11:49Austin Marchese · Tutorial
How Anthropic Founders ACTUALLY Pick What to Build with Claude
A former startup COO reverse-engineers the four decision rules behind Anthropic's industry-leading shipping velocity.
June 5thThe argument in one line.
Most Claude Code projects fail because builders chase the wrong ideas or execute as technicians instead of leaders; follow three rules—avoid the idea trap, build where you have earned domain judgment, and lead like a CEO rather than execute—then filter every project through fo.
Read if. Skip if.
READ IF YOU ARE…
- A non-technical founder or side-project builder who wants to use Claude Code to ship fast and needs a framework for picking projects that won't waste 6 months on dead ideas.
- A technical person with domain expertise in a specific vertical who's unsure whether an AI project idea fits YC's thesis or if you're about to compete directly against OpenAI and Anthropic.
- A COO, operations leader, or someone with earned credibility in a niche market who's considering building an internal tool or specialized solution and wants to validate the decision before coding.
SKIP IF…
- You're looking for technical tutorials on Claude Code syntax, API integration, or how to actually build the thing — this is strategy, not implementation.
- You're an experienced AI product builder who's already shipped multiple projects and validated product-market fit — this is foundational thinking for first-time builders.
The full version, fast.
Before opening Claude Code, decide whether anyone but you will use what you build, and whether the project sits in the path of frontier AI labs or alongside them. The framework is three rules drawn from Y Combinator's Garry Tan: avoid the idea trap by scoping a precise user and refusing to compete with steamrolling foundation models, build where you have lived long enough to evaluate good versus great output since that judgment is the real moat, and operate as the CEO of an AI team rather than the executor. Lead by writing a CLAUDE.md onboarding doc, planning before prompting, granting scoped permissions, assembling specialized expert agents, reviewing volume instead of producing it, and automating yourself out with hooks, scheduled agents, and loops.
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 – 00:42
01 · Cold open — the gate
Claude Code logo animation, YC credentialing, three-rule roadmap preview, Austin intro as $25M+ startup COO.
00:42 – 03:11
02 · Rule 1 — Avoid the Idea Trap
Two failure modes: unclear user identity and jumping in front of the AI steamroller. Path 1 (only user) vs Path 2 (distribution-first). Self-check questions.
03:11 – 05:52
03 · Rule 2 — Build Where You Live
T-shape model. Surface knowledge vs deep vertical. Evals as the real moat. Garry Tan ethnographer clip. 49.7% of AI tools concentrated in one category.
05:52 – 07:32
04 · Anti-SLOP agreement + Rule 3 intro
Subscribe ask framed as mutual agreement. Shift from execution to leadership layer.
07:32 – 11:11
05 · The 6 Moves of an AI Leader
CLAUDE.md onboarding, pre-prompt planning interview, agent permissions, specialized sub-agents, manager review, hooks/scheduled-agents/loops. BuildPartner.ai mention.
11:11 – 11:21
06 · 4-Question pre-project test + outro CTA
Four filters before starting any project. Cross-promotes Andrej Karpathy video.
Lines worth screenshotting.
- Building a cybersecurity tool to audit codebases puts you in direct competition with Anthropic, OpenAI, and the biggest labs — a battle you will lose.
- If you are the only user of what you're building, you should optimize for speed and function, build it ugly, and stop making it look pretty.
- Distribution is the only thing that matters when you want other people to use your product — solving the right problem for the wrong audience is still failure.
- The Y Combinator concept of jumping in front of a steamroller means working on a problem that frontier AI labs are already solving better than you can.
- A non-technical person used these three rules to build an app generating $400,000 per year in under a month with Claude Code.
- Rule: build where you have earned domain depth — your lived experience in an industry is a competitive advantage that no amount of AI access can replicate for someone without it.
- Operating as an AI CEO means directing Claude rather than executing alongside it — the builder who prompts sets strategy, Claude handles implementation.
- The four pre-project filter questions exist to kill bad ideas before you spend any tokens — most projects fail at question one (who exactly uses this?).
- Y Combinator's core doctrine in the AI era: scope tightly, know exactly who you're solving for, and build the thing that's perfect for that specific set of people.
- Building something that gets better as AI models get more capable requires being anchored to a domain problem, not a technical approach.
Steal the framework, own the vertical.
LFB playbook
The moat is not the AI — it is twenty years of watching funnels work and fail that nobody can prompt their way into.
- Record a 'build where you live' video using direct-response conversion as the domain — it's a story nobody else in the Claude Code tutorial market can tell.
- Turn the 4-question pre-project test into a one-page PDF lead magnet — immediately usable and brands Joe as the strategist, not the tutorial guy.
- The CLAUDE.md onboarding angle is content Joe already lives; make a harder, more specific version with real examples from JoeFlow, MCN, and Clip Lab.
- Rule 3 move 6 (hooks/scheduled agents/loops) maps directly to JoeFlow's morning-batch-launcher thesis — this is a product demo, not just a concept.
- Use the T-shape model in MCN+ positioning: members get Joe's vertical (evals from 20 years of direct response) plus the tools — not just the tools.
Terms worth knowing.
- Y Combinator (YC)
- A Silicon Valley startup accelerator that has funded companies including Airbnb, Stripe, and DoorDash. It provides seed funding, mentorship, and a network in exchange for a small equity stake.
- Domain depth
- Deep firsthand knowledge of a specific industry or problem, earned through years of working in it — considered a key advantage when deciding what products to build, because it reduces the risk of solving the wrong problem.
- AI CEO
- A framing for how a non-technical builder should operate when using AI coding tools: directing the AI like a CEO delegates to employees — setting strategy, reviewing outputs, and making decisions — rather than getting lost in execution details.
- COO (Chief Operating Officer)
- A senior executive responsible for overseeing a company's day-to-day operations, typically reporting to the CEO and managing the execution of business strategy.
- Idea trap
- The mistake of building a product based on what sounds exciting rather than on validated evidence that real users need and will pay for it.
Things they pointed at.
00:00channelY Combinator / Garry Tan interviews
11:12channelAndrej Karpathy Claude projects video
Lines you could clip.
03:54
“Being able to do evaluations of what models and what prompts are good — that's actually turning out to be the moat for many startups.”
05:05
“49.7% of all AI tools being built are in one category. The rest is wide open.”
05:52
“In the AI era, you already have a team at your disposal. The question isn't whether you have a team because you do. The question is whether you're actually ready to lead them.”
11:11
“Stop waiting for somebody to save you. Stop waiting for permission to do these things. You can just do these things.”
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.
metaphoranalogy
00:00Claude Code lets you build anything you want. Luckily, Y Combinator, the company that helped start Airbnb, Stripe, and DoorDash is very public about what projects you should and what projects you should not start in the age of AI. So after analyzing what their CEO, Gary Tan, has said, I've uncovered three rules you need to follow whether you're working on a side project for yourself or a business you want to grow.
00:20The first two rules cover what to build so you save time and money, and the last rule covers how to build it so you're successful. By the way, I'm Austin. I was a COO of a tech startup worth over $25,000,000.
00:30And about a month ago, I used the three rules that I'm gonna cover here to help a nontechnical person use Claude code to build an app that makes over 400,000 a year. So I know with certainty that these rules actually work.
00:42Rule number one is avoid the idea trap. Before we talk about what to build, you have to understand what not to build. Because unfortunately, most projects are dead on arrival because people just build the wrong thing.
00:53And there are two distinct ways that most people fall into this. The first way is that the user isn't clear. Simply put, you need to have a deep understanding of who is actually going to use what you're building.
01:03So there are typically two paths. Path one is you're the only user. Let's say it's an internal tool, an automation that saves you three hours a week, a side project where you're just learning.
01:12In this case, you don't have to worry about distribution or getting people to use it because it's only for you. And as a result, you shouldn't worry about making it pretty and you shouldn't worry about scale. You are the only user.
01:25You should optimize for speed and function. You need to build it ugly and fast and stop making it look pretty because that's just a waste of time. And path two on the other hand is you want other people to use it.
01:35And in this case, distribution or getting people to use it is the only thing that matters. To solve this challenge, you have to understand who you're building for and what problems they actually have.
01:45Here's Gary, the CEO of Y Combinator, talking about the importance of tightly scoping whatever you're building. One thing that we're seeing is that if you, like, scope what you're doing and make the thing that is perfect for that set of people, there you can't just take ChatGPT and have it do this type of work yet.
02:03He's essentially saying, you need to know exactly what problems you're solving and solve those problems in an elegant way. The second way people fall into the idea trap is they jump in front of a steamroller. Let me walk you through a simple example of what this means.
02:16Let's say you wanted to build a cyber security tool to help audit a code base. This is a great idea. Right?
02:22It could work just for you or it could work for thousands of others. It's really valuable. This is exactly what Anthropic, OpenAI, and all of the Frontier Labs are working on.
02:31You're directly in line competing against the biggest and smartest people on the planet. A battle that you will lose. You're essentially that's rolling towards you.
02:41So if me and you are David, how do we compete against the Goliaths? Well, the truth is you don't. You need to ask yourself, how can I build something that as AI models get better, this becomes more valuable, not less?
02:52But to help you avoid the idea trap and keep you from building the wrong thing, ask yourself these questions. Is this just for me? If not, can I name five specific people who'd use this today?
03:02Two, is this adjacent to AI progress? Will it become more valuable, not less over time? That's rule number one, so you know what not to build.
03:10But what should you build? Rule number two is build where you live. We've discussed the importance of knowing who you're building for, but what should you actually focus on?
03:17The obvious answer is you should focus on where you have expertise or the most domain knowledge. And this is correct, don't get me wrong, but not for the reason you think.
03:24In today's world, just knowing how to build a website or how to make a Facebook ad isn't valuable. And why isn't it valuable? Because you can prompt AI to get any of that.
03:34That's just surface level understanding. But what is valuable is the evaluation letter. Knowing what makes a website actually convert versus a website that doesn't or knowing what makes a Facebook ad successful versus not.
03:46That's where the deep value lives. It's the judgment between good and great. A fancy term for this is evals.
03:51Evals are basically knowing how to tell what's good and bad output from AI. Here's Gary talking about why evals are the real moat, which essentially means the real differentiation.
04:00You know, being able to do evaluations of what models and what prompts are good. It's like that's actually turning out to be the moat for many startups.
04:07Yep. Part of design, I think, is actually the empathy for the user. Like, you sort of have to be like an ethnographer.
04:13He said founders as ethnographers, which I had to look up because we both know that I didn't know what that meant. And here is the definition.
04:21A person who studies and describes the culture of a particular society or a group. That is the juice. So Gary's saying the moat or your unfair advantage isn't the AI itself.
04:30The moat is where AI can't replicate. The judgment you've earned from watching things actually work and fail in your domain of expertise. A mental frame that I use for this is .
04:40The top of the t is your surface knowledge. It's broad. It covers a lot of width, but it doesn't have a lot of depth.
04:46It's shallow. This is essentially the same as anyone who's just asking AI prompts. The vertical of the t, that is where you've gone deep.
04:53That's where you've watched something work, where you've watched it fail, and learned the difference between good and great. That's where you've lived. The vertical is your moat.
05:01That's where you build. And this should really excite you because of this number. 49.7% of all AI tools being built are in one category.
05:09Healthcare, 1%. Legal, less than 1%. Education, less than two percent.
05:14Half the market is fighting over the same slice. The rest is wide open. So an engineer might be a 10 in technical skills, but in your domain, they're a zero.
05:22And you might be a one in technical things, but in your domain, you're a 10. And that's the d part. And that's what's really valuable.
05:29Now, before we get to rule three, which will provide the playbook for you to build something successfully. If this is your first video of mine, welcome to the channel. If this is your second or more, here is our anti SLOP agreement.
05:39The visuals, the testing, the time I put into this video, that's for humans. It's not for AI robots or data scrapers. So all I ask is you subscribe as part of this agreement to help this content reach more people so I can keep making videos like this.
05:52Moving to rule number three where everyone is a CEO now. So you know what not to build and what to build. But how do you actually build something successfully in this new world?
06:00Most people are still operating with the old mindset. I do the work. I execute.
06:04In the AI era, that's the wrong game. The new mindset is I orchestrate, I direct, I review, much like a CEO. This shift is from the execution layer to the leadership layer.
06:14In the AI era, you already have a team at your disposal. That's Claude Codes, specialized models, custom skills. They're sitting there waiting.
06:20The question isn't whether you have a team because you do. The question is whether you're actually ready to lead them. Here's Gary pointing out what the best leaders can do in the AI world.
06:28Really
06:29super young teams that basically are starting out with nothing. And they can go from really 0 to $10,000,000
06:38a year in revenue, sometimes in the course of less than twelve months. And they can do it with less than 10 people. And this point isn't just about startups.
06:47Whether you're at your day job, working on a side project, or building a 100 person company, everyone needs to level up to leadership. I actually saw this firsthand with Nick, a nontechnical founder at BDGE, when we were working on launching a Vibe coded app that was entirely built with Claude Coder. One day he called me and he's like, I'm more productive than my engineering team.
07:06I'm building fashion than them. They're way too slow. And this was a nontechnical founder who was outpacing his own full time engineers.
07:12And it wasn't because he was outcoding them. He was outleading them. He just directed AI better than they did.
07:17And after we worked together for about forty five days, we were able to launch an app worth over 400 k. If you're interested in how that actually happened, I do have that link below. I made a whole YouTube video on it.
07:28But with that, what is operating at the leadership leadership layer actually look like day to day? And what do successful leaders do? Well, there are six things you need to follow.
07:34The first is onboard AI like a new hire. Don't just open Claude code and start prompting a cold. Write a Claude dot m d file first.
07:41Think of this file as your AI's onboarding doc. The better context you give it, the less time you spend correcting it later. It's essentially the same way a manager onboards a new employee.
07:49Here's a prompt you can use to help create this. The second move is write a plan before you do any work. Don't just prompt and hope.
07:56Have AI interview you first to figure out exactly what you wanna build. What's the core problem this feature is solving? What does success look like?
08:03What should this not do? Ten minutes of planning could save you hours and hours of time. Here's a prompt you can try to help get information out of you to make sure whatever you're building is actually what you wanna build so you can move faster.
08:14Moon number three is give AI employee level permissions. Much like when you onboard employee, you have to give them permissions to do things. And if you've used Cloud Code, you know how annoying it can be to constantly approve permissions for things that should just have access to.
08:27For reversible actions, just let the AI agents flow. For anything that's destructive, make it stop and ask for permission. This is a lot like how you'd give employees permissions to do things without your approval or not do things.
08:39You wanna protect the agent from itself, but you also don't wanna have to approve everything it does, which is very annoying. So try this prompt to give your agents proper permissioning while working on your computer. Move number four is build a cabinet of specialized experts.
08:52Start thinking about creating your own panel of advisers that specialize in a specific task. One trained on your sales playbook, one on your content, one on your finances. Specialized employees be one generalist every time, much like building a team.
09:05Here's a prompt you can try to do exactly this. Now if this does sound complex, I built a tool called buildpartner.ai that solves this exact problem.
09:12You can run a skill called slash b p colon expert advice that will take whatever you're working on and run it through an expert in that field so it gives you specific advice from that expert. I love this tool. I built it because I ran into this issue all the time, so you can check that out.
09:26Move number five is review like a manager. Have AI bring you volume and you pick the winner. Don't have it do end to end.
09:32Have it do tasks that are middle to middle. You want an idea, have it bring you ideas and you approve them. Here's a prompt you can try to help with this.
09:38Move six is remove yourself as the bottleneck using Claude's power user features. These are three things I lean on a ton. Hooks.
09:45These fire automatically when something happens. Like, every time you finish a session, Claude will log what works and what doesn't. Scheduled agents.
09:52These run on a timer remotely. Let's say you wanna run something daily or weekly, Claude can do whatever you want, whenever you want using this functionality. And then loops.
10:00These are things that Claude will run automatically on your computer however often you want. This is how you create a system that works while you sleep. End of the day, the more you're a bottleneck, the less you're leading.
10:09Here's a prompt to help integrate these into your system. That's rule number three. Everyone is a CEO now, which means you're a CEO now.
10:16Stop waiting for somebody to save you. Stop waiting for permission to do these things. You can just do these things.
10:21So you now know these three rules. But before you start any project, you need to be able to complete this four question test. Who exactly is this for?
10:29Yourself, your team, your clients, or external users? Be specific or kill it. Anyone interested in next isn't an answer.
10:37Is this in front of an AI steamroller? Don't pick up a penny in the path of AI disruption. Build adjacent to it.
10:42Do I understand this in practice, not just on paper? If you can't tell me what you've watched fail in real world, you're not the expert yet. Understand your t shape.
10:51And the fourth and maybe the most important, is this congruent with the rest of my work? Are you working on something that complements everything else in your life? Set up things that compound over time.
11:01If your idea is a one off that pulls you sideways, maybe it's not worth doing. Now if you like this video, you'll love this video where I break down the exact system that Andrea Carpathi, the former director of AI at Tesla, uses to 10 x his clawed projects.
11:14Once you know what to build and how to build it, this system will optimize your setup. I'll see you over there. Peace.
The Hook
The bait, then the rug-pull.
The title is a gate, not a promise. Austin Marchese opens with a threat: most Claude Code projects are dead on arrival because people build the wrong thing. The first thirty seconds name-drop Y Combinator, Airbnb, Stripe, DoorDash, and a $25M startup COO backstory. By the time the first rule appears, you already believe he has something real.
Named ideas worth stealing.
00:42concept
Avoid the Idea Trap
Two failure modes: user not defined, or competing directly against frontier AI labs. Forces a binary: are you the only user (optimize for speed/ugly) or do you need distribution?
Steal forPre-project filter — kills side-project scope-creep before it starts
03:51model
The T-Shape Moat
Top of the T = broad surface knowledge anyone can prompt for. Vertical of the T = earned judgment from watching things fail in your domain. The vertical is where you build.
Steal forPositioning argument for why Joe's direct-response depth beats any AI generalist
03:54concept
Evals as Moat
Knowing what makes good vs. bad AI output is the actual competitive advantage. Garry Tan: 'that's actually turning out to be the moat for many startups.'
Steal forMCN+ positioning, LFB framing — domain experts win the AI era
07:48list
6 Moves of an AI Leader
- Onboard AI like a new hire — write CLAUDE.md first
- Write a plan before prompting — have AI interview you
- Give AI employee-level permissions (reversible=auto, destructive=ask)
- Build a cabinet of specialized sub-agent experts
- Review like a manager — AI brings volume, you pick the winner
- Remove yourself as bottleneck via hooks, scheduled agents, loops
Steal forContent series: Joe already does 5 of 6 — the 6th maps directly to JoeFlow's morning-batch-launcher
10:23list
4-Question Pre-Project Test
- Who exactly is this for? (specific or kill it)
- Is this in front of an AI steamroller?
- Do I understand this in practice, not just on paper?
- Is this congruent with the rest of my work?
Steal forAny project-scoping conversation with clients or in LFB sessions
How they asked for the click.
VERBAL ASK
06:10subscribe
“The visuals, the testing, the time I put into this video — that's for humans. It's not for AI robots or data scrapers. So all I ask is you subscribe as part of this agreement.”
Framed as a mutual social contract (anti-SLOP agreement) rather than a standard ask. Converts subscribe from obligation to reciprocity.
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.
