Modern Creator
Austin Marchese · YouTube

Stop Building Claude Projects Nobody Wants. Copy Alexandr Wang Instead

Four repeatable steps Meta's Chief AI Officer used to ship Muse: write the memo, prove it fast, harden it, make it memorable.

Posted
4 days ago
Duration
Format
Tutorial
educational
Views
14.7K
268 likes
Big Idea

The argument in one line.

Wang's process for building Muse reduces to four steps anyone can copy: write a memo, build a proof of concept fast, harden it against a test table, then make it memorable.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • A solo builder who keeps starting Claude projects and abandoning them because the goal was never written down.
  • Someone with an AI automation or app that works when they watch it but breaks when it runs unattended.
  • A founder building a consumer product who wants a concrete way to think about onboarding and shareability.
SKIP IF…
  • You want deep technical detail on how Muse was trained or evaluated; the video stays at the level of process, not model internals.
  • You already run a formal eval and launch-gate process and need more than a five-to-ten row test table.
TL;DR

The full version, fast.

Austin Marchese studied Alexandr Wang's public talks on building Meta Muse and distilled four steps. First, write a memo answering what it is, who it's for, what it isn't, what done looks like, and the smallest version that proves it. Second, build a proof of concept in a week or two by cutting scope and timeline. Third, harden it with a spreadsheet of behaviors, a test per row, and launch-blocking versus good-enough thresholds. Fourth, make it memorable by shrinking time to magic and designing for the screenshot.

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 →
Chapters

Where the time goes.

00:00 – 00:48

01 · How Wang starts every project

Sets up Muse, Wang's public interviews, and the four-step framework.

00:48 – 02:30

02 · Step 1: Write the memo

Five questions plus a feature list that turn an idea into concrete deliverables.

02:30 – 07:39

03 · Step 2: Build the proof of concept fast

POC versus MVP, the one clear point of view, the more trap, and three ways to move faster.

07:39 – 12:51

04 · Step 3: Harden it

The behaviors spreadsheet, evals, launch-blocking thresholds, and a five-step small-scale version.

12:51 – 15:52

05 · Step 4: Make it memorable

Why good isn't enough, time to magic, and the screenshot test.

15:52 – 18:26

06 · The adulthood you dreamed of in childhood

Wang's quote on AI and agency, and the one action to take: write a memo.

Atomic Insights

Lines worth screenshotting.

  • A memo works like plugging a destination into GPS: you decide where you're going before you hit the gas.
  • A proof of concept answers whether it can be done; an MVP answers whether it's ready for users. The first comes before the second.
  • Wang had a Muse prototype in front of Meta's board within one to two weeks, and many first-prototype decisions survived into the final product.
  • Without one clear point of view, a product turns into a Frankenstein smoothie that users feel immediately.
  • The 'more' trap, more automation, more features, is how builders end up with something only they can use.
  • Cut your proof-of-concept timeline by 75 percent: if you think eight days, plan two; if eight hours, plan two.
  • Building can become a dopamine loop with AI, so ask whether you need to build it at all before you start.
  • 'It worked when I tried it' and 'it works continuously' are two very different things, and the gap is what hardening closes.
  • AI is non-deterministic, so one bad output can make a user decide the whole product is trash.
  • Wang's team hardened Muse with a spreadsheet of 100+ specific behaviors, an eval for each row, and regular reviews of the weakest rows.
  • Set two bars per row: what blocks launch and what is good enough to go live.
  • If an AI feature only produces good output 80 percent of the time, you're often better off removing it, since heavy users will hit the failures.
  • You won't catch everything before launch: people used Muse for 3D demo games Meta never anticipated.
  • If the result of your product can't be screenshotted and sent to someone, it won't grow organically.
  • The strategy for memorability is to reduce time to magic: reliable gets used, memorable gets shared.
Takeaway

Write it down, prove it, harden it, make it shareable.

WHAT TO LEARN

Big AI projects start with a written memo, a fast prototype, a test table that makes them reliable, and a design that makes the first result memorable.

  • Answer what it is, who it's for, what it isn't, what done looks like, and the smallest provable version before you build anything.
  • Cut your proof-of-concept timeline by 75 percent and decide up front what you won't build or can plug in from existing tools.
  • Ask whether you need to build it at all, because building with AI can become a dopamine loop with no user at the end.
  • Keep one clear point of view so the product doesn't become a Frankenstein of every feature you thought of.
  • Turn 'it worked when I tried it' into 'it works every time' with a table of 5 to 10 behaviors and a test for each.
  • Set a launch-blocking bar and a good-enough bar for each row, then ship when you clear them and learn from real users.
  • Remove any AI feature that only gets it right most of the time, since repeat users will meet the failures.
  • Judge your product by what someone gets in the first couple of minutes and whether they would screenshot it.
Glossary

Terms worth knowing.

Proof of concept (POC)
A quick, rough build that answers whether the idea can be done at all, before anyone worries about users or polish.
MVP
A minimum viable product, the smallest version that is ready for real users, as opposed to a POC that only proves feasibility.
Eval
A test for one specific behavior a model or product must get right, scored so you can track improvement.
Hardening
Turning something that works when you watch it into something that works reliably without you monitoring it.
Time to magic
How long it takes a new user to get the first valuable or surprising result from a product.
Resources

Things they pointed at.

Quotables

Lines you could clip.

06:40
“It worked when I tried it and it works continuously are two very different things.”
Short, universal, no setup needed→ TikTok hook↗ Tweet quote
10:40
“If you have an AI feature that only produces a high quality output 80 percent of the time, you're better off not having it at all.”
Contrarian and specific→ IG reel cold open↗ Tweet quote
14:40
“If the result can't be screenshotted and sent to someone, then it's not going to organically grow.”
Clear rule of thumb→ newsletter pull-quote↗ Tweet quote
The Script

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.

Alexander Wang, Meta's chief AI officer, is one of the youngest and brightest minds in AI. But until recently, we almost had no insight into how he starts and builds his AI projects. But that changed when Meta launched Muse, because since then, he's been everywhere.
Podcasts, interviews, thousands of tweets. So I went through every piece of content he's made, and surprisingly, his building process breaks down into four repeatable steps that anyone can copy. So in this video, I'll walk through each step and give you the exact prompts that you can use today.
And then in the end, I'll dive into this. quote from Wang that I just have not been able to stop thinking about.
they dreamed of in childhood. He describes it as AI can give everyone the adulthood they dreamed of in childhood. And we'll dive into that later in the video.
But first, the starting point for building any big project with AI is step one, write the memo. After Muse launched, there were millions of downloads overnight and Metastock rose more than 11%, which equates to about $190 billion in valuation. But what did Wang do before he took the lead on building Muse?
First, he wrote a memo. And if you've never worked inside a big company, a memo is a short document that explains what you're building and why.
Zuckerberg loves them. Jeff Bezos has centered every key decision for Amazon around memos. These are extremely effective.
And in February of this year, Wang sat down and sent a memo to Zuckerberg. What he's saying here is before anyone built anything, he wrote down what he believed this thing was. And then based on that memo, that's when he got the green light from Zuck.
And so mentally think of a memo like plugging in your destination into GPS. You want to take a minute and figure out where you're going and the route that you're going to take to get there before ever hitting the gas. And whether it's just you or a team, this will align your thoughts on why and what you're building so to write your own memo answer these five questions what it is who is it for and ideally name a real person so you bridge the abstract to concrete number three what it is not four what does done look like and five what is the smallest version that proves this is possible and then one additional question that i personally like to add is what features do i want in this so if it's a complex automation what is in the automation if it's a product what is the core functionality again the point of this is to bridge the abstract to concrete deliverables that help you visualize what the final product will be.
Here's a prompt that interviews you to build the memo. And if you want to iterate on it visually, I built a tool called build pad inside build partner .ai, where you can refine the memo right in the pad and land on something you're confident in. So at this point, you know what you're building and why you're building it.
And this is where Wang did the opposite of what you would expect from a company the size of meta. Step two is build the proof of concept fast. If the memo is a thinking step two is the action.
And in this case, we're building a proof of concept or a POC. In short, a POC answers the question, can this be done? Whereas another term I use on this channel is a minimum viable product or MVP.
And this answers, is this ready for users? A proof of concept comes first and then an MVP follows. And it's important to understand the clarification.
Now, I don't want to get bogged down in the difference, but I have said those terms before. So that's how I think of it. But the reality is that people are naturally bad at this step.
We like to overthink. And the longer you overthink something, the longer it takes to find out if it will actually work. And so here's how ruthlessly fast Wang was about building his proof of concept.
prototype within like a week maybe two weeks and demoed it to the board like it was this like very fast cycle and many of the decisions that made its way into the final muse product were present in the first prototype and right here this is where action produces information comes back wang and his team made a prototype in one to two weeks to show real people and this is the board of directors at meta it wasn't meant to be perfect it was meant to find out if the point of view was right and this is where steps one and two connect so once you establish the why and what in the memo there's a clear single point of view that will determine how the proof of concept actually gets built because if there's ambiguity there it leads to issues and here's wang on what happens on when you don't have a clear single point of view.
Become kind of this Frankenstein, like smoothie almost. People feel it as users and consumers. You've had to be the exact opposite, which is like, there's one clear point of view.
Here, he's making it clear the importance of having a clear direction as you build. And I know that this sounds obvious, but in reality, it isn't actually that easy. And the reason is because this issue presents itself disguised as the more trap.
I want the automation to do more. I want the product to have more features, more, more, more. And then the next thing you know, you've built a Frankenstein that only you know how to use.
And frankly, I fall into this with buildpartner .ai. I'm still learning what I want it to be. So as I experiment, at times I can feel it a bit Frankenstein -esque.
With that caveat in mind, how do you actually go ahead and build proof of concept quickly? In a recent video on my channel, I interviewed a head of engineering at WAP, a company that's now worth over $5 billion. And there are two points that I wanted to bring into this video.
The first is you need to decide what you don't want to build. Does your proof of concept really require you to solve every problem? For example, if you're building a product, you probably don't need to figure out payments from scratch.
Just plug in a tool like WAP and move on. And this may sound obvious, but most of us don't do it. So here's a prompt that what really matters for your proof of concept and calls out any external tools that can streamline it and the second question from that interview was ask whether you need to build it at all are you solving a real problem or just excited that you can solve it building can feel like a drug especially with ai and we get this dopamine loop where we just want to build more and more things but you never stop and ask do i really need to build this and then the third thing is just cut your timeline by 75 take however long you think a proof of concept will take you and just reduce it if you said going to take eight days make it two days if you said eight hours make it two hours just do it i'm telling you it won't take as long as you think if you put your mind to it in this case in two weeks wang had something that worked you can build faster trust me push yourself but in this case for meta and muse that's not where it gets interesting yes he spent the two weeks working on the proof of concept but he spent the next seven months doing one thing before the public could touch it but before we get to that and removing the barrier to build things is why ai is so exciting to me and this isn't just about building a website it's also remove the
barrier for building and scaling businesses, which brings us to today's video sponsor WAP. So I don't think there's ever been a more exciting time than right now to start and scale a business. You guys know that I use WAP for buildpartner .ai, but I'm seeing more and more people around me use WAP to help them launch their businesses.
So my one friend runs a bakery off their front porch every Saturday, and then another friend built a site for learning games for kids, but both of them hit the same wall. And that's all the other stuff that comes from creating a business. That's collecting payment, creating an LLC, tracking customer data.
conversion history, analytics, the list goes on and on. So I texted the two of them and was like, you need to use WAP to collect payment and use their business intelligence tools to help them get set up. And then after a couple of texts, they added it immediately.
And for them, they like saving money, which isn't a hot take. Everyone likes doing that. So WAP's low payment fees related to other service providers made it an absolute no brainer.
But the reality is that's actually just the tip of the iceberg. So let's say you were starting a business from scratch. You don't actually have to start from a blank slate anymore.
Within WAP, you can pick a business business blueprint like e -commerce or service business, or you can just describe what you want directly to WAP's AI. And then it'll then give you a fully operational business off the shelf.
So that's the site, the products, the pricing already set up, and then you just edit it by prompting. This feels like magic and it is crazy. And compared to when I started my first business as a SAT math tutor when I was 16 years old, WAP and its AI tooling has made it quite literally a hundred times easier.
So whether you're starting something new or you want to scale your existing business faster and save money in the process, sign up for WAP using the link in the description. Okay, moving to step three, you need to harden it over time.
So you have the prototype, you try it, it does the thing and you're excited about it. But there is a catch with prototypes and Wang described it like this. Prototype where it like sometimes worked and was magical, but most of the time didn't work and would constantly break.
It worked when I tried it and it works continuously are two very different things. And this is the difference between a proof of concept versus something that's ready to go into other people's hands. So in the case of you're building an automation, a proof of concept works on your machine while you watch it and maybe you fix some small issues.
But in order for it to be ready to run automatically without you monitoring it, you need to harden it. And in the AI world, this matters more than ever. And here's Wang describing why.
You're dealing with models. You're dealing with something non -deterministic. You're dealing with agents.
They can break. If people have these unreliable experiences, they think it's trash.
And what he's saying is AI isn't normal software. The same requests can render different results. And we've all seen this ourselves.
But the reality is that the person on the other end of a bad output doesn't care why it happened. One bad experience and they just decide it's trash. So how did Wang fix this?
It wasn't with a better prompt. It was actually with a spreadsheet. And we had like these long spreadsheets of.
hundreds or even thousands of different, not thousands, but more than a hundred like specific behaviors and capabilities that we knew the model needed to have to be able to build the ideal personal agent. Like it was very, very meticulous. And then.
Over time, we built evals for all of those. We figured out how to train the model to be really, really good at every single one of those things. And we have these reviews where we look at that spreadsheet and we figure out the things that we're bad at and make sure that we're improving on them.
It really is a great explanation, but each row is one specific thing it has to get right and it needs to be tested. And an eval is just a test for that one row. The rows in that spreadsheet are based on the key things that they have to get right.
And then the person who actually identifies what needs to be in the row is the person who's best at it. identifying what should be tested. Now for you, those rows are what you want the product to produce and what the output should be.
Now, how much rigor do you actually need? Now, I'm not telling you to spend seven months hardening something. No, Muse is very different from what you're building, but it does depend on the cost of error in terms of how deep you go.
At Metascale, they set two bars on every row, what's launch blocking and what's good enough for go live. And here's them explaining how they knew when it was ready for customers. We had set for each of those rows, what threshold is launch blocking and above.
what threshold is good and we had many checkpoints that like were like green on 80 of the rows but red on 20 of the rows we knew that wasn't good enough and we had our first model checkpoint which was green across the board so there was like the very numerical part of it and then there was just the experiential part which is we tried the model again you're not building muse and most of you aren't at that level and frankly i'm not at that level either so instead for what you're building write down the top things it must do and make sure it hits those every time and if you're building something and there's a certain component where ai doesn't get it right perfectly remove that entirely one of the lessons i've learned through buildpartner .ai as over 5 000 users have gone through it is if you have an ai future that only produces a high quality output 80 of the time you're better off not having it at all because statistically that means every person who continuously uses your product will have a bad experience and that means that that single bad experience could lead them to not trusting the entire product so how do you harden what you're working on over time here is a simple version first ruthlessly cut scope second make a table with five to ten rows of things that you want to test third ask claude create a skill that tests this table and produces a report of what worked and what didn't fourth iterate until you're confident on those specific use cases and then
And five, good is good enough. It's time to launch. Now that last step that I mentioned, this one matters because the sheet and your planning, it's not gonna be magical.
When Muse launched, people started using it in ways that Meta had never thought of. For example, doing 3D demo games. And this was only apparent after launch.
And Wang described this as obvious in retrospect, we probably should have done a bit of that. And he's talking about testing. And ultimately, you won't catch everything.
So personally, I buy us toward getting whatever you're building in a state that's good, not perfect, so you can learn. from other people. So here is a single prompt that will help you harden whatever you're building based on the steps above.
Wang spent seven months turning a demo that sometimes worked into something he could put in front of billions of people. And we can do a smaller version of that for anything that we build. And yes, being reliable makes for a good user experience.
But there's one more step that makes the biggest difference. But before we get to that, if this is your first video of mine, welcome to the channel. Now, if this is your second or more, you know the deal.
This is our anti -slap agreement. The visuals, the testing, the hours of research that went this video this is entirely built for humans by humans so all i ask is you subscribe to help this content reach more people and if you're already subscribed make sure to sign up for post notifications that really helps the channel and also we just hit 100k subs thank you everybody for the support this is actually i never would have expected this so thank you everybody truly and as a thank you for all your support every video i give away a cloud max subscription and last video's winner is fix it with chris for building an asset management platform for this video comment below with what you're building a recent feature you built, or any problems you're running into.
You're all legends, and I look forward to chatting with you in the comments. Now let's get to step four, which is about making it memorable. The question you have to ask is why would someone care?
We've all built things that even we don't end up using. So if you're trying to get yourself, let alone anyone else to care about what you're doing, you have to make it memorable. Here's this problem described by Wang.
It felt important that we figured out a way to make Muse breakthrough. it's like this torrential stream of updates like for most people they're just like constantly bombarded with new ai products and startups and releases and here's another model and here's another product and here's another feature and so being good isn't enough anymore there's a new ai tool every single day and this isn't just a meta problem a tool for a client an automation for your team a report for a customer nobody uses it unless they care so what did the muse team and wang have to do he kept coming back to one question one thing that we came back to a lot in in both building music as well as just like building stuff in general at meadow was just like why would someone care people don't care about you or your company everyone is self -interested so nobody by default cares what you built they care what it does for them so wang did two specific things about it with muse first the wow moment how do you get people to match the experience like how do you get someone to to
A wow moment with the product as early as possible. Product teams call this the aha moment, which is the point where the user feels the value first. For Airbnb, it's the first booking.
A good onboarding gets people there as fast as possible. I personally call it time to magic. And here's actually a conversation that I had with my team about a feature we were building.
I said, what's the least amount of steps we can get for them to experience the most valuable and unique thing that'll get them hooked? That's the magic. And the second thing that Wang does is the screenshot.
This one is specific to scale and consumer. products but there are lessons in it so make sure to pay attention modern product marketing is like very screenshot driven like i think to have a modern consumer product like it has to work on the screenshot like the screenshot has to has to fly so to speak either through like group chats or word of mouth or like online in plain english if the result can't be screenshotted and sent to someone then it's not going to organically grow so packaging this together before you call anything done ask yourself what is the first thing someone gets from this in the first couple of minutes and would they screenshot it so how can you apply this framework depending on what you're building let's say you're building something that's literally just for you how can you reduce the friction to actually use the product so simplify how you actually interact with it and that's personally why I love cloud skills for automations.
They just live right in my cloud session, so it's super easy for me to use it. And if you're building yourself your own app or command center, make it dead simple to get up and running. And if you're building for other people, ruthlessly focus on onboarding.
Personally, I use BuildPartner's expert advice skill to analyze my onboarding and suggest improvements. And I also use BuildPartner's web research to identify what other people are doing in a similar space so that we can learn from them. Ultimately, the strategy is reduce the time to magic.
If it's reliable, it'll get used. But if it's memorable, it'll get shared. So let's just recap.
First, we write the memo. Then we build a proof of concept fast. Then we harden it.
And then we make it memorable. Now, before you start just ripping these memos, I mentioned a quote earlier that Alexander Wang said that I haven't been able to stop thinking about.
And this is about how AI can give everyone the adulthood they dreamed of in childhood. I'm going to let this clip run a bit longer than usual, but pay attention. There's a reason for it.
I think Muse can give everyone the adulthood they dreamed of in childhood. Everyone, when they're kids, they have like big ideas and big dreams. They imagine this life where they can like be an astronaut or they can save the planet or they can change the world.
And then by the time people finish school, they enter the workforce, all that hope and ambition and ability to dream has been sucked out of them. You kind of become a zombie. Maybe at some point you want to go tackle your dreams, but then it's like really, really hard because you have all these commitments.
You maybe have a family. You become trapped in this sort of like... restating the way I said it at the start.
AI can give everyone the adulthood they dreamed of in childhood. I just felt this clip at my core in everything that I've seen around me. I've had friends start businesses now who would have never done it before.
I know current business owners who are thinking bigger and bigger. As a kid, I dreamed of starting my own business, creating content, and helping others, and AI has helped make that possible. And pulling from one of Wang's posts, the world until now has been shaped by the small number of fanatics who have somehow found a way to make their wants real.
but we've never seen humanity with every single person's agency fully switched on and building on that concept. For all of us, we have a mix of things that we want to do and things that we have to do. And the promise of AGI is that we're going to spend almost none of the time doing things that we have to do because the AIs will just start.
doing a lot of that for us and we can spend all of our time on things that we want to do you're a kid you have big dreams use ai help make those things happen then like dream bigger and that's why these four steps matter they aren't just how meta ships a product they're how you'd build the thing you wanted to build as a kid everyone has mentally shelved ideas they parked because of time skill or cost ai changed all three of these a lot of what's on that shelf is now doable so here's the one thing to do right now write a memo for the most impactful thing that you want to build what it is who it's for what it's not, what done looks like, and the smallest version that's possible.
If you want the exact memo format my team fills out that you can fill out yourself, I linked it below which is entirely free. Now if you like this video, you'll love this video where I break down in detail the five questions you need to answer with Claude to build 10 times faster. Now if you're new here, my name is Austin Marchese.
Welcome to the channel and I'll see you over in the next video. Peace.
The Hook

The bait, then the rug-pull.

Meta's Chief AI Officer rarely explains how he starts a project, until Muse launched and he went on every podcast. Austin Marchese read and watched all of it and found four repeatable steps.

Frameworks

Named ideas worth stealing.

00:20list

Wang's four steps

  1. Write the memo
  2. Build the proof of concept fast
  3. Harden it
  4. Make it memorable

The sequence Marchese extracts from Wang's public comments on building Muse.

Steal forany AI project from automation to product
01:45list

The memo questions

  1. What is it
  2. Who is it for (a real person)
  3. What it is not
  4. What does done look like
  5. What is the smallest version that proves it
  6. What features do I want

Five questions plus one added by the presenter to bridge the abstract to concrete deliverables.

Steal forkicking off a Claude project
11:40list

Five-step hardening

  1. Ruthlessly cut scope
  2. Make a table of 5-10 things to test
  3. Have Claude build a skill that tests the table and reports
  4. Iterate until confident
  5. Good is good enough, launch

A small-scale version of Wang's spreadsheet-and-evals process.

Steal formaking an automation reliable
CTA Breakdown

How they asked for the click.

VERBAL ASK
18:05link
“If you want the exact memo format my team fills out, I linked it below which is entirely free.”

Soft closing pitch tied to the one action the video asks for, writing a memo, plus a pointer to a related video.

Storyboard

Visual structure at a glance.

open
hookopen00:00
memo questions
valuememo questions01:57
POC vs MVP
valuePOC vs MVP03:06
hardening
valuehardening07:56
memorable
valuememorable12:51
CTA
ctaCTA18:05
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.