment.
So she showed.
(05:23):
At the language model level,like at Gemini's level, at
Claude's level, how to buildthese applications.
These are basically what we calla vertical agent that's on top
of those models.
What that means in plain Englishis bolt lovable and replicate
here in order they've all builtinfrastructure on top of these
language models and try to giveas many tips and tricks and
(05:47):
guides behind the scenes to makethese into junior software
engineers.
And that's why you can vibe codewith these in a lot more
powerful way than you can withjust Claude or Gemini or
anything else because they havefocused on marrying software
engineering principles andinfrastructure with language
models.
And just to give a, a visual,instead of just using, um, nerd
(06:08):
speak, if I had to give you avisual of what this looks like,
imagine you have lovable boltand rept and behind the scenes,
behind the curtains, you havethe real star working away,
which is Claude.
And for most of them, similar towhat Lisa mentioned, Claude is
one of the best, if not thebest, right now at writing code,
especially production level codefor software engineering.
(06:30):
So behind the scenes, this isall powered by a brain, and
right now all of them are usingClaude four, which is the latest
and greatest version of Claude.
If I were to give you a quickcomparison before we actually go
into and see how they work andhow they operate and the types
of apps they can make inliterally minutes.
This is a quick comparison.
So Bolt is really good at beingfast and speedy, and the actual
(06:54):
code it writes is pretty wellthought out.
Lovable is the most beautiful.
So if you look at the design, itwins design every single day.
In terms of beauty, in terms ofease of use, it's also really
easy to use as well.
So if you are starting outcompletely from scratch, lovable
is the easiest way to go fromzero to idea on paper that
you're really excited aboutBold.
(07:16):
You have to work a little bitmore on the prompting side to
get there, but Repli is one ofthe most robust, so imagine rep
lit as having a.
Its own repository of tools atits disposal.
They're all rep'sinfrastructure.
What that means in plain Englishis Repla actually used to be
like a cloud provider before,but now they've kind of built on
top of their database tools,their server tools.
(07:37):
So repli can not only build appsfor you, but it can natively
integrate databases, everythingwithout you having to worry
about going the next step andlearning new tools to connect to
these other providers.
So Rep is one of the mostrobust, but it also takes
sometimes four business days tofinish, and I'm exaggerating,
but from the time you send out aprompt, it could be 15 to 20
minutes before you actually getyour first iteration.
(07:59):
So if you're less inclined toknow what you're doing or how to
prompt, you'll get a way fasterfeedback loop with bold and
lovable and really get yourskills upskilled.
Once you're ready, you cangraduate to try to work with Rep
Lit, because then you'll be ableto plan out your prompts a lot
more effectively and have thatquicker feedback loop you're
looking for.
(08:20):
And then this is the last startall show before we actually go
into the platforms and show youwhat I'm actually talking about.
So in terms of MVPs, whichstands for minimum viable
projects or prototypes rather,um, bolt and Lovable are
exceptional for you being in themiddle of a meeting.
Most of the times when I'mspeaking with clients, you we're
trying to visualize what theirdream outcome looks like.
(08:42):
I will use lovable on the side.
I'll take the transcript of myfireflies, dump it into lovable
and say, try to visualize whatthis client's actually looking
for.
And I'll have that run on theside while we're having the
conversation and then I'll beable to show it and say, is this
kind of what you're thinking?
So the power of being able touse that speed is very impactful
and similar to what Isarmentioned you, you wanna know
(09:05):
when it's time to bring a knifeto a gunfight and when to bring
a nuclear bomb to a gunfight.
And it really depends on whichtype of tool you want to use.
So if you wanna build somethinghighly robust in a quick amount
of time, rep's probably the bestway.
If you want to build justsomething quick, a mockup to get
the ideas flowing, lovable is agreat option.
(09:26):
So without further ado, let'sactually take a quick preview of
what these tools look like whenyou log in and then I'll show
you some apps already built out,and I ran it ahead of time just
for time's sake.
But for Bolt, right now, this iswhat it looks like.
You'll notice all of them have abig glaring text box in the
middle of the screen where youcan ask away your dreams in
terms of what you're lookingfor.
(09:46):
You can attach an image by theway to all of them and say,
here's a website I really like,or, here's a design I really
like.
I think you can paste up to fourto five images, send it with a
prompt and say, try to recreatethis to the best of your
ability.
And then one key feature here isthis button.
If I say something really poor,like build me an app to automate
(10:08):
my manufacturing business with asurveillance dashboard, I'm
gonna misspell it on purpose.
You can see here they, theyhave, it's something called
enhanced prompt.
So if you're getting started,this will take your very vague
prompt and make it moreproduction level ready.
So this will also help youbecome a prompt engineer in a
(10:29):
very compressed short amount oftime, because you'll start
reading and understanding theflow, how it expects to be
spoken to.
Now, is this perfect?
No.
And I'm gonna leave as a littleteaser at the very end here.
I'm gonna show you how you canuse perplexity to make you a
prompt engineering deity withoutactually learning the
fundamentals yourself.
But until then, you, you wouldjust send this and then it would
(10:51):
actually pop up the firstresponse.
Isar Meitis (10:55):
One more thing just to add.
Yeah, go ahead.
One short thing.
Yeah.
Uh, because you said aboutattaching files, what I do every
time is I have like a brandguidelines, fancy PDFI upload
that and then it knows how tograb my logo, my colors, my font
and everything.
And it, you don't even have tosay anything.
You literally just upload itthere and then it figures out
what to do with it.
And it works on all thedifferent applications in a, in
(11:15):
a very solid way.
Mark Kashef (11:17):
Exactly.
Yeah.
No, that's great.
Add-in.
Thank you.
And on the lovable side, samething.
You can upload files, uh, interms of images as well.
Um, you can connect the wholeworkspace.
You can connect to somethingcalled super base.
This is basically a databasethat a lot of SaaS platforms use
to store data, but also enablefunctionality, which I'll get to
shortly.
(11:38):
And this one's prettystraightforward.
You send the, the prompt as welland you'll get some form of user
interface.
And the last one is rept, wherethis one's a slightly different.
So Rept on the left hand sidehere, you can actually click on
this and you can select exactlywhat kind of platform you're
trying to optimize for.
So if you wanna build a modernweb app, you can just click this
and it will know behind thescene.
(11:59):
It'll get a cheat sheet injectedbehind the scene that that's
exactly what you're going forversus having to guess.
If you wanna just pick autobased on what you write in
natural language, it will guesswhat you're trying to build.
This is the only one that I'lltry to run live, just to show
you how it's different from theothers.
So if I use a, the prompt wejust made in Bolt.
Okay.
And you run it here, like Isaid, this will take.
(12:22):
A while to get to the pointwhere the interface is there.
But one thing it's exceptionalat is it plans out loud and it
plans a prototype before it evenbuilds the prototype.
So it tells you basically, thisis kind of what I'm thinking.
This looks like, is this kind ofwhere we're at?
Or should I pivot my thinkingbefore I even build the first
version?
(12:42):
And the reason why it's sothoughtful, you can see right
here it's now building a visualpreview of a prototype of my
prototype to see if we're in theright direction.
It's giving you a list of allthe items.
It's planning on adding asfeatures.
And what's also cool is even ifyou didn't come up with a
feature yourself, it tries tosuggest features that could be
richer.
So rep is, like I said, can be amuch richer experience overall.
(13:06):
And then it always asks for yourpermission, which I love.
I feel like an actual boss to ajunior software engineer where
you have to AP approve and startthe project in order for it to
actually go and build it out.
And once this finishes running,we should get some form of
interface here that willdisplay.
You'll see right now productiondashboard, there's something
loading here.
It's building these componentsas we speak right in front of
(13:29):
me.
There we go.
We got some modals here.
And by the end of this I can saylike, this looks kind of boring,
or I wish it was something else.
Yep.
Once I click a proven start,then it'll be on its way.
It'll probably take around 10 to15 minutes because one thing rep
(13:49):
likes to do is it tries to makeeverything functional, not just
build a mockup.
It tries to actually make thedatabase work.
It tries to pop up a modal thatsays you says you wanna use
OpenAI.
Great.
Go get me your API key.
You wanna use stripe?
Go get your stripe key.
So it's very comprehensive inits responses.
So that's what I wanted to showas a preview before we actually
look at apps.
(14:10):
Now I'm gonna show you theprompt that I've used.
I'm going to show you two setsof apps.
This prompt is eight pages.
Is it because I'm so inherentlygifted and smart and intelligent
that I wrote an eight pageprompt?
No, I actually dictated a verysmall brain dump to perplexity,
and I gave it a prompt that I'llshow you shortly, and it helped
(14:33):
me build a very comprehensiveprompt.
And the crux of the prompt isbuild a comprehensive AI powered
IT operations command centerthat combines real-time
infrastructure monitoring,cybersecurity thread detection,
and gamified collaboration in amodern interactive web app.
And then it goes through a bunchof requirements here that again,
I would love to take credit for,but all of these eight pages of
(14:56):
requirements and designs I justapproved by going back and forth
with perplexity.
So I took this exact prompt andI threw it into each platform to
give you a sense of how it wouldlook in different platforms.
So Numera Uno Bolt, I purposelyjust showed the code instead of
the, the web app in Bolt.
It took around probably, uh,five minutes or six minutes and
(15:18):
we came up with this, which isactually pretty beautiful.
It's responsive.
Um, it has pretty much all theelements I asked for.
You can notice here that thispart looks really similar to
what we, we just saw in Rept.
There's a reason for that.
A lot of these vibe coding appsuse certain what are called
libraries like, like a normallibrary you think of, but let's
(15:39):
say in the coding world whereyou have a series of frameworks
or where all of these platformslike to get icons from to
create, populate these iconshere.
And it put together not only onetab, but if you click on
multiple tabs, you'll see heresecurity center, you have
security alerts.
Obviously this is allfictitional data that you can
make production data, but itstarts out also drafting out
(16:02):
what the profile would looklike, et cetera.
And this is only one shot.
Now with Vibe coding, I like tosay that you're only as good as
your last prompt.
So the most important thing whenyou Vibe code is you'll notice
this is my very first prompt.
Is every prompt I send to theseplatforms the same, no.
(16:22):
The first prompt for any ofthese platforms is kind of like
building a house.
So if you had a constructionteam working for you and you
told them the blueprint of thehouse, right?
If you don't tell them at thebeginning that you want a pool
in the middle of the courtyardof your house and they start
planning the building, thehouse, going back later, halfway
through the app and tellingthem, oh, I want a pool.
(16:44):
Even though there's already tworooms there in a kitchen and the
plumbing's already been put out,that that becomes a tricky
problem to go through.
So if you don't want to pivothalfway and now you have to do a
change order like you would inreal life, and they have to go
scrap a bunch of code, you startgetting what's called spaghetti
code, where now you have allthese types of codes on top of
each other, and you go into thisendless loop where you're not
(17:06):
accomplishing what you'relooking for, you're getting
constant errors.
And if you're not as tech savvy,you can't even read those errors
and you keep clicking, try fix,praying that it knows what to
do, and then you eventually say,this vibe coding stuff sucks.
Bye to prevent that fromhappening, you wanna be as
thoughtful as possible beforeyou even kick off your journey.
And then the likelihood there'llbe more accurate will be much
(17:29):
higher.
So that was both.
Isar Meitis (17:31):
So to pause it just for one second and connect what
you said before aboutperplexity, I do the same thing.
I have a project in Claude,which is I did a lot of research
on what needs to go into productdesign, and then I add.
That as the input into thatthing.
And then what I do is say, okay,this is my idea.
I want you to ask me multiplequestions to punch holes and,
(17:54):
and figure out the detailsbecause I don't know the
details.
The other thing is it has allthe things you need to think
about when you design a softwarelike the user interface.
Does it need to be mobile or notmobile, like all you got, all
the gazillion other things thatyou want to be able to do.
And then it has just asks me allthese questions.
Now, some of them I don't knowand I said, well, I don't know.
(18:14):
What do you suggest?
And then it makes a suggestionand I can pick from it
suggestions.
But think about the AI as yoursoftware design assistant that
most likely knows about softwaredesign.
Way more than you do, even ifyou're a software engineer like
Mark.
Definitely if you're somebodylike me now, I have the benefit.
I ran software companies foralmost 20 years, so I
(18:36):
understand.
I don't write code, but Iunderstand how software is built
and how it works.
But if you don't, it's stillfine.
You can still have thatconversation back and forth.
And only after I have a longconversation and then I have a
very solid prompt, again, likeMark said, explaining that needs
to be a pool in the middle ofthe courtyard.
Then I go into one of the toolsand actually let it run and the
(18:57):
results are significantly afaster be better.
And see, like Mark said, forpeople like me, significantly
less frustrating because I don'tknow how to fix it.
And I'll say one more thingabout the whole getting stuck in
the loop of not being able tofix.
When you're doing this five orsix times and it doesn't solve
the problem, just ask it to goin a different path.
(19:17):
'cause you're ch you don'tunderstand how the code runs.
You say, you know what, forgetit.
Let's start this component fromthe beginning.
Yes.
And then just that part startsscratch.
Instead of trying to fix 15different things, your chances
of getting it done issignificantly higher.
Mark Kashef (19:34):
Yep.
Exactly.
No, that was amazing.
Isar Meitis (19:36):
You know, one, one last thing, sorry.
Uh, most of these tools havelike an assistant that doesn't
actually write code that you canjust say, okay, let's brainstorm
this for a minute.
Uh, inside the tool itself,right?
I do this in replicate all thetime and then it doesn't write
code.
It just helps you kinda liketroubleshoot and come up with
new ideas on how to solve theproblem.
And then you go back and try toedit.
Mark Kashef (19:55):
Yeah, you stole my thunder.
Yeah, I was gonna show thatnext.
So, um, oh, awesome.
Perfect.
No, you're good.
So, um, I was just gonna showthe app first in lovable and
then repla and then show thatfunctionality.
So in lovable, that exact sameprompt I showed you, if I put
this back, you'll see visually,although it's somewhat similar
to what I just showed you, it isever so slightly prettier.
It, it focuses on the littlethings like the size of the
(20:17):
font.
You'll notice bolt here, it kindof put a lot on one screen and I
have a bit of overload,cognitive overload when I take a
look at it.
So even though this is pretty,there's a lot of going, there's
a lot going on.
There's titles, there'ssubtitles, there's headers.
So lovable is very lovable interms of the designs it creates.
And as you go from tab to tab,you'll just notice that it's
(20:38):
very easy on the eyes.
Uh, lemme see.
This one works.
At least see this one.
Um, and you'll see by the way,if, if there's an error in the
code, you'll always see thispopup that says, try to fix.
So either you can look at thelogs if you're a nerd like me,
or you can click on try to fix,and it will try to figure out
what is happening in theunderlying code.
And then last but not least,this is what Repla put together.
(20:58):
So repla, I will say, willalways be slightly less pretty
than both those platforms.
But in terms of functionality,you'll see like it jumps ahead,
it asks permission.
It's asking me right now, howdoes the application look and
feel to you?
Are all the features working asexpected?
So it's literally waiting for myfeedback to tell it, yeah, you
did a good job or not a goodjob.
The other ones are kinda like,here, take it.
(21:19):
Uh, here's my report.
Uh, lemme know if you like it.
If we go back to what Azar wasmentioning, all of these tools
have something that's calledchat mode.
In chat mode, I can actually aska question like, what do you
think we should add next?
Okay.
And then instead of actuallyexecuting and writing code and.
(21:41):
Making the application morebloated.
It just talks to me.
I'm talking to Claude now, butI'm talking to Claude in the
context of my project.
So if it comes up with an ideaor a plan, I can read through
it.
If it looks logical, all I haveto do is click on implement this
plan right here and then now itwill go and execute it.
But at least now we're on thesame page.
(22:02):
I noticed that a lot of timesthat I will spend 70% of my time
even in lovable right here.
It'll be called uh, chat righthere.
If you click on chat mode, I'llspend 70% of my time here.
I'm only 30% of my time actuallywriting the code.
'cause I'll spend five 10messages to make sure we are
100% on the same page before wego and write the thing or do the
(22:22):
thing.
And then rep has, uh, the samething as well.
And then rep's just really goodat dealing with feedback in
general.
All of these have also, uh,visual editors.
So if I go to here, um, I canclick on this button right here
and then I can go select aspecific element.
And you'll notice on the lefthand side, it's now referring to
(22:43):
this exact element.
So instead of.
Praying that I send the promptand it knows exactly what I'm
talking about.
I can say, no, this one, this iswhat I wanna change.
Same thing with, uh, lovable.
You can click on visual edit andI can say, this button, this
button's ugly.
And I don't even have to send aprompt.
Sometimes if I wanna just, um,change the size, I could just do
(23:03):
it here without a prompt.
And that's how a lot of theseplatforms differ ever so
slightly.
You can't do this exactfunctionality in Bolt, but in
lovable, you're allowed to bespecific.
I can even, by the way, Google aspecific font name, paste it
here, click save, and it willmake the change in the code,
which is very powerful becauseit saves you from having to
guess how to tell it the thingin natural language.
(23:25):
And then in terms of rept, Idon't think there's, uh, is
there a prompt helper?
I don't recall that, um, myself.
But at the very beginning, youcan get some help going back and
forth on flushing out theprompt.
And one thing I just wanna showyou on that rept, again, small
nuance is you'll notice itactually takes physical
(23:46):
screenshots of the webapplication while it's
processing, whereas the otherones are solely depending on
code.
So where I say it's more robustis it's not just looking at its
code, it also is takingscreenshots.
Like I would send to OpenAI orClaude and saying, analyze and
see what's wrong.
Let's see, before the usercatches it, let's catch it
ourselves.
(24:07):
And what I mentioned before interms of being more robust, you
can notice here if I go on.
Uh, database deployments, allthis stuff.
I go to all tools.
All of these tools are pre-builtin rept.
So if you want a database, ifyou want the ability to use
Google Sign in very easily, allthat's built into what's called
rept authentication.
They've kind of made one wholeecosystem of tools.
(24:29):
If I go back to Bolt andLovable, if I want to enable
something like Google Sign in,I'm gonna have to go to
something like integrationshere.
And then I'm gonna have tocreate what's called a superb
base account, which is like adatabase account.
And I'm gonna have to connect tothat superb base account, then
give it some credentials, thenit can start to build functions
that allow me to doauthentication.
So that's just one smallmicrocosm of using these tools,
(24:52):
the differences.
These can get me a prototypeimmediately without me having to
think about anything in depth.
Rept will allow me to go muchdeeper.
And what another cool thingabout Rept is they have a mobile
app.
So if you're on, uh, the bus, ifyou're on a plane, if you're
waiting in the airport, I couldhave built this on a mobile app.
And just prototyped on the go.
(25:12):
Vibe coded while waiting for mydelayed flight, and then I'd be
on my way.
So that's just one more idea.
And the last thing is on theRepla side, I've used Repla as a
developer for years to deployapplications on a server.
The fact that it's all built in,again, in that one ecosystem,
that means that they're reallygood at cloud, they have secure
cloud, they're already trustedby hundreds of organizations,
(25:35):
and now they're just marryingtheir whole ecosystem to cloud.
So that's where I'd give Rep theedge on robustness.
Now, being mindful of time, um,I'll see how much time I have
left.
I'll just show you the hack andperplexity.
If you're looking at all thisand you're like, I'm still stuck
on how you wrote that prompt,I'd rather unblock that for you
before we end up.
So in perplexity, and you canuse what, whatever language
(25:58):
model you wish that's connectedto or hooked up to, uh, a web
search.
So you can use, I usually usedeep research, but just for
time, I'm just gonna use, uh, avery quick search.
And what I'll do is thefollowing, and I'll just dictate
in natural language so you cansee how poor I am as a prompt
engineer.
So I want you to act as a promptengineer with a specialty in
(26:19):
product design as well.
And I want you to look up allthe documentation, watch YouTube
videos on the platform's,lovable rept and Bolt.
And I want you to come up with auniversal prompt that I can
create for my e-commercebusiness that would work for all
of them.
And build a beautifulminimalistic e-commerce app that
lets me manage my Shopify ordersprogrammatically in just using
(26:43):
my natural language, typing insome form, some form of text
box.
And I want you to output thatprompt in an un unformatted code
block in mark down so I can justeasily copy paste it.
So in this case, in plainEnglish, I just said go and
watch YouTube videos, go andread documentation on all these
apps, and it will know that it'susing Claude behind the scenes.
(27:04):
So in a way, it's also gonnaresearch that.
And when I send this, you'regonna see what's gonna do.
It's gonna go through lovabledocumentation, it's gonna go
through, uh, rapiddocumentation.
It watched the ultimate guideon, oh, if I go back, my bad, it
watched the Ultimate Guide onusing, uh, repli and Bolt.
So let me just try that again.
(27:27):
Worst case I'll just do, therewe go.
So lemme go back to what it did.
Go to sources.
So I went to Reddit, uh, went toYouTube videos and watched, uh,
I watched videos on.
Shopify integrations.
It went through gpt, it wentthrough GitHub, which is
basically a repository fordifferent apps.
(27:47):
And then it came up with, afterdoing all that this prompt, and
you can see here, this exposesmy lack of genius'cause it did
all the hard work.
Build a beautiful minimalistice-commerce web app that connects
to Shopify store and allows meto manage orders
programmatically.
And then it gives a list of allthe different requirements,
clean, modern, minimalistic UIwith responsive design, and
(28:09):
basically a laundry list of allthe things we want in this app.
And then you could just copythis and throw it into whatever
web app builder you're lookingfor, and you'd be on your way.
And the cool thing is, one extrapro tip I'll give you is if you
get stuck, what Isar mentionedis amazing, which is trying to
pivot or going to chat mode todiscuss.
But one other thing you could dois, I like to actually, since
(28:32):
you have access to the code,even if you can't read the code,
I like to ask it, okay, where doyou, which file has this problem
you're talking about?
And it'll tell me this AIanalytics tsx, and I'll just go
copy, paste this whole thing,give it to maybe Gemini or
Claude, uh, a brand new versionof Claude with virgin eyes on
the matter.
And maybe they'll say, you knowwhat?
It's missing this.
And you're like, aha, this isthe problem.
(28:54):
So the fact that you have accessto code, even if you don't know
how to read code, you can use itto help you get back on track if
things go off the chain.
So that's on the prompting side.
I'm gonna pause
Isar Meitis (29:05):
This was nothing short of mind blowing, like the
chop people, like this isphenomenal.
This is incredible.
I can't believe I did not knowabout this.
Good.
And I did a littlequestionnaire, like a little
survey, who knows about vibecoding.
And most people didn't have aclue or kind of like knew about
it, but didn't.
So I'm, I'm sure this is gonnatake a lot of sleeping hours
from people to play over thenext few days, uh, which was the
(29:26):
whole point.
Uh, mark, again, this wasnothing short of exceptional,
uh, as expected.
So I set the bar high, but youjumped like 10 feet above the
bar.
Thank you.
So thank you so much.
I appreciate it.
Uh, if people wanna work withyou, follow, you know, you
connect with you, what are thebest ways to do that?
Mark Kashef (29:43):
Sure.
Yeah.
So, uh, YouTube is just MarkCash if, uh, if you wanna follow
my Mad Scientist experiments.
Um, otherwise, uh, promptadvisors.com if you wanna engage
on a project.
I just threw it in the chatthere.
And yeah, I mean, YouTube willgive you 90% of my life, uh, and
what I do and what I'm workingon.
So that'll be the best way togo.
This is the end of part two ofthe Ultimate AI Showdown.
(30:06):
But don't worry, there's partthree and four of this
incredible series of theshowdown.
So next week we will release thenext episode.
That will be episode 2 0 4 ofthis podcast in which Rory
Flynn, one of the top AI imagegeneration professionals in the
world today, will compare AIimage generation tools and how
(30:27):
to pick them and how to use themin the most effective way.
And until then, have an amazingrest of your week.
Welcome to part two of theUltimate AI Showdown.
(00:04):
If for some reason you missedpart one, go back to episode 200
of this podcast.
We hosted Lisa Adams and she wasdoing the ultimate AI showdown
comparing tools for doing dataanalysis for businesses.
It's absolutely brilliant.
So go back and check that outand now two part two, comparing
vibe coding tools with theamazing Mark Kashef.
Isar Meitis (01:08):
our next topic today is maybe the most exciting
aspect of AI development thatexploded in 2025, and that is
Vibe Coding.
Now, those of you who don't knowwhat Vibe coding is, and you
never heard the term before,first of all, the term was
coined by Andres Carpathy, who'sa big name in the AI world,
(01:29):
who's been part of the, one ofthe founders in OpenAI.
He left OpenAI a while back, uh,but he's still, uh, followed by
a lot of people.
But what it actually means, itmeans using AI development
platform and explain to it insimple English or other
languages if you prefer.
What is it that you want todevelop and the platform will
develop the thing for you.
(01:51):
Now, when I say develop fornon-developers like me, that's a
pretty broad term.
Well, depending on the tools youpick, it does the design, it
does the development, it doesthe QA and solving bugs.
It does the deployment.
It creates an application thatyou can actually use.
And again, not all the tools doall these things, but just like
(02:13):
we spoke before, you can mix andmatch these tools to create
applications that you, that youcan use.
Now, the cool thing about it,people like me who are not
coders, who are not computerengineers, can develop
applications for anything thatyou want.
If you are a software engineer,you now have rocket fuel in your
back pocket to do a lot morestuff, a lot faster now.
(02:36):
The interesting thing here isthat if you know what you're
doing, you can do magic If youdon't know what you're doing,
this could be really, reallyscary.
But the cool thing is if youknow English, or like I said,
many other languages as wellwould work.
You can write applications thatwill solve either basic or
complex problems in yourbusiness.
(02:57):
But there's a wide range oftools right now that do a huge
variety of things, and they eachhave a specific target audience
in mind, and they still have.
Pros and cons.
And so our guest today, uh, markKashif comes from a background
of data science.
He's a data scientist, asoftware engineer, but he became
(03:19):
a vibe coding ninja.
He shares priceless videos onYouTube, on how to use AI tools
to do really incredible stuff.
A lot of it around vibe coding,but not only that, but he's
definitely somebody to follow ifyou want to go that next step in
your AI development.
And he brings together bothworlds one and a deep
understanding of how softwareactually works and how data
(03:42):
needs to be analyzed.
And on the other hand, he'sreally, really good in
explaining it to monkeys like meso I can follow and do the same
stuff and develop my ownapplications and do some cool
things with it.
And so from my perspective, vibecoding is the next version of
superpowers that AI gives us.
After we learn how to chat andlearn how to build some basic
(04:04):
agents, now you can use it toactually build applications that
you need.
And why would you wanna buildapplications?
Because every application youuse today has between 20,000 to
172,000 functions when you needfour, and yet you use the really
complex application becausethat's what you have.
Well, that's not the caseanymore.
You can create your ownapplications that do exactly
(04:26):
what you do and connect exactlyto the data that you need and
display it in exactly the waythat you want.
Um.
Consistently.
And so you can start creatingapplications as of today.
And we're gonna do this with theamazing help of the amazing Mark
Cashew.
And so Mark, welcome toleveraging ai.
Mark Kashef (04:46):
Thank you so much.
That's a very humbling intro.
The bar's way too high for whereI axia, so I appreciate it.
Um, I,
Isar Meitis (04:53):
I, that was underplaying it.
Like I, I, uh, what you do isreally amazing, so I appreciate
you.
I appreciate it.
Right on.
Should I share my screen?
Yeah, let's, let's get started.
That's why we're here.
Mark Kashef (05:03):
Alright, cool.
So, hey everyone, I'm gonna walkthrough three main tools right
now when it comes to vibecoding.
There are, if you look it up,probably now a hundred copycats
of these tools, but these arethree of the most powerful ones.
And this basically builds on topof what Lisa just showed in the
last segment.
So she showed.
(05:23):
At the language model level,like at Gemini's level, at
Claude's level, how to buildthese applications.
These are basically what we calla vertical agent that's on top
of those models.
What that means in plain Englishis bolt lovable and replicate
here in order they've all builtinfrastructure on top of these
language models and try to giveas many tips and tricks and
(05:47):
guides behind the scenes to makethese into junior software
engineers.
And that's why you can vibe codewith these in a lot more
powerful way than you can withjust Claude or Gemini or
anything else because they havefocused on marrying software
engineering principles andinfrastructure with language
models.
And just to give a, a visual,instead of just using, um, nerd
(06:08):
speak, if I had to give you avisual of what this looks like,
imagine you have lovable boltand rept and behind the scenes,
behind the curtains, you havethe real star working away,
which is Claude.
And for most of them, similar towhat Lisa mentioned, Claude is
one of the best, if not thebest, right now at writing code,
especially production level codefor software engineering.
(06:30):
So behind the scenes, this isall powered by a brain, and
right now all of them are usingClaude four, which is the latest
and greatest version of Claude.
If I were to give you a quickcomparison before we actually go
into and see how they work andhow they operate and the types
of apps they can make inliterally minutes.
This is a quick comparison.
So Bolt is really good at beingfast and speedy, and the actual
(06:54):
code it writes is pretty wellthought out.
Lovable is the most beautiful.
So if you look at the design, itwins design every single day.
In terms of beauty, in terms ofease of use, it's also really
easy to use as well.
So if you are starting outcompletely from scratch, lovable
is the easiest way to go fromzero to idea on paper that
you're really excited aboutBold.
(07:16):
You have to work a little bitmore on the prompting side to
get there, but Repli is one ofthe most robust, so imagine rep
lit as having a.
Its own repository of tools atits disposal.
They're all rep'sinfrastructure.
What that means in plain Englishis Repla actually used to be
like a cloud provider before,but now they've kind of built on
top of their database tools,their server tools.
(07:37):
So repli can not only build appsfor you, but it can natively
integrate databases, everythingwithout you having to worry
about going the next step andlearning new tools to connect to
these other providers.
So Rep is one of the mostrobust, but it also takes
sometimes four business days tofinish, and I'm exaggerating,
but from the time you send out aprompt, it could be 15 to 20
minutes before you actually getyour first iteration.
(07:59):
So if you're less inclined toknow what you're doing or how to
prompt, you'll get a way fasterfeedback loop with bold and
lovable and really get yourskills upskilled.
Once you're ready, you cangraduate to try to work with Rep
Lit, because then you'll be ableto plan out your prompts a lot
more effectively and have thatquicker feedback loop you're
looking for.
(08:20):
And then this is the last startall show before we actually go
into the platforms and show youwhat I'm actually talking about.
So in terms of MVPs, whichstands for minimum viable
projects or prototypes rather,um, bolt and Lovable are
exceptional for you being in themiddle of a meeting.
Most of the times when I'mspeaking with clients, you we're
trying to visualize what theirdream outcome looks like.
(08:42):
I will use lovable on the side.
I'll take the transcript of myfireflies, dump it into lovable
and say, try to visualize whatthis client's actually looking
for.
And I'll have that run on theside while we're having the
conversation and then I'll beable to show it and say, is this
kind of what you're thinking?
So the power of being able touse that speed is very impactful
and similar to what Isarmentioned you, you wanna know
(09:05):
when it's time to bring a knifeto a gunfight and when to bring
a nuclear bomb to a gunfight.
And it really depends on whichtype of tool you want to use.
So if you wanna build somethinghighly robust in a quick amount
of time, rep's probably the bestway.
If you want to build justsomething quick, a mockup to get
the ideas flowing, lovable is agreat option.
(09:26):
So without further ado, let'sactually take a quick preview of
what these tools look like whenyou log in and then I'll show
you some apps already built out,and I ran it ahead of time just
for time's sake.
But for Bolt, right now, this iswhat it looks like.
You'll notice all of them have abig glaring text box in the
middle of the screen where youcan ask away your dreams in
terms of what you're lookingfor.
(09:46):
You can attach an image by theway to all of them and say,
here's a website I really like,or, here's a design I really
like.
I think you can paste up to fourto five images, send it with a
prompt and say, try to recreatethis to the best of your
ability.
And then one key feature here isthis button.
If I say something really poor,like build me an app to automate
(10:08):
my manufacturing business with asurveillance dashboard, I'm
gonna misspell it on purpose.
You can see here they, theyhave, it's something called
enhanced prompt.
So if you're getting started,this will take your very vague
prompt and make it moreproduction level ready.
So this will also help youbecome a prompt engineer in a
(10:29):
very compressed short amount oftime, because you'll start
reading and understanding theflow, how it expects to be
spoken to.
Now, is this perfect?
No.
And I'm gonna leave as a littleteaser at the very end here.
I'm gonna show you how you canuse perplexity to make you a
prompt engineering deity withoutactually learning the
fundamentals yourself.
But until then, you, you wouldjust send this and then it would
(10:51):
actually pop up the firstresponse.
Isar Meitis (10:55):
One more thing just to add.
Yeah, go ahead.
One short thing.
Yeah.
Uh, because you said aboutattaching files, what I do every
time is I have like a brandguidelines, fancy PDFI upload
that and then it knows how tograb my logo, my colors, my font
and everything.
And it, you don't even have tosay anything.
You literally just upload itthere and then it figures out
what to do with it.
And it works on all thedifferent applications in a, in
(11:15):
a very solid way.
Mark Kashef (11:17):
Exactly.
Yeah.
No, that's great.
Add-in.
Thank you.
And on the lovable side, samething.
You can upload files, uh, interms of images as well.
Um, you can connect the wholeworkspace.
You can connect to somethingcalled super base.
This is basically a databasethat a lot of SaaS platforms use
to store data, but also enablefunctionality, which I'll get to
shortly.
(11:38):
And this one's prettystraightforward.
You send the, the prompt as welland you'll get some form of user
interface.
And the last one is rept, wherethis one's a slightly different.
So Rept on the left hand sidehere, you can actually click on
this and you can select exactlywhat kind of platform you're
trying to optimize for.
So if you wanna build a modernweb app, you can just click this
and it will know behind thescene.
(11:59):
It'll get a cheat sheet injectedbehind the scene that that's
exactly what you're going forversus having to guess.
If you wanna just pick autobased on what you write in
natural language, it will guesswhat you're trying to build.
This is the only one that I'lltry to run live, just to show
you how it's different from theothers.
So if I use a, the prompt wejust made in Bolt.
Okay.
And you run it here, like Isaid, this will take.
(12:22):
A while to get to the pointwhere the interface is there.
But one thing it's exceptionalat is it plans out loud and it
plans a prototype before it evenbuilds the prototype.
So it tells you basically, thisis kind of what I'm thinking.
This looks like, is this kind ofwhere we're at?
Or should I pivot my thinkingbefore I even build the first
version?
(12:42):
And the reason why it's sothoughtful, you can see right
here it's now building a visualpreview of a prototype of my
prototype to see if we're in theright direction.
It's giving you a list of allthe items.
It's planning on adding asfeatures.
And what's also cool is even ifyou didn't come up with a
feature yourself, it tries tosuggest features that could be
richer.
So rep is, like I said, can be amuch richer experience overall.
(13:06):
And then it always asks for yourpermission, which I love.
I feel like an actual boss to ajunior software engineer where
you have to AP approve and startthe project in order for it to
actually go and build it out.
And once this finishes running,we should get some form of
interface here that willdisplay.
You'll see right now productiondashboard, there's something
loading here.
It's building these componentsas we speak right in front of
(13:29):
me.
There we go.
We got some modals here.
And by the end of this I can saylike, this looks kind of boring,
or I wish it was something else.
Yep.
Once I click a proven start,then it'll be on its way.
It'll probably take around 10 to15 minutes because one thing rep
(13:49):
likes to do is it tries to makeeverything functional, not just
build a mockup.
It tries to actually make thedatabase work.
It tries to pop up a modal thatsays you says you wanna use
OpenAI.
Great.
Go get me your API key.
You wanna use stripe?
Go get your stripe key.
So it's very comprehensive inits responses.
So that's what I wanted to showas a preview before we actually
look at apps.
(14:10):
Now I'm gonna show you theprompt that I've used.
I'm going to show you two setsof apps.
This prompt is eight pages.
Is it because I'm so inherentlygifted and smart and intelligent
that I wrote an eight pageprompt?
No, I actually dictated a verysmall brain dump to perplexity,
and I gave it a prompt that I'llshow you shortly, and it helped
(14:33):
me build a very comprehensiveprompt.
And the crux of the prompt isbuild a comprehensive AI powered
IT operations command centerthat combines real-time
infrastructure monitoring,cybersecurity thread detection,
and gamified collaboration in amodern interactive web app.
And then it goes through a bunchof requirements here that again,
I would love to take credit for,but all of these eight pages of
(14:56):
requirements and designs I justapproved by going back and forth
with perplexity.
So I took this exact prompt andI threw it into each platform to
give you a sense of how it wouldlook in different platforms.
So Numera Uno Bolt, I purposelyjust showed the code instead of
the, the web app in Bolt.
It took around probably, uh,five minutes or six minutes and
(15:18):
we came up with this, which isactually pretty beautiful.
It's responsive.
Um, it has pretty much all theelements I asked for.
You can notice here that thispart looks really similar to
what we, we just saw in Rept.
There's a reason for that.
A lot of these vibe coding appsuse certain what are called
libraries like, like a normallibrary you think of, but let's
(15:39):
say in the coding world whereyou have a series of frameworks
or where all of these platformslike to get icons from to
create, populate these iconshere.
And it put together not only onetab, but if you click on
multiple tabs, you'll see heresecurity center, you have
security alerts.
Obviously this is allfictitional data that you can
make production data, but itstarts out also drafting out
(16:02):
what the profile would looklike, et cetera.
And this is only one shot.
Now with Vibe coding, I like tosay that you're only as good as
your last prompt.
So the most important thing whenyou Vibe code is you'll notice
this is my very first prompt.
Is every prompt I send to theseplatforms the same, no.
(16:22):
The first prompt for any ofthese platforms is kind of like
building a house.
So if you had a constructionteam working for you and you
told them the blueprint of thehouse, right?
If you don't tell them at thebeginning that you want a pool
in the middle of the courtyardof your house and they start
planning the building, thehouse, going back later, halfway
through the app and tellingthem, oh, I want a pool.
(16:44):
Even though there's already tworooms there in a kitchen and the
plumbing's already been put out,that that becomes a tricky
problem to go through.
So if you don't want to pivothalfway and now you have to do a
change order like you would inreal life, and they have to go
scrap a bunch of code, you startgetting what's called spaghetti
code, where now you have allthese types of codes on top of
each other, and you go into thisendless loop where you're not
(17:06):
accomplishing what you'relooking for, you're getting
constant errors.
And if you're not as tech savvy,you can't even read those errors
and you keep clicking, try fix,praying that it knows what to
do, and then you eventually say,this vibe coding stuff sucks.
Bye to prevent that fromhappening, you wanna be as
thoughtful as possible beforeyou even kick off your journey.
And then the likelihood there'llbe more accurate will be much
(17:29):
higher.
So that was both.
Isar Meitis (17:31):
So to pause it just for one second and connect what
you said before aboutperplexity, I do the same thing.
I have a project in Claude,which is I did a lot of research
on what needs to go into productdesign, and then I add.
That as the input into thatthing.
And then what I do is say, okay,this is my idea.
I want you to ask me multiplequestions to punch holes and,
(17:54):
and figure out the detailsbecause I don't know the
details.
The other thing is it has allthe things you need to think
about when you design a softwarelike the user interface.
Does it need to be mobile or notmobile, like all you got, all
the gazillion other things thatyou want to be able to do.
And then it has just asks me allthese questions.
Now, some of them I don't knowand I said, well, I don't know.
(18:14):
What do you suggest?
And then it makes a suggestionand I can pick from it
suggestions.
But think about the AI as yoursoftware design assistant that
most likely knows about softwaredesign.
Way more than you do, even ifyou're a software engineer like
Mark.
Definitely if you're somebodylike me now, I have the benefit.
I ran software companies foralmost 20 years, so I
(18:36):
understand.
I don't write code, but Iunderstand how software is built
and how it works.
But if you don't, it's stillfine.
You can still have thatconversation back and forth.
And only after I have a longconversation and then I have a
very solid prompt, again, likeMark said, explaining that needs
to be a pool in the middle ofthe courtyard.
Then I go into one of the toolsand actually let it run and the
(18:57):
results are significantly afaster be better.
And see, like Mark said, forpeople like me, significantly
less frustrating because I don'tknow how to fix it.
And I'll say one more thingabout the whole getting stuck in
the loop of not being able tofix.
When you're doing this five orsix times and it doesn't solve
the problem, just ask it to goin a different path.
(19:17):
'cause you're ch you don'tunderstand how the code runs.
You say, you know what, forgetit.
Let's start this component fromthe beginning.
Yes.
And then just that part startsscratch.
Instead of trying to fix 15different things, your chances
of getting it done issignificantly higher.
Mark Kashef (19:34):
Yep.
Exactly.
No, that was amazing.
Isar Meitis (19:36):
You know, one, one last thing, sorry.
Uh, most of these tools havelike an assistant that doesn't
actually write code that you canjust say, okay, let's brainstorm
this for a minute.
Uh, inside the tool itself,right?
I do this in replicate all thetime and then it doesn't write
code.
It just helps you kinda liketroubleshoot and come up with
new ideas on how to solve theproblem.
And then you go back and try toedit.
Mark Kashef (19:55):
Yeah, you stole my thunder.
Yeah, I was gonna show thatnext.
So, um, oh, awesome.
Perfect.
No, you're good.
So, um, I was just gonna showthe app first in lovable and
then repla and then show thatfunctionality.
So in lovable, that exact sameprompt I showed you, if I put
this back, you'll see visually,although it's somewhat similar
to what I just showed you, it isever so slightly prettier.
It, it focuses on the littlethings like the size of the
(20:17):
font.
You'll notice bolt here, it kindof put a lot on one screen and I
have a bit of overload,cognitive overload when I take a
look at it.
So even though this is pretty,there's a lot of going, there's
a lot going on.
There's titles, there'ssubtitles, there's headers.
So lovable is very lovable interms of the designs it creates.
And as you go from tab to tab,you'll just notice that it's
(20:38):
very easy on the eyes.
Uh, lemme see.
This one works.
At least see this one.
Um, and you'll see by the way,if, if there's an error in the
code, you'll always see thispopup that says, try to fix.
So either you can look at thelogs if you're a nerd like me,
or you can click on try to fix,and it will try to figure out
what is happening in theunderlying code.
And then last but not least,this is what Repla put together.
(20:58):
So repla, I will say, willalways be slightly less pretty
than both those platforms.
But in terms of functionality,you'll see like it jumps ahead,
it asks permission.
It's asking me right now, howdoes the application look and
feel to you?
Are all the features working asexpected?
So it's literally waiting for myfeedback to tell it, yeah, you
did a good job or not a goodjob.
The other ones are kinda like,here, take it.
(21:19):
Uh, here's my report.
Uh, lemme know if you like it.
If we go back to what Azar wasmentioning, all of these tools
have something that's calledchat mode.
In chat mode, I can actually aska question like, what do you
think we should add next?
Okay.
And then instead of actuallyexecuting and writing code and.
(21:41):
Making the application morebloated.
It just talks to me.
I'm talking to Claude now, butI'm talking to Claude in the
context of my project.
So if it comes up with an ideaor a plan, I can read through
it.
If it looks logical, all I haveto do is click on implement this
plan right here and then now itwill go and execute it.
But at least now we're on thesame page.
(22:02):
I noticed that a lot of timesthat I will spend 70% of my time
even in lovable right here.
It'll be called uh, chat righthere.
If you click on chat mode, I'llspend 70% of my time here.
I'm only 30% of my time actuallywriting the code.
'cause I'll spend five 10messages to make sure we are
100% on the same page before wego and write the thing or do the
(22:22):
thing.
And then rep has, uh, the samething as well.
And then rep's just really goodat dealing with feedback in
general.
All of these have also, uh,visual editors.
So if I go to here, um, I canclick on this button right here
and then I can go select aspecific element.
And you'll notice on the lefthand side, it's now referring to
(22:43):
this exact element.
So instead of.
Praying that I send the promptand it knows exactly what I'm
talking about.
I can say, no, this one, this iswhat I wanna change.
Same thing with, uh, lovable.
You can click on visual edit andI can say, this button, this
button's ugly.
And I don't even have to send aprompt.
Sometimes if I wanna just, um,change the size, I could just do
(23:03):
it here without a prompt.
And that's how a lot of theseplatforms differ ever so
slightly.
You can't do this exactfunctionality in Bolt, but in
lovable, you're allowed to bespecific.
I can even, by the way, Google aspecific font name, paste it
here, click save, and it willmake the change in the code,
which is very powerful becauseit saves you from having to
guess how to tell it the thingin natural language.
(23:25):
And then in terms of rept, Idon't think there's, uh, is
there a prompt helper?
I don't recall that, um, myself.
But at the very beginning, youcan get some help going back and
forth on flushing out theprompt.
And one thing I just wanna showyou on that rept, again, small
nuance is you'll notice itactually takes physical
(23:46):
screenshots of the webapplication while it's
processing, whereas the otherones are solely depending on
code.
So where I say it's more robustis it's not just looking at its
code, it also is takingscreenshots.
Like I would send to OpenAI orClaude and saying, analyze and
see what's wrong.
Let's see, before the usercatches it, let's catch it
ourselves.
(24:07):
And what I mentioned before interms of being more robust, you
can notice here if I go on.
Uh, database deployments, allthis stuff.
I go to all tools.
All of these tools are pre-builtin rept.
So if you want a database, ifyou want the ability to use
Google Sign in very easily, allthat's built into what's called
rept authentication.
They've kind of made one wholeecosystem of tools.
(24:29):
If I go back to Bolt andLovable, if I want to enable
something like Google Sign in,I'm gonna have to go to
something like integrationshere.
And then I'm gonna have tocreate what's called a superb
base account, which is like adatabase account.
And I'm gonna have to connect tothat superb base account, then
give it some credentials, thenit can start to build functions
that allow me to doauthentication.
So that's just one smallmicrocosm of using these tools,
(24:52):
the differences.
These can get me a prototypeimmediately without me having to
think about anything in depth.
Rept will allow me to go muchdeeper.
And what another cool thingabout Rept is they have a mobile
app.
So if you're on, uh, the bus, ifyou're on a plane, if you're
waiting in the airport, I couldhave built this on a mobile app.
And just prototyped on the go.
(25:12):
Vibe coded while waiting for mydelayed flight, and then I'd be
on my way.
So that's just one more idea.
And the last thing is on theRepla side, I've used Repla as a
developer for years to deployapplications on a server.
The fact that it's all built in,again, in that one ecosystem,
that means that they're reallygood at cloud, they have secure
cloud, they're already trustedby hundreds of organizations,
(25:35):
and now they're just marryingtheir whole ecosystem to cloud.
So that's where I'd give Rep theedge on robustness.
Now, being mindful of time, um,I'll see how much time I have
left.
I'll just show you the hack andperplexity.
If you're looking at all thisand you're like, I'm still stuck
on how you wrote that prompt,I'd rather unblock that for you
before we end up.
So in perplexity, and you canuse what, whatever language
(25:58):
model you wish that's connectedto or hooked up to, uh, a web
search.
So you can use, I usually usedeep research, but just for
time, I'm just gonna use, uh, avery quick search.
And what I'll do is thefollowing, and I'll just dictate
in natural language so you cansee how poor I am as a prompt
engineer.
So I want you to act as a promptengineer with a specialty in
(26:19):
product design as well.
And I want you to look up allthe documentation, watch YouTube
videos on the platform's,lovable rept and Bolt.
And I want you to come up with auniversal prompt that I can
create for my e-commercebusiness that would work for all
of them.
And build a beautifulminimalistic e-commerce app that
lets me manage my Shopify ordersprogrammatically in just using
(26:43):
my natural language, typing insome form, some form of text
box.
And I want you to output thatprompt in an un unformatted code
block in mark down so I can justeasily copy paste it.
So in this case, in plainEnglish, I just said go and
watch YouTube videos, go andread documentation on all these
apps, and it will know that it'susing Claude behind the scenes.
(27:04):
So in a way, it's also gonnaresearch that.
And when I send this, you'regonna see what's gonna do.
It's gonna go through lovabledocumentation, it's gonna go
through, uh, rapiddocumentation.
It watched the ultimate guideon, oh, if I go back, my bad, it
watched the Ultimate Guide onusing, uh, repli and Bolt.
So let me just try that again.
(27:27):
Worst case I'll just do, therewe go.
So lemme go back to what it did.
Go to sources.
So I went to Reddit, uh, went toYouTube videos and watched, uh,
I watched videos on.
Shopify integrations.
It went through gpt, it wentthrough GitHub, which is
basically a repository fordifferent apps.
(27:47):
And then it came up with, afterdoing all that this prompt, and
you can see here, this exposesmy lack of genius'cause it did
all the hard work.
Build a beautiful minimalistice-commerce web app that connects
to Shopify store and allows meto manage orders
programmatically.
And then it gives a list of allthe different requirements,
clean, modern, minimalistic UIwith responsive design, and
(28:09):
basically a laundry list of allthe things we want in this app.
And then you could just copythis and throw it into whatever
web app builder you're lookingfor, and you'd be on your way.
And the cool thing is, one extrapro tip I'll give you is if you
get stuck, what Isar mentionedis amazing, which is trying to
pivot or going to chat mode todiscuss.
But one other thing you could dois, I like to actually, since
(28:32):
you have access to the code,even if you can't read the code,
I like to ask it, okay, where doyou, which file has this problem
you're talking about?
And it'll tell me this AIanalytics tsx, and I'll just go
copy, paste this whole thing,give it to maybe Gemini or
Claude, uh, a brand new versionof Claude with virgin eyes on
the matter.
And maybe they'll say, you knowwhat?
It's missing this.
And you're like, aha, this isthe problem.
(28:54):
So the fact that you have accessto code, even if you don't know
how to read code, you can use itto help you get back on track if
things go off the chain.
So that's on the prompting side.
I'm gonna pause
Isar Meitis (29:05):
This was nothing short of mind blowing, like the
chop people, like this isphenomenal.
This is incredible.
I can't believe I did not knowabout this.
Good.
And I did a littlequestionnaire, like a little
survey, who knows about vibecoding.
And most people didn't have aclue or kind of like knew about
it, but didn't.
So I'm, I'm sure this is gonnatake a lot of sleeping hours
from people to play over thenext few days, uh, which was the
(29:26):
whole point.
Uh, mark, again, this wasnothing short of exceptional,
uh, as expected.
So I set the bar high, but youjumped like 10 feet above the
bar.
Thank you.
So thank you so much.
I appreciate it.
Uh, if people wanna work withyou, follow, you know, you
connect with you, what are thebest ways to do that?
Mark Kashef (29:43):
Sure.
Yeah.
So, uh, YouTube is just MarkCash if, uh, if you wanna follow
my Mad Scientist experiments.
Um, otherwise, uh, promptadvisors.com if you wanna engage
on a project.
I just threw it in the chatthere.
And yeah, I mean, YouTube willgive you 90% of my life, uh, and
what I do and what I'm workingon.
So that'll be the best way togo.
This is the end of part two ofthe Ultimate AI Showdown.
(30:06):
But don't worry, there's partthree and four of this
incredible series of theshowdown.
So next week we will release thenext episode.
That will be episode 2 0 4 ofthis podcast in which Rory
Flynn, one of the top AI imagegeneration professionals in the
world today, will compare AIimage generation tools and how
(30:27):
to pick them and how to use themin the most effective way.
And until then, have an amazingrest of your week.
Popular Podcasts
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.