YT YouTube2Text

YouTube2Text → Transcripts → I Built a $1M/y SaaS with Claude Code, Here's How

I Built a $1M/y SaaS with Claude Code, Here's How — Transcript

by Nick Saraev · 5,510 words · 787 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hey, so we just hit a million dollars in
0:01ARR with our SaaS product, which we use
0:03Cloud Code to build. And I know a lot of
0:04people here are probably interested in
0:05using Cloud Code either independently or
0:07within an organization to put together
0:09some sort of SaaS app and then take it
0:10to market. So, I figured in this video
0:12I'd run you through basically everything
0:13that we did in order to get to where we
0:14wanted to and uh also share all the
0:16learnings along the way. So, what is the
0:18SaaS? It's called Clarivo. It is
0:19essentially an AI-enabled power dialer.
0:23And just to unpack those words, what
0:24this does is it allows us to make more
0:26calls per unit time and then have more
0:28of those calls picked up on the back
0:30end. And that works really well and is
0:32very powerful if you're in an industry
0:33that is traditionally pretty call-based.
0:35So, either you have some sort of funnel
0:37where you have inbound leads and you
0:38need to call them very quickly and en
0:40masse or uh you know, you're doing like
0:42traditional cold calling or outbound
0:43calling to try and acquire clients whom
0:45you don't have preexisting relationships
0:47with. And so, anytime you're starting
0:48any business, whether it's a SaaS,
0:49e-com, you know, service company,
0:51whatever, you need to have a very
0:52clearly defined problem that you're
0:53trying to solve. And so, I'm going to
0:54run you guys through exactly how we
0:55picked this problem later, but
0:57essentially at a high level, we picked
0:58this because it has very high a lifetime
1:01meaning that a single client that we get
1:02on our service will pay us a lot of
1:04money over the course of the next few
1:05years. Uh it's very low churn because
1:07once we install it into a company, it's
1:09very unlikely that they're going to just
1:10bow out. Their whole infrastructure
1:12depends on us. And then, it's also very
1:14straightforward and easy to do in a
1:16market that didn't have a lot of uh
1:17other entrants. Cloud Code helped us
1:19come up with every single way and I'll
1:20run you through a quick step-by-step on
1:22how to do it in a second. But just so
1:23that we're all clear, essentially, you
1:25know, an industry competitor might make
1:27100 calls an hour. These might be
1:28outbound calls to try and close some
1:30deals to strangers they've never met
1:31before or could be inbound calls calling
1:33a list of people that opted into some
1:35offer. From those 100 calls, because of
1:37dial times, connect times, people aren't
1:38present, people aren't picking up the
1:40phone from numbers they don't recognize,
1:42maybe only 40% of those will actually
1:43pick up. So, if you think about it right
1:45off the bat, a salesperson's making 100
1:46calls, only 40 people are picking up,
1:48there's sort of a 2.5x drop-off right
1:50there. And so, if you just do the math,
1:51you have a salesperson working 8 hours a
1:52day, they're capable of getting 40
1:54pickups an hour, it's like how many
1:55actual conversations are you having?
1:57Let's say they do that every day for a
1:58month, maybe they make 10K a month. What
2:00Clara does is allows you to make more
2:01calls in the front end. So now we're
2:02capable of doing what say 200 calls an
2:04hour instead, and then it also increases
2:06the fraction of people that pick up
2:07because the calls are more recognizable.
2:09We use a couple of cool cloud code-based
2:11algorithms to like dial multiple numbers
2:13simultaneously and then also double and
2:15triple dial if needed. And then
2:17basically at the end result is you just
2:18make more money. So in our case we have
2:20more calls, we have a higher pick up
2:21rate and so there's significantly more
2:22people that are actually on the phone.
2:24And uh right now we're capable of
2:25generating, you know, somewhere between
2:2750 to 80% improvements to the companies
2:29that we work with. We took a pretty
2:30sizable business from Texas from
2:32somewhere between 3 to 5 million dollars
2:34per month, uh which is almost uh you
2:35know, double their their revenue. And so
2:37this is the sort of value proposition
2:39that Net Clever has. So how do you
2:40actually use Claude here? Well, I should
2:41note that we didn't actually know how to
2:42solve this problem when we started. Uh
2:44we actually had Claude walk us through
2:46every possible way that it knew of to
2:48improve pick up rates and increase the
2:50total number of calls we could make per
2:51unit time. And uh most of the ideas were
2:54absolute trash. But after mining Claude
2:56for 200, 300 ideas, a couple of them
2:59were actually pretty good. And so the
3:00process, if you're interested, is we
3:01literally said, "Hey, we're building
3:03insert product here. You know, it is in
3:06our case an AI-powered dialer for local
3:08service businesses like HVAC, plumbing,
3:10roofing, et cetera. Our core metric to
3:12optimize is call pick up rate, which is
3:15defined as the percentage of dialed
3:16numbers that result in a live human
3:18answering within say 10 seconds." So
3:20here we have the current baseline, we
3:22have the industry ceiling, and then we
3:23even had our target. So what I told it
3:25to do was spawn 10 parallel sub agents.
3:27Each one should propose 10 distinct
3:28mechanisms we can use to increase pick
3:30up rate. I also want you to diverge each
3:32of these wildly. So do algorithmic,
3:34behavioral, infrastructural, regulatory,
3:35psychological, time-based,
3:36identity-based mechanisms. Don't
3:38self-censor for any feasibility. I'm
3:40going to do all this later. And so after
3:42it comes up with all of these ideas, and
3:43it's going to come up with a lot of
3:44ideas as I mentioned, what we're going
3:45to do is we're just going to take them
3:47and then verify, "Okay, is this a like a
3:48total BS idea or is it like an okay
3:50idea?" And so here we go. We now have a
3:52variety of results. A lot of them are
3:54hard duplicates, as well. But, just
3:55going top to bottom, the first is a
3:57temporal propensity model, which is
3:59basically using AI to determine
4:02an optimal call window, aka when to call
4:04people. So, this is legitimately
4:05something that we do at Clara. We have
4:07optimal call windows based off of
4:09average pickup times per, you know, time
4:11of day, essentially. But, at the same
4:13time, some of these other ideas are
4:14total BS. So, weather times pickup
4:16regression, you know, can we run a
4:17regression, which is a statistical
4:19analysis, on historical pickup rates
4:21versus hyper-local weather. Uh you know,
4:23just off the top of my my head, that's
4:25probably not going to be anywhere near
4:26as valuable as doing some sort of like
4:27call-based on time, let's say. And so,
4:30you're going to get tons of ideas like
4:31these, and yeah, the majority of them
4:32are going to be junk. But, you're going
4:34to find a couple that work. And so, in
4:35our case, this is literally what we did.
4:37We ideated over all of the possible ways
4:39to improve something. After you're done
4:40with that, we shortlist one of these
4:42ideas. And so, in our case, predictive
4:44pacing was actually a pretty well-known
4:45idea. It's not something we invented.
4:47Clara could definitely didn't invent.
4:49But, you know, it's an idea that we
4:50wanted to explore and see, okay, what
4:51sort of alpha would there be if, you
4:53know, rather than just call one person,
4:54they actually call multiple people
4:55simultaneously. Essentially, because the
4:58amount of time it takes to dial somebody
4:59is very fixed. Like, if you think about
5:01it, you enter your phone number in, and
5:03then you stand on the line, it goes
5:04din-din-din, din-din-din.
5:06What that means is if the person doesn't
5:07pick up, you just wasted all that time
5:09as a salesperson. So, if your your goal
5:11is optimally to be more efficient, the
5:13actual optimal play is not just to call
5:15one person and have the din-din-din,
5:17din-din-din. It's actually to call two
5:19people and have the din-din-din,
5:21din-din-din. Because if one of those
5:22people doesn't pick up, well, no
5:24problem, you've taken the total amount
5:26of time it would have made to make that
5:27dial, and then you connected with this
5:28person anyway.
5:29And so, this isn't just limited to two
5:30people. We actually use an algorithmic
5:32model that specifically imbues like
5:35offsets into our multiple call thing
5:38that is proven, and we've seen it in our
5:40data, to call and get picked up by the
5:43optimal amount of people per unit time.
5:45Do some people pick up at the same time,
5:47And then that results in kind of an up
5:48weird awkward situation? Yeah, but we
5:50also have a built-in call routing so
5:52that if, you know, we make multiple
5:53dials here, one of them doesn't get
5:55picked up. It actually goes to an agent
5:56that might actually be available. So,
5:58it's a queuing system which, you know,
5:59Claude code obviously helped us build.
6:00But it all started like right here. This
6:02is the exact same approach that we use
6:03in order to figure all that out. And so,
6:06once you have this simulation harness,
6:07you know, you feed it in a bunch of data
6:08on historical call times, which we
6:10accumulated through our own businesses
6:11and then businesses of other people. Now
6:13we have something we can run stats on.
6:15And we can figure out, okay, what's the
6:16optimal offset for this, you know, batch
6:18of 50,000 calls, let's say, in order to
6:20determine, you know, what our what our
6:22offset needs to be. Once you're done
6:23with that, you feed it in another prompt
6:25that says, "Hey, I want you to now
6:26implement this predictive pacing
6:28simulation from the spec above. Here is
6:30some historical data. I want you to
6:31optimize for these things using, in this
6:33case, Bayesian optimization." Obviously,
6:35this is going to depend on the specific
6:36problem you're trying to solve. But what
6:37I'm trying to say is, we just had Claude
6:39code, you know, figure out
6:41the ways to improve what we wanted to
6:43improve, and then actually implement
6:44that the simulated environment. Finally,
6:46you build the thing, which in our case
6:47was this predictive pacer, and then you
6:49roll it out in real businesses. And, you
6:51know, I think this is probably the thing
6:52that's going to trip up a lot of people
6:54because they don't have real
6:55pre-existing businesses that are
6:56currently live right now that they can
6:58test things out on. And that's why data
7:00is ultimately the quite the moat. If you
7:01have the data and then you also have the
7:02means to deploy something and do, you
7:04know, parallel testing, you can you can
7:06usually get through this sort of thing
7:07way faster. Okay, and that takes me to
7:08this general sort of loop. In order to
7:11do this sort of thing effectively, what
7:12you always start with is you start by
7:13defining a problem. Of course, you're
7:15going to have Claude code help you do
7:16the idea mining and the problem
7:18definitions. That's okay. Um but in our
7:20case, we just knew this was a problem
7:22that a lot of people were willing to pay
7:23a fair amount of money for. Then you
7:24say, "Hey, Claude, how can we solve this
7:26problem? I want you to enumerate, aka
7:28list, all possible solutions to, you
7:31know, the problem of let's say call
7:32pickup rates."
7:34Then what you do after that is you apply
7:35your little human brain, your little
7:37sponge, and you say, "Okay, which one of
7:39these are total and which one
7:40of these are actually somewhat
7:41feasible?" And so, in our case we had a
7:43short list of maybe five or six out of
7:44several hundred that were actually
7:46feasible. And you know, over time we're
7:48going through the the the the rest of
7:49them as well just to verify if this is
7:51something that can actually add some
7:52alpha, some delta to, you know, call
7:54pickup rates. But the vast majority of
7:56the time it's one of those things that
7:56you'll just read and you'll be like,
7:57"Okay, yeah, this is obviously the one."
7:59Once we're done, we design some
8:00simulations with Claude code, usually
8:03based off some form of historical data,
8:04and then we run a statistical model, in
8:06our case the predicted pacing algorithm,
8:08in order to actually have that perform
8:09better. Then we iterate a simulator,
8:12okay, we have Claude code just like
8:13change the the the parameters of our
8:15models so that it gets better and better
8:16and better. And then finally we have
8:18like a real life stress test where we
8:19actually roll it out. And I mean, it can
8:21fail at any step along these lines here.
8:23We've had a variety of, you know, pretty
8:25cracked out approaches that we thought
8:26were going to work really well in the
8:28sim because we saw better improvements
8:29in our stats, but then when we rolled
8:31them out to real life we're like, "Oh my
8:32god, wait a second, there's actually
8:33this third variable here that confounds
8:36and kind of ruins everything." So, you
8:38know, it's not easy. If it was easy,
8:39you'd have everybody doing it, and if
8:40everybody was doing it, nobody would be
8:41making any money, but this is how we
8:44ideated on the set of core features of
8:47Clarvo that ultimately ended up making
8:48us a fair amount of money. But the
8:49pricing is 250 bucks a month, which is
8:51not like a scientifically determined
8:53price. We started by pricing close to
8:55like 100 bucks a month, and we figured
8:56out that people were willing to pay for
8:57it, so then we increased the price,
8:59figured out people were still willing to
9:00pay for it, increased the price. Uh you
9:02know, I think people that are trying to
9:03use these big statistical pricing models
9:05or have AI like determine what the best
9:07price is are usually just wrong. The
9:08much easier and simpler way is just like
9:10pick a price and then sell it to a bunch
9:12of people, and if it's easy and they say
9:14yes, then just keep increasing the price
9:15until eventually it gets hard. In
9:16general with SaaS companies there's a
9:18big spectrum of possible prices. Um if
9:20this is our spectrum here, at the very
9:22left is basically what is called um low
9:25touch. Low touch SaaS businesses,
9:27generally speaking, are like self-serve.
9:29What that means is it's like a
9:31self-guided onboarding. There's like
9:32maybe a video from the founder. You pay
9:33like 5, 10, 15, 20 bucks a month, and
9:36then everything's is kind of done for
9:37you. And, you know, these can be really
9:39good, but my head cannon, my my personal
9:41belief is in an era where Claude code
9:44and other AI agents are capable of
9:45whipping up basically any SaaS,
9:47you know, like you got to ask yourself
9:48at a certain point any business owner
9:50will be willing or able to make the
9:52trade-off of just paying money for
9:53tokens to actually just rebuild the
9:54whole thing. So, rather than us sort of
9:57going really cheap and really small and
9:59solving a tiny problem, we decided to go
10:01the exact opposite direction, um and we
10:02ended up solving a pretty big problem
10:04kind of close to the enterprise
10:06uh with what's called a high-touch SaaS.
10:08So, Clervo sits sort of right around
10:09here, and typically we don't just sell
10:11individual licenses. It's not like uh
10:13you know, a single user can't sign up if
10:14they want to. But, in general, we work
10:16with companies and then roll this out to
10:17a pre-created team of people that are
10:19doing calling. So, for instance, you
10:21know, we sign a 100-seat deal at $250 a
10:24month, well, if you think about it kind
10:26of mathematically, that's $25,000 MRR,
10:28which is 300k ARR. So, that's more or
10:30less what we've done. We've closed a
10:30handful of deals with sort of like
10:32mid-market uh uh to maybe larger uh
10:34businesses that operate in a variety of
10:36very call-heavy industries. Only takes a
10:38couple of those people to say yes to
10:40roll it out to their team and then make
10:41a fair amount of money. On the pricing
10:42point, my big take on a lot of this is
10:44nowadays anybody can build virtually
10:46anything. If you look at the total
10:48number of commits over time, okay, they
10:51are skyrocketing, and that's because AI
10:52is doing the vast majority of the
10:53intellectual heavy lifting now. So, it's
10:56no longer can you build insert software
10:58product here, cuz we can all build it.
11:00The the the bottleneck, the mode, like
11:02the value that you have is what should
11:04you build, and you know, essentially how
11:06should you price. So, what you quickly
11:07realize is that the vast majority of
11:09frameworks are total fluff. Now, we
11:11tried a lot of agent frameworks for
11:13Clervo. We tried Hermes, we tried Open
11:16Claw, we tried a bunch of these context
11:18libraries, uh basically made like vector
11:20DBs of your memory. We probably tried
11:22like 50 different approaches. And I can
11:24definitively say for the purposes of
11:27creating a software product that later
11:29generates revenue, basically every
11:31additional framework you use is
11:33inversely correlated with the amount of
11:34money you make. Cuz every time you jump
11:36on a different framework, you are not
11:38only distracting yourself and pulling
11:40away from like the thing that you're
11:41trying to build. Uh typically, you have
11:44like regression within whatever the code
11:45base is because now the prompt is being
11:48understood or mediated a little bit
11:49differently than it was before. For
11:51those of you guys that don't know,
11:51regression is just where, you know, you
11:53had an approach previously that worked
11:54really well. Let's say some vanilla
11:56thing with like a small little cloud and
11:57MD. Uh but because now you're you're
11:58doing it through a different framework,
12:00like a lot of the assumptions and
12:01memories and and and things that the
12:02model used to know about your code base
12:03no longer works. Uh which is quite
12:05unfortunate. So, you know, rather than
12:07jump around a lot and try and like
12:09uh aim for that 100% quality uh or like
12:12a 100% score uh IQ test of the model, I
12:16would rather have the model work 90% as
12:19well of like its total potential, let's
12:21say, but I'd have it work consistently
12:23and be the same every single time. The
12:24real value that I think not a lot of
12:26people understand is that, you know,
12:29the intelligence comes from the model
12:30itself these days. It does not come from
12:33the shiny framework that wraps around
12:34it.
12:35You slapping on some new framework to,
12:38you know, the way that your your team is
12:39building on cloud code is kind of like
12:41uh people that put a fuzzy cover on
12:42their steering wheel and then they
12:44pretend that that's the reason why their
12:45car works so good. Like obviously,
12:46that's not the reason why your car works
12:48so good. Your car works good because it
12:49has wheels, it has an engine, it has a
12:51chassis, and so on and so forth. It's
12:52the craftsmanship of the person that
12:54built all of that. Uh but, you know, you
12:57cuz you want to be all special and and
12:58new and stuff like that, uh put put your
13:01little fuzzy steering wheel on and then
13:02go like, "Oh yeah, this is way better."
13:04It does not a genuine improvement.
13:05That's just your subjective improvement.
13:07And so, I think human beings, we want to
13:08take credit for everything even if it's
13:09not necessarily ours. And so, we do the
13:11uh virtual equivalent of slapping on a
13:13bunch of like fancy fuzzy covers, aka
13:15all these Hermes agents and and and open
13:17claw tools and stuff like that. Uh when
13:20in reality, the thing that's making the
13:21car go is is the is the base model. And
13:23so, that's why if you guys look deep
13:25into the people that actually like
13:26created a lot of these technologies.
13:28Like Boris Cherny for instance, who's
13:30one of the creators of Claude code.
13:31These people typically have like nothing
13:33of substance in their Claude.md files.
13:36They have nothing in their system
13:38prompts. They're literally just using
13:40the vanilla intellect of the model. And
13:42the vanilla intellect of the model is
13:43usually, for all intents and purposes,
13:45pretty damn good. You'll only get
13:46marginal improvements applying one of
13:48these frameworks. And what you find is,
13:49you know, Claude code's getting so good
13:51so quickly nowadays that if there is a
13:53marginal improvement that gives you like
13:54a 5% a plus ROI, the next generation of
13:57the tool, maybe like three or four days
13:58later, will actually already include
14:00that. Either hardcoded into its system
14:02prompt or maybe actually just part of
14:03like the training of the model. The
14:05second thing is to pick problems that
14:06actually pay. And so, the idea is, okay,
14:09you can build more or less anything. And
14:12so, this left-hand side of the Venn
14:13diagram are all of the things that you
14:15could build, and every green dot is a
14:16thing that you've decided to build.
14:18You're not going to make any money.
14:20What you want to do, okay, is find that
14:22small little slice of the Venn diagram
14:25on the right-hand side that people will
14:26actually pay for. So, these are things
14:28like red-hot problems. They're
14:30industries and niches that have big
14:31budgets. It's people with a pre-existing
14:33pain. And then what you want to do is
14:34you just want to focus all your time
14:35over here.
14:36And so, with Clarabridge, that's what we
14:37did. We saw just how inefficient a lot
14:39of sales people were and how literally
14:41just getting on a power dialer, cuz this
14:44isn't a new idea to power dialer,
14:46but we saw like the difference between
14:47not having a power dialer and then
14:48having a power dialer was like 3x
14:51effectiveness. Then we're like, "Okay,
14:52what if we could just make actual
14:53pre-existing power dialers even better?"
14:55And we're like, "Okay, if we can
14:56generate even like a 2x effectiveness,
14:57we'll be able to to take a large portion
14:59of the value that we provide for
15:00companies."
15:01And so, that's that's the most That's
15:03sort of where you need to sit if you
15:04really want to crush it in SaaS
15:05nowadays. And so, everything exists on
15:06this problem-value spectrum. You know,
15:09on the left-hand side, you have a bunch
15:10of lukewarm problems. These are things
15:11that are nice to have, but they're not
15:13necessary to have.
15:15And this is unfortunately where probably
15:17like 90% of people spend their time.
15:19And I'd built, you know, a a of demos
15:21showing you how you could put together
15:22to-do apps and simple browser extensions
15:25and simple productivity tools and so on
15:26and so forth. But, the harsh reality is,
15:29you know, if the problem isn't big
15:30enough to justify somebody
15:32uh you know, choosing your SaaS over
15:34like building it all themselves because
15:36as mentioned, software's now quite easy
15:37to build. Anybody can just
15:39uh convert tokens into product just at
15:42some sort of exchange rate. You know, if
15:44it's not a big enough problem, people
15:45are just going to do that and the
15:47longevity of your SaaS is going to be
15:48significantly smaller than if picked up
15:50a red-hot burning problem.
15:51So, in our case, we picked uh something
15:53that is currently costing organizations
15:54millions of dollars a year. They'll pay
15:56anything to fix their to fix their
15:57pick-up rates or improve it if they know
15:59that it's an option. And uh so, this is
16:01more or less what what we've done.
16:03So, instead of solving a, you know,
16:06I don't know, uh marketing for dog
16:07walkers where it's like the average dog
16:09walker probably makes like a thousand
16:10bucks a month or something like that.
16:12You know, solve a core need for a large,
16:16usually mid-market and up style company.
16:18Uh people that actually have budgets and
16:20typically also have many seats that
16:21would need to subscribe to these budgets
16:23in order to solve said problem. So, as
16:24mentioned, uh we implemented this on one
16:26of our eight-figure clients and it says
16:27uh a year here, but it's it's literally
16:28a month. I think the AI just didn't
16:30believe me when I said it was
16:31legitimately a month. Uh and we took
16:33them basically uh we increased their
16:35their monthly revenue by 66%.
16:38And so, if you think about it, like what
16:39did we do? The delta there is two
16:41million a year in revenue.
16:42And typically the way that it works is
16:43if you solve a problem, okay, you are uh
16:47I don't want to say entitled to, but you
16:48can typically negotiate or ask for
16:50somewhere between to 15% of the total
16:53amount that you are providing. And so,
16:55we provide two million dollars a month
16:57to this company, 24 million a year. It
16:59is not unreasonable for us to ask for or
17:02at least be in a position where we can
17:03negotiate a tenth of that or 2.4 million
17:06dollars a year. And so, this is the sort
17:07of problem that ultimately you want to
17:08solve. You know, you want to find people
17:10that have the means to pay for uh this
17:13red-hot burning thing. But, you also
17:14need the problem itself to be quite
17:15valuable. If it's not, probability of
17:17you, you know, getting anywhere with
17:18that is quite low. Another hack is to
17:20pick an industry or a SaaS type that
17:23requires some form of human
17:25implementation or like human onboarding.
17:28What I mean by this is, you know, if
17:30everything that you do is entirely
17:32digital, then it is pretty reasonable to
17:36expect that in the next couple of years
17:38AI will be able to do it better than
17:39your team.
17:40And so, you know, your onboarding your
17:42tool into the company is nowhere near as
17:44valuable just like, "Hey Claude, can you
17:45do it all for me?" Claude will be able
17:47to do that for most things fairly
17:48shortly.
17:49But the one thing that AI can't
17:50currently do is it can't upend like
17:52regulation. You know, if you need, in
17:55our case, a bunch of numbers applied
17:57for, you you need A2P registration. And
18:00that's just like a fixed thing, that's
18:01like a law, that's like a regulation.
18:03You can't just say, "Claude, screw screw
18:05the A2P registration, get me 5 million
18:07phone numbers." Because both for moral,
18:09ethical, and programmed-in reasons,
18:10Claude will will say no. But also,
18:13there's just no way to get the number
18:14unless you actually go through this like
18:15pretty bureaucratic process.
18:17And so, what I mean by that is like in a
18:19future where there's no moat to to
18:21doing, you need to look for natural
18:23moats that are created by regulatory
18:24environments. In our case, things like
18:26numbers, for instance. Another great
18:28example of that is like in healthcare.
18:31Everybody complains about HIPAA all the
18:32time, myself included, because, you
18:34know, it's it's quite the blocker to US
18:35healthcare implementing any sort of or
18:37building any sort of like cool
18:38transcription service. It will require
18:40you to like fastidiously adhere to HIPAA
18:42principles, and that can slow you down a
18:43lot. You need to anonymize your data,
18:44and so on and so forth. But viewed
18:46another way, that's actually a major
18:47opportunity in like an AGI world because
18:50that's the only thing that is currently
18:51stopping us from being able to, you
18:52know, do things.
18:54Legitimately having some sort of like
18:55certification, let's say, or some sort
18:57of board approval of rolling something
18:59out. And so, as a as a company, as a
19:01SaaS, if you could build some form of
19:03human implementation, human onboarding,
19:06you know, a human responsible for
19:08maintaining the relationship between you
19:09and the advisory board that needs to to
19:11rubber stamp the thing, then you'll go
19:12way further.
19:14And so in our case, you know, we have a
19:15bunch of relationships and connections
19:16with people that know how to do these
19:18things and facilitate them a lot faster.
19:19And that that's one of the moats that I
19:21think will actually carry us forward in
19:22the next couple of years as opposed to,
19:24you know, big AI just pulverizing the
19:26vast majority of these low-touch,
19:27low-ticket SaaS's. Finally, one last tip
19:29is to make whatever your code base is
19:31model agnostic. So I know the whole
19:33point of this video is that we built it
19:34with Claude code. Um, I would say that's
19:36like 90% true. In addition to Claude
19:38code, we obviously tried a variety of
19:39other models. We tried a deep seek to
19:41arbitrage token costs on like constant
19:43long-running 24/7 uh uh like
19:45restructuring and refactoring and stuff
19:47like that. Constant like bug fixes and
19:49and and so on. And uh that worked okay.
19:52We tried Codex a number of times. Um,
19:54our team is increasingly using Codex
19:55just as we've run into like some um
19:58token issues. And the the tokenomics
19:59essentially are the main thing that that
20:00are holding us back from going all in on
20:02Claude code 24/7.
20:04But also, I think uh over the course of
20:05the next few months, you'll probably see
20:06fluctuations in the quality of each of
20:08these models and the availability of
20:10each of these models because uh you
20:11know, like the major AI companies are
20:13starting to get very compute restrained
20:15because everybody on planet Earth wants
20:16one of these models now. They're
20:17realizing how economically effective
20:19they are. And so you need to be able to
20:20just like hot swap your code base at
20:22will from let's say like a Claude code
20:24base project to like a Codex project.
20:26And this isn't really that hard at all.
20:27It's just like a a little bit of
20:28friction that I think slows people down.
20:30But uh Clarifai, we just made our our
20:31code base totally model agnostic. And
20:33what that means is like, you know how
20:34Claude code has like a skills spec and
20:36it expects a Claude.md and so on and so
20:38forth. Uh we just have like, you know,
20:40an agents.md. We have the agents skills
20:42spec. We have uh you know, the thing
20:44things for Gemini, Gemini.md. Just in
20:47case at any point in time we want to hop
20:48over or maybe employ a different model
20:51to see if maybe that model can solve a
20:52problem that we're struggling with. Um
20:54you know, it's just like that. And
20:55anybody in our team has the ability to
20:57to do so. And so the real actionable tip
20:58here is just duplicate everything and
21:00then probably have Claude go through the
21:02specs of each of these models and just
21:03like make sure to prepare the workspace
21:05so that at any point in time you have
21:07the ability to, you know, instant
21:08preload all of your system prompts and
21:09so on. And um MCP specs and then skill
21:13specs are actually currently understood
21:15differently from like Claude versus
21:17other uh platforms. Like not all
21:18platforms do the YAML front matter
21:20tuning for instance where they'll only
21:22preload uh like the name and the
21:23description of the skill. Um some of
21:25them will actually load the entire
21:26thing. These are just slight little
21:27model differences that you can optimize
21:29around that will uh you know, allow you
21:31and other people within your company to
21:32operate much faster. Okay, I hope you
21:34guys like the video. Had a lot of fun
21:36putting it together for you. Um as
21:37mentioned, obligatory pitch for the SaaS
21:39company. That was sort of a case study
21:41for this whole video, Clearvo. If you
21:42guys want to improve your pickup rates,
21:43definitely check that out um because,
21:45you know, we're experimenting with with
21:46pricing and a variety of different
21:47things. Um you know, I'll I'll add a
21:49link to the top of the description so
21:50you guys can give it a quick click and
21:52go through if you like. More generally,
21:53if you guys want to learn how to
21:54monetize AI automation and SaaS apps in
21:57this way, definitely check out Maker
21:58School. It's my 90-day accountability
22:00program where we'll guarantee you that
22:02you get your first customer for an AI or
22:04automation-related service within that
22:05time period or I give you your money
22:07back. And if you guys have any ideas for
22:08future videos or if you guys want me to
22:10record something on specific topic that
22:12is trending, interesting, or just sort
22:14of stream of consciousness, uh feel free
22:15to let me know. I take most of my video
22:17ideas at this point from people in the
22:18comments, okay? Thank you again for
22:20watching and I'll catch all y'all in the
22:21next video.

About this transcript

This page contains the full transcript of I Built a $1M/y SaaS with Claude Code, Here's How by Nick Saraev, generated from the public captions YouTube serves with the video. The transcript has 5,510 words across 787 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.