11:21Austin Marchese · Tutorial
Don't Start ANY Claude Code Project Until You Watch This
Three rules from YC CEO Garry Tan translated into a six-move AI leadership playbook — and the four questions that kill bad projects before they start.
May 20thThe argument in one line.
The three barriers that once killed most ideas -- cost, domain expertise, and time -- have collapsed simultaneously, so the only useful filters left are whether you can verify the output, who it is actually for, and whether a human stays at both edges of the workflow.
Read if. Skip if.
READ IF YOU ARE…
- You are building something with Claude and keep second-guessing which ideas are worth pursuing.
- You have a backlog of shelved ideas that felt too expensive, too technical, or too slow to build before AI.
- You want a clear rule for when AI-generated output is reliable enough to ship versus when it will fail in production.
- You are trying to narrow your target audience and stop building things that are only sort of valuable to everyone.
SKIP IF…
- You are not actively building anything with Claude -- this is a decision framework for builders, not a general AI introduction.
- You want tactical Claude prompting techniques -- the video stays at the strategy layer.
The full version, fast.
Anthropic ships faster than almost any company in history because they run every build decision through four filters: cost, skill, and time to ship have all collapsed so most shelved ideas are now viable; anything unverifiable when the cost of failure is high should not be released; every build must serve a clearly defined ICP and explicitly exclude an anti-audience; and AI should only own the middle of a task while a human frames the start and judges the end. These filters compound -- each build sharpens the last -- which is how Anthropic grew 80x in Q1 2026 against a 10x internal plan.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:23
01 · Cold open -- premise and shipping velocity
Hook claim: most people build the wrong things. Introduces four rules and shows Anthropic's release cadence as evidence of disciplined prioritization.
01:23 – 02:50
02 · Rule 1: Recalibrate what's possible
Three factors that have collapsed: cost, domain expertise, time. Dario quote on software becoming free. Daniela quote on building a website. Shelf audit as the practical action.
02:50 – 05:15
03 · Rule 2: Kill anything you can't verify
The cost-of-error filter. Anthropic held back Claude Methos because they could not prove consistent safety. Two steps: define verification before building; give Claude a self-check loop. Boris Czerny tweet on feedback loops.
05:15 – 08:45
04 · Rule 3: Know who you're building for (and who you're not)
ICP vs. audience anti-goal. Anthropic ICP = developers. Anti-goal = image/video for creatives. Steve Ballmer clip as comedic illustration. Compounding builds argument. 80x Q1 2026 growth stat.
08:45 – 11:49
05 · Rule 4: Build middle to middle, not end to end
Every task has start / middle / end. End-to-end AI = human absent. Middle-to-middle = human at both edges, AI in center. Autonomous weapons as extreme end-to-end example. Dario's 5% / 95% comparative advantage math.
Lines worth screenshotting.
- The premise that software must be amortized across millions of users is starting to be false -- meaning you should build a lot more, not less.
- AI has made it possible for anyone to reach level-one understanding in almost any domain, eliminating the expertise barrier that killed most ideas before they started.
- Claude Cowork -- one of Anthropic's biggest products -- was built in a week and a half, almost entirely with Claude Opus.
- The verification filter is not whether something works in testing but whether it can be proven to work consistently when the cost of error is high.
- Defining how you will verify an output before you start building is the step most people skip, and it is why AI automations fail in production.
- Giving Claude a feedback loop to verify its own work 2-3x the quality of the final result, according to the creator of Claude Code.
- An audience anti-goal -- explicitly naming who you are NOT building for -- makes every downstream decision faster and removes the temptation to chase scope creep.
- When you commit to one audience, each build compounds on the last: output quality improves, the relationship deepens, and the next project starts from a higher baseline.
- End-to-end AI workflows where AI owns framing, execution, and final judgment reliably produce lower-quality results than workflows where a human is present at the start and end.
- Doing just 5% of a task -- the framing and judgment -- while AI handles the 95% middle makes you roughly 20x more productive through comparative advantage.
- Anthropic refuses autonomous weapon use cases because the cost of error -- a human life -- is too high. The same logic applies to any automation where a wrong output causes real damage.
- Anthropic grew 80x in Q1 2026 against a 10x internal plan, and the driver was a focused ICP (developers) rather than trying to serve everyone simultaneously.
Four filters that decide which AI builds are worth starting.
WHAT TO LEARN
Before starting anything with AI, the question is not whether it is technically possible but whether you can verify it, who it is actually for, and whether a human is still at both edges of the workflow.
- Cost, domain expertise, and time to ship have all collapsed simultaneously, which means most ideas shelved in the past two years are worth reconsidering from scratch.
- The relevant verification question is not whether an output works in testing but whether it works consistently when the cost of a failure is high -- and that test should be defined before building, not after.
- Giving AI a feedback loop to check its own work multiplies output quality by two to three times because the model will catch and correct errors it would otherwise hand off.
- Defining who you are not building for is as important as defining who you are -- without an explicit anti-audience, each new build pulls in a slightly different direction and none of them compound.
- Focus on one audience long enough and each build sharpens the last: output quality improves, the audience relationship deepens, and the next project starts from a higher baseline.
- End-to-end AI workflows where AI owns the framing, execution, and final judgment reliably produce lower-quality results than workflows where a human is present at the start and end.
- Human judgment at 5% of a task creates roughly 20x leverage when the 95% middle is handled by AI cheaply -- the valuable skill is not doing more work but placing better decisions at the edges.
Terms worth knowing.
- ICP (Ideal Customer Profile)
- The specific type of person or organization a product is designed for. Defining it tightly means every feature decision has a clear test: does this help that person?
- Audience anti-goal
- The customer you are explicitly not building for. Naming this is as important as naming your ICP because it prevents scope creep and keeps each build compounding on the previous one.
- Middle to middle
- A workflow model where a human frames the task at the start and reviews the output at the end, with AI handling only the execution in the middle. Contrasts with end-to-end, where AI owns all three stages.
- Cost of error
- The real-world consequence if an AI-generated output is wrong. When this is high, verification before release is non-negotiable and should be defined before building begins.
- Shelf audit
- A personal review of every idea parked because it once seemed too expensive, too technical, or too slow -- re-evaluated through the lens of what AI has made newly viable.
- Comparative advantage
- The economic principle that even a small contribution to a task creates disproportionate leverage when the rest is handled more cheaply by another party. Applied here to human judgment at the edges of an AI workflow.
Things they pointed at.
Lines you could clip.
01:05
“Software is gonna become cheap, maybe essentially free. The premise that you need to amortize a piece of software you build across millions of users -- that may start to be false.”
02:10
“AI has made it so that anyone can get to level one understanding in almost any domain, which opens up a whole realm of possibilities for what individuals can actually build.”
10:22
“Even if you're only doing 5% of the task, that 5% gets super amplified because the AI does the other 95%, and you become twenty times more productive.”
04:00
“The filter isn't does this work in testing. It's can I prove this works consistently?”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogystory
00:00I listened to Dario and Daniela, the cofounders of Anthropic, speak at a Claude conference in San Francisco, and I learned something that I wasn't expecting. Most people are building the wrong things with Claude. So I dug deeper, and after studying everything the Anthropic founders have recently said, I uncovered four rules for how they actually decide what to build.
00:18And it turns out that these rules apply to you whether you're building a side project or your dream business. The first rule is you have to recalibrate what's possible. If you look at Anthropic's product release schedule, they are shipping faster than almost any company ever.
00:30You can see each release highlighted here, and this isn't normal. Most companies will do one of these releases every quarter if they're lucky, and they're doing this almost every single day. So how do they actually do this and decide what's worth building?
00:41Well, they've successfully recalibrated three key factors that help them decide what they can and can't build. The first is cost.
00:48Can you afford to build it? Before AI, a lot of builds died because you had this great idea, but you had to hire someone or it just cost too much to build for the opportunity. But the math has now changed.
00:58Listen how Dario framed this new reality in a recent interview. Software is gonna become cheap, maybe essentially free.
01:05The premise that you need
01:08to amortize a piece of software you build across millions of users, that may start to be false. The calculus has entirely changed for what you should build because you don't need millions of users for it to be successful. That means the bar for what should be built is a lot lower, meaning you should build a lot more.
01:23The second factor that they've recalibrated is domain expertise. Essentially, do you know how to build what you wanna build? Here's a clip from Daniella at Stanford who is one of the cofounders of Anthropic, and it's also Dario's sister if you're wondering that because I was also.
01:36I didn't think I could build a website. Right? And now using Cloud, I'm like, oh, man.
01:39That's so easy. Like, I just click a couple buttons and Cloud, like, builds a website for me. AI has made it so that anyone can get to level one understanding in almost any domain, which opens up a whole realm of possibilities for what individuals can actually build.
01:50And the third factor is a time constraint. Do you actually have enough time to build it? A lot like cost, the total time required to deliver features is entirely collapsing.
01:58And as a result, the Anthropic team can now elect to ship more features and products that they may have never built before. For example, Claude CoWork, one of their biggest products. Here's Dario talking about how long it took to build it.
02:08This was a a version of our tool, Claude Code, for non coding. This was built in a week and a half, almost entirely with Claude Opus. So they're building quicker and more things, but does the data back it up?
02:19According to their internal data, 27% of Claude assisted work that would have never been completed without Claude. So how can you do this with whatever you're working on?
02:27Well, run a shelf audit. Every person watching this video has a mental shelf of ideas that you've parked because of one of these three factors. Right?
02:35The time, the skill, or the cost. Now a lot of these items on the proverbial shelf are very doable, and you just haven't thought to look or you haven't recalibrated what is actually possible. Now before you start building everything on the shelf, rule number two could actually kill a lot of the ideas that you have.
02:50Rule two, kill anything you can't verify. Simply put, for certain projects, if you can't verify it, you can't release it. At Anthropic, the most extreme example of this happened earlier this year.
02:59They had to hold back one of their most capable models, Claude and Mythos, because they weren't certain that it was reliable and secure enough. Daniella spoke about this recently at Stanford. But we're just not confident enough yet.
03:09It's irresponsible of us to release it until we are confident that all of the patching that needs to be done has been done. So the filter isn't does this work in testing. It's can I prove this works consistently?
03:21Now are you gonna ship an AI model that could change the trajectory of humanity? Probably not. Maybe.
03:25Maybe. But probably not. But this concept a thousand percent holds true for you.
03:29For example, if you're building an email response automation or a financial analysis tool, can you prove it works consistently? I had to learn this the hard way. I built an email drafting automation tool for one of my clients, and none of the emails were done properly.
03:41I couldn't verify that it was doing things correctly, so I probably should have never built it. Now you don't have to run everything through this lens, but you need to when the cost of error is high. What this means is if there is an error, what is the cost of this error?
03:53In Anthropix case for Methos, the cost of error was extremely high. And I'll dive into this cost of error concept later in this video as well. So if the cost of error is high, before you actually do anything, here are the two things you have to consider.
04:03First, define how you'll verify before you build. Figure out what the final output needs to be to confirm that it actually works. If you can't figure out what is needed to verify before you start building it, how will you know it works at the end of building?
04:15Step two is set a way for Claude to verify its own work as it builds. This is a tip from Boris Czerny, the creator of Claude Code, where he tweeted, the most important thing to get great results out of Claude Code is to give Claude a way to verify its work. If Claude has a feedback loop, it will two to three x quality of the final result.
04:32If you do these two things, and you'll build outputs actually provide value. Not stuff that looks cool and breaks the second it's running because the cost of error is just too high. So to help with this, here's a prompt that you can add to the end of anything you do to force Claude to think about how to verify the output.
04:46You can also add this to your Claude dot m d file, so it's a general rule across your entire project. Before we get to rules three and four for how Anthropic actually decides what to build, we're going through a lot of concepts quickly here. So if you want something that you can go at your own pace, I have a free five day email course that walks through a lot of these concepts, as well as the AI system that I used as a COO of a $25,000,000 startup.
05:08To get that, click the first link in the description. It's entirely free and based on over 5,000 people who have gone through it, I'm confident you'll love it. Rule number three is know who you're building for and who you're not building for.
05:18Most people try and do everything for everyone. And the result is they build a bunch of stuff that is just sort of valuable when they should have focused and provide value on a specific group of people.
05:28There's two high level concepts you need to understand for this rule. The first is your ICP, and the second is your audience anti goal.
05:35An ICP stands for your ideal customer profile. Who are you building this for? And for people watching this video, the answer can be just you.
05:41That's totally fine. And second is and an audience anti goal is the customer you are explicitly not building for.
05:47And what has helped Anthropic be so successful is they have clearly defined both of these. First, listen to what Daniella said at Code with Claude about their ICP. I think in many ways developers are the most important users of Claude.
06:00And here's Dario talking about what they're not building or their audience anti goal. We we don't make, you know, models that generate images and videos and for many reasons. They know who they're building for, which is developers, and they know what they're not building, image and video generation for creatives and consumers.
06:15If you look on the other hand, right, OpenAI, they've gone for general consumers, image generation, video generation, enterprise, social media, essentially trying to do everything at once. And this is likely why Anthropic has been able to grow much faster than OpenAI.
06:28Listen to Steve Ballmer, the then CEO of Microsoft, talk about who they were building for. Developers. Developers.
06:35Developers. Developers. Developers.
06:40Developers. Developers. Developers.
06:42Developers. Developers. The tenacity is amazing.
06:47Not to mention the sweat stains. I may have to hop on my bike before the next YouTube video I make so that I can just freaking bleed that intensity. But anyway, the filter is clear.
06:56They know who they're building for and who they're not building for. And by drawing that line in the sand, every decision in terms of what you're building becomes faster and more effective. And the added benefit of this is something that most people miss.
07:07When you pick one customer, your build start to compound and complement each other over time. Each new build sharpens a relationship with that one audience. The next build is easier and more valuable than the last.
07:17For Anthropic, they built for developers, which gets more developers to use it, which helps train their clawed model, which then makes it better for developers. They then use that same clawed model to build clawed code, which helps developers. And this clearly worked.
07:28Right? Anthropic grew 80 x in q one twenty twenty six against a 10 x plan. Eight zero.
07:34That's a lot of x's. Now here's where it gets interesting for me and you. This concept scales down to whatever you're working on.
07:39You just have to understand who you're building for and how it helps them. Let's say you're building for yourself and you stay laser focused on what you need. You shouldn't chase shiny objects that solve problems that you don't actually have.
07:48If you're building for a specific audience, focus on that audience. For me, this is my audience that I look at every single day. That ends up being 30 to 50 year old males who live in The United States.
07:58Honestly, shout out everyone watching this. You guys are legends. And if you're unsure who you're building for, a good starting point could be identify who you aren't building for and set your anti audience goal.
08:06At the end of the day, the most successful people I know only work on things that complement each other thing that they're working on. And that's because there's no waste of time and effort. And that's what Anthropic did, and that's what you should do as well.
08:17So that's the first three rules that Anthropic follows when deciding what and who to build for. And before we get to rule number four, which dives into how to build what we're working on. If this is your first video, welcome to the channel.
08:26But if it's your second or more, here is our anti slop agreement. The visuals, the testing, the time I put into this, this is all for humans, not for AI clanker robots. So all I ask is that you subscribe as part of this agreement to help this content reach more people.
08:39Also, every couple of weeks, I give away a clawed max subscription, so comment below with what you're building to enter. Rule number four is build middle to middle, not end to end.
08:47Most people when they think about AI, they're thinking about end to end. They prompt it, they walk away, and they hope something good comes out on the other side. Anthropic is looking at it a different way, middle to middle.
08:57Every task you start has three parts. A start where you frame the problem, a middle where the work gets done, and then an end where you deliver actual final product. End to end is when AI takes all three parts.
09:06The start, the middle, and the end. The human in general is just gone. Middle to middle is where AI only takes the middle.
09:12A person starts, AI does the middle, and a person reviews at the end. In a morbid example, here's Dario explaining a use case they refuse to support, a classic end to end example. If you have a large army
09:25of drones or robots that can operate without any human oversight, where there aren't human soldiers to make the decisions about who to target, who to shoot at, that that presents concerns. And we need to have a conversation about about how that's overseen, and we haven't had that conversation yet.
09:42And so we feel strongly that, you know, for for, you know,
09:47those two use cases should, uh, should not be allowed. Autonomous weapons are end to end. AI decides who to target.
09:53AI fires. AI delivers the outcome. I know that's an extreme example, but it's where the cost of error, losing a human life, is too high.
10:00They drew the line and they refused to build for use cases like this. Now, the same logic shows up in smaller stakes. Listen to how Dario explained the math that makes middle to middle actually work.
10:09Comparative advantage is surprisingly powerful. Right? Even if you're only doing, like, you know, 5% of the task, like, you know, that 5% gets super amplified and levered because it's like you're only doing 5% of the task,
10:22the AI does the other 95%, and so you become, you know, twenty twenty times more productive. The 95%
10:28is the middle. Claude takes that. And then the 5% is the stuff that matters, the start and the end.
10:33The decision at the front, the judgment on the back. That's what stays human, and it gets 20 x more leverage because the middle just got cheap and more effective. And if you look at two of Claude's most successful products, Claude Code and Claude Cowork, these are tools that are optimized for middle to middle workflows.
10:47Interface assumes that you're at both ends of the work for most of the use cases, and that's exactly how I use it. I'm at the start of every interaction, and then I'm reviewing the final output. So how can you use this concept when evaluating which Claude projects to build?
10:59Well, before you start building anything, think about the three parts of the process. The start, what decision or framing do you need to bring before Claude actually does any work? The middle, what routine work do you wanna hand off?
11:10And the end, where do you review, judge, and sign off on the final output? If Claude is doing all three parts, you've built end to end, and this will likely produce low quality work. My recommendation would be to redesign it so you're at the start and end.
11:21And just this mental tweak of thinking about things middle to middle makes everything easier because the output doesn't actually have to be perfect. You can take it and make it perfect at the end. So if you follow the four rules I covered here, you will get amazing results with whatever you're building.
11:34And look no further than the success that Anthropic has had. Now if you like this video, you'll love this video where I break down how Anthropic's own engineers actually prompt Claude code.
11:43That's the tactical layer underneath all of this, which which will streamline whatever you're building. I'll see you over there. Peace.
The Hook
The bait, then the rug-pull.
Most people are building the wrong things with Claude. That is the claim Austin Marchese opens with after attending a founder talk in San Francisco -- and the four rules he unpacks to fix it turn out to be borrowed directly from how Anthropic itself decides what is worth shipping.
Named ideas worth stealing.
01:23list
The 3 Factors to Recalibrate
- Cost: can you afford to build it?
- Domain Expertise: do you know how to build it?
- Time: do you have enough time to build it?
Before AI, these three factors killed most ideas. All three have now collapsed.
Steal forAny go/no-go decision framework for new projects
05:15model
ICP + Audience Anti-Goal
- ICP: the exact person you are building for
- Audience Anti-Goal: the person you are explicitly NOT building for
Defining both sides of the audience filter simultaneously. The anti-goal is as load-bearing as the ICP.
Steal forProduct positioning, feature prioritization, content strategy
08:42model
Middle to Middle
- Start: human frames the problem
- Middle: AI executes
- End: human reviews and judges
A workflow model that preserves human judgment at the edges and hands off only execution. Contrasted with end-to-end where AI owns all three stages.
Steal forDesigning any Claude automation or Claude Code workflow
02:50concept
Cost of Error Filter
Before building any AI automation, ask: if this output is wrong, what is the cost? High cost = define verification first. Low cost = ship and iterate.
Steal forEvaluating any AI automation before building it
How they asked for the click.
VERBAL ASK
05:08link
“Click the first link in the description. It's entirely free and based on over 5,000 people who have gone through it, I'm confident you'll love it.”
Mid-video pause between rules 2 and 3. Frames the email course as a slower-paced companion. Secondary CTA at the very end for the next video in the series.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
