The argument in one line.
The compounding value in AI-driven experimentation is not each individual iteration but the accumulated experiment log — once you can read a metric automatically, Claude Code will improve it indefinitely without human intervention.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code for tasks but have never pointed it at a repeating business metric to optimize.
- You run a community, email list, or landing page and want to test copy variations without manually A/B testing every week.
- You want to automate data collection from a platform that doesn't expose an API (Skool, Kajabi, custom dashboards).
- You're comfortable with IDEs and Python repos but haven't yet built an agentic improvement loop.
SKIP IF…
- You have no existing traffic or conversions — AutoResearch needs volume to produce signal.
- You're looking for a no-code solution; this tutorial assumes comfort with a code editor and CLI.
- You need a pre-built product, not a framework to implement yourself.
The full version, fast.
Karpathy's AutoResearch repo encodes a single idea: let an AI run controlled experiments on a metric, log every result, and compound the improvements indefinitely. The video builds this loop live — clone the repo, pick one measurable objective (Skool conversion rate), load your business DNA into the AI's context, then solve the hardest part: getting data from platforms that have no API. The solution is Claude CoWork browser automation feeding a Notion page, which Antigravity polls on a schedule. A Pinecone vector database pre-loaded with expert content (Hormozi, community surveys, your own transcripts) makes the generated copy specific rather than generic. The first iteration produces a "pain block" hypothesis with a full testing roadmap.
Members feature
Chat with this breakdown.
Modern Creator members can chat with any breakdown — ask for the hook, quote a framework, find the exact transcript moment. Unlocks at T2: refer 3 friends + add your own API key.
Create a free account →Where the time goes.
00:00 – 01:29
01 · Hook — the WD-40 principle
Loss-frame opener, host intro, WD-40 iteration metaphor as organizing frame for the entire video.
01:29 – 03:46
02 · What AutoResearch is
The three-part loop: generate → deploy → harvest. The log as compounding asset. Baseline vs. challenger pattern.
03:46 – 04:24
03 · What makes a great candidate
Objective metrics only — pass/fail, number, percentage. Anything with a 'vibe' fails the test.
04:24 – 06:27
04 · Repo setup in Antigravity
Clone AutoResearch from GitHub into Cursor; connect Claude via subscription (cheaper than API); repo structure overview.
06:27 – 10:14
05 · Step 1 — Choose the metric + context
Pick Skool conversion rate. Load context: community copy pasted in, business.md prompt, survey data. Claude asks clarifying questions.
10:14 – 13:57
06 · Business DNA and survey data
business.md template walk-through. Survey responses as voice-of-customer training data. Feed everything to Claude before first iteration.
13:57 – 15:46
07 · Step 2 — Measure: baseline data without an API
Skool has no public API. Strategy: screenshot + manual paste for first baseline. Claude asks what data it needs.
15:46 – 20:24
08 · Frequency + Pinecone knowledge base
250 events = stable window. Build a vector DB from Hormozi books, YouTube transcripts, community posts to make copy specific.
20:24 – 23:15
09 · Step 3 — Automate with Claude CoWork
Install Claude browser extension; give CoWork a task to visit Skool dashboard on a schedule, copy analytics, return data to chat.
23:15 – 26:26
10 · Pipeline: CoWork → Notion → Antigravity
CoWork writes scraped data to a Notion page; Antigravity polls that page; full closed loop without manual intervention.
26:26 – 29:34
11 · Business guide to choosing where to start
Decision flowchart: measurable? → retrievable? → frequent enough? Great vs. bad candidate matrix. Start where volume is highest.
29:34 – 31:20
12 · First iteration output + CTA
Claude generates v0→v1 hypothesis: pain block between headline and offer. Testing roadmap with priorities. CTA to knowledge-base video.
Lines worth screenshotting.
- WD-40 is named WD-40 because 39 formulations failed first — the iteration count is the product, not the accident.
- The log of past experiments is the compounding asset; each new Claude session inherits all prior context and never repeats a failure.
- AutoResearch requires exactly three things: one thing to change, one objective metric, and an automated way to read the result.
- Platforms without APIs are not blockers — Claude CoWork's browser automation can scrape any dashboard your browser can open.
- 250 conversion events is roughly the threshold for a statistically stable test window; below 100 and the signal is too noisy.
- Feeding Claude a vector database of domain expert content (books, transcripts, survey data) replaces generic copy with voice-matched specificity.
- The best AutoResearch candidates have high volume, fast feedback cycles, and numeric or binary outcomes — not 'does it feel good' judgments.
- Claude's clarifying questions at context-loading time are a feature: the specificity of the question predicts the quality of the output.
- You can run the data loop even when your laptop is closed — if CoWork misses a scheduled run, it retries on next open.
- A cookie-based scraper can replace CoWork entirely for certain endpoints, removing the need for a running browser session.
- The hypothesis Claude generated — a 'pain block' between headline and offer — was grounded in survey data, not generic CRO advice.
- Starting with the single biggest constraining metric (not a random one) determines whether the whole system produces real business impact.
Any metric you can read automatically is a metric Claude can improve.
WHAT TO LEARN
The AutoResearch framework's real unlock is not the AI — it's designing your measurement setup so the loop can run unsupervised.
- The accumulated log of past experiments is the compounding asset, not any single iteration — the AI inherits all prior context and never repeats a failure.
- An objective metric (a number, a pass/fail, a percentage) is a hard prerequisite; anything that requires a human judgment call will stall the loop.
- Platforms without APIs are not blockers — browser automation can scrape any dashboard a browser can open, then relay the data to the agent via a shared document.
- 250 conversion events is a reasonable stability threshold before drawing conclusions; below 100, variance will mask real signal.
- Loading domain-specific context (expert books, customer surveys, your own transcripts) into a vector database moves AI-generated copy from generic to voice-matched.
- Start your first AutoResearch project on the highest-volume, fastest-feedback metric you have — the system compounds fastest where data is most abundant.
- Changing one variable per iteration is the discipline that makes results interpretable; radical redesigns reset the log's accumulated signal.
Terms worth knowing.
- AutoResearch
- A GitHub repo by Andrej Karpathy that structures AI-driven iterative experimentation: generate a variant, deploy it, measure the result, log it, repeat. The log compounds as the primary asset.
- Baseline / Challenger pattern
- A controlled testing structure where the current best-performing version (baseline) is held constant while an AI-generated variant (challenger) is evaluated; the challenger replaces the baseline only if it wins.
- Claude CoWork
- A browser extension and desktop integration that lets Claude autonomously perform actions in a web browser — clicking, scraping, filling forms — on a user-defined schedule, without requiring an API.
- Antigravity
- The presenter's name for the Cursor IDE (an AI-native code editor); used interchangeably throughout the video.
- program.md
- The single markdown file in the AutoResearch repo that serves as the standing instruction set for the agent — defines the metric, the change process, and the loop frequency.
- Pain block
- A short copy block positioned between a headline and an offer that names the reader's specific frustration, used to increase self-identification and reduce bounce before the pitch.
- Vector knowledge base
- A database (here, Pinecone) that stores content as numerical embeddings so an AI can retrieve the most relevant passages by semantic similarity rather than keyword match.
Things they pointed at.
19:00bookAlex Hormozi books
30:40toolGlider (speech-to-text)
Lines you could clip.
00:26
“WD-40 is called WD-40 because WD-39 failed.”
02:38
“The log is the asset.”
14:00
“The question you ask is more important than the answer sometimes. In fact, most times.”
Word for word.
00:00If you're using Claude Code without auto research, you might be leaving money on the table. So in this video, I'm gonna show you new strategy from Andre Kapathi that could not only just improve any metric in your business, but also save you hours and get you light years ahead of your competitors. And if you don't know who I am, my name is Jack Roberts.
00:16I built and saw my last tech startup for the 60,000 customers, and now I run an AI automation business. So if you haven't already, grab that coffee unless I've straightened. Bloop code auto research can be best explained by w d 40, the tool that every man in the world needs.
00:31Now what most people don't realize is it's called w d 40 because w d 39 failed. They had to do 40 different iterations of something before they got the actual winning combination. The TLDR here is that the more we iterate on something and try and experiment, the better the thing actually can be, which can drive incredible benefits to our business or personal lives.
00:51For example, if we work on the copy and the positioning, we can get higher email conversions. We can do the exact same thing on our websites, better title, better blogs, and get more sales. Likewise, if you want people to join your community or you wanna increase your click through rate on your social media or get better hooks or even sales scripts where your team are selling constantly or your clients are selling constantly.
01:11You literally name it. This process of iteration gathering data and making better decisions is what helps us to generally improve in anything we're doing in business. So what is it and who the hell is this guy?
01:20Well, this is Andre Kapathi, one of cofound of OpenAI. He also worked in in I think it was a Tesla AI as well. Essentially, it is a GitHub repo that Andre has created that essentially enshrines some of these core principles of this sort of self improvement cycle.
01:32It's blown up in GitHub. It's got many, many thousands and thousands of stars quite recently, but effectively, it has a couple things. We just need something to change.
01:39We need an easy way to measure that thing, and then we just kind of let our cold code agents run it and report back to us on a specified frequency that we actually give to it. And to do this, we need three wonderful things.
01:50First of all, we need one thing to change. We need one metric to measure, and we need a way to read the result that it gives us. So the way this works essentially is a loop is we generate something, we deploy that thing, and when that thing's deployed, then we grab the data.
02:03We in effect harvest the data. The AI are gonna do this on a loop. It could be every two minutes.
02:08It could be every day. It could be whatever frequency we want to. But that's the kind of triangle loop that we create something, we then see its performance, and then based on performance and data, we try a new experiment.
02:18And there is that the actual quality of the thing goes up over and over time. And if you think about it like this, day one, it's like you're pushing, like, know, a boulder up a mountain or a snowball up mountain. But the core idea here is that every iteration compounds and every experiment is logged and it never forgets.
02:32So we'll already know what we tried and didn't work. And so by day three six five, you've got all these beautiful success that you could never have gotten before.
02:40Remember, 1% improvement. If you improve a 100 by 1% every day, I think the total output is, like, over 300. It's really, really impressive.
02:47So log is the asset. Essentially, if you think about it like this, right, hand over from v one to v two. Model v one compiles initial research and context, but then passes over to v two.
02:55So v two optimize, enhance, and builds on knowledge, then v three does the same thing, and we keep giving this playbook of excellence together so we can actually figure what the optimum is. Now if you think about this, the way that it essentially works is that we want some kind of control, a control sample if possible.
03:10This won't always be possible, and I'm gonna show you a circumstance in which that might not be possible. But a typical idea is that we have two groups of things. We have a baseline control, which is a blue card, and then we have the challenger variant that Claude code will create for us.
03:21And the idea with this is if the challenger variant beats the control, it becomes the new baseline, and so on and so on and so forth ad infinitum, as they say. Now, what makes a great candidate for this auto research process?
03:34Well, it needs to be something that is objective. In other words, there can be no deliberation or debate about whether or not it was achieved. So does it feel good?
03:41Does it have a certain kind of vibe? Not great. Is it a metric?
03:44Is it a conversion rate percentage? Is it something that's a one or a zero or a number or a pass or a fail? These are things that make it, but fundamentally, needs to be an objective metric, and I'll touch on that in the video as well.
03:54But also show you how you can actually come up with what you should be focusing on the very first project to do for your business or your personal life. Now anything measurable can be optimized essentially. So it could be landing pages, email subject lines, ad creative, YouTube titles and thumbnails, ecommerce checkout flow, chatbot scripting, and the idea is that we build up all these different systems to run through as a parallel so we can improve our entire business and workflows.
04:14And some very simple steps, we're gonna clone the repo, write a program dot m d, define our metric, and then set the loop. And even if the thing that you wanna track does not have an API, we're still gonna cover that and show exactly how to do it, because APIs are not gonna stop us here on the Jack Roberts YouTube channel.
04:29Wonderful. So now I'm an anti gravity, which is my IDE of choice, my integrated development environment. And then once you double click in the middle, you'll see an orange thing appear at the top.
04:37If you don't have Claude installed, you can come on the left hand side to I believe it is just under extensions. You type in Claude, it's there, and then you download it, and you're gonna be good to go. So we can choose how we wanna connect Claude.
04:47I'm gonna do it via my subscription because I've got a max plan. You can do it with the pro plan as well. We'll save you on credit, so let's get that set up.
04:53We say, yes, Claude. That is absolutely fine. We're gonna authorize you.
04:56And once you've done that, you'll see Claude appear on the right hand side. And by the this is by far the cheapest way to access Claude code is just get your pro plan. You don't need a max plan unless you've taken it into Stella, which I think a lot of you are because you're watching this video and you're kicking butt.
05:08But ultimately, the pro plan is gonna be so much cheaper for you than API. So this is the GitHub that we're gonna copy in there. And it basically explains auto research, how it works altogether, and what it does.
05:16So the repo is deliberately kept small and only has three files that matter. Prepare dot p y, fix constraints, one time data prep, train dot p y, the single file that the agent edits, and program dot m d, which is based on instructions for one agent. Point your agent here and let it go.
05:30And it's got quick start instructions, a load of wonderful thing. The best way to integrate this, as with all these things, is come up with code. And we're just gonna copy this, and time recording has over 44,000 stars.
05:40So let's go back over then to anti gravity, and then let's give it here. We're gonna say, hey, that. What I'd like you to do is go ahead, grab this repo, and clone it into this project folder, please.
05:48Cool. So it's basically just gonna go straight over to GitHub, grab that, and drop it in. We're using glider for speech to text.
05:53And if I come over here, can see now in the explorer section, we're gonna have everything's gonna be added. So what we first will wanna do is just open up a folder, so we're gonna be rocking and rolling. You can also just climb a repo here, the way, on the left hand side if you want You literally just put that down there, climb from this URL, and it will just automatically do it.
06:08And then we need to pick a folder, of course. So if you come to any folder you want to, I can come down here and I just say auto improve. Okay.
06:14Cool. And then hit create. And then select the repo's destination, and then it should literally climb the whole thing.
06:19Yep. We're gonna do that. We'll open that one up, and then then we have everything here rocking and rolling.
06:22It's an alternative option if you don't wanna just yap to Claude. It's possible. As you can see, now we have everything there for us instantaneously on the left hand side.
06:30Now this is the anti gravity window, which we can remove by doing control and l. You can speak with Claude if you want to. Within here, it's completely your call, but we're gonna be using Claude within the main section here by clicking in on this guy, which is wonderful.
06:41So now we've done that. The first thing we need to do is essentially decide what is the thing that we want to track. Now there's many things you could do.
06:48You could do things like email conversion rates. We have different split up things. Now to apply these incredible auto research principles to your business, you only need to follow a very simple three step framework that will essentially work for anything that you want to track and improve.
07:02The first one being actually choosing an important metric to change in the first place. And number one is something to change. So let's take an example of a platform.
07:10Now let's say for instance that you were on school. Now many people are on school these days. So let's say that you were on here and you wanted to track your conversions, and you thought, actually, one of the things I wanna do is actually see what my conversion rate is.
07:22So let's say you have copies somewhere. Again, this could be on your website. The example doesn't actually matter.
07:27All that really matters is that you're picking a thing. So you may look at this copy and say, hey, this copy is cool. But what if you could trial this on a weekly basis and have AI actually use this exact same auto research principle to trial out different copies based on all this information?
07:42Well, that would be pretty cool. But in order to do that, we need a metric. We need a conversion metric.
07:46But the problem is not everything in the world connects through an API. If you're connecting to Go High Level or an email service, we can grab that. You can't grab that via score unless you use the strategy that I'm gonna show you in this example, which is why remember, the example doesn't matter, it's how we actually do it.
08:01So step one is basically decide the metric, which is going to be the thing we're actually changing, which is going to be this copy. Okay? I'm just gonna copy all of this like so, and now we're in back into anti gravity.
08:10The first thing I wanted to do is to familiarize itself with the information on the left hand side. So hey.
08:16The first thing I'd like to do is to familiarize yourself with this entire project on the left hand side. And the spirit of what I'm trying to do, we are going to spin up several projects to track systematically improvements of a key metric in my businesses.
08:30And, essentially, we're going to dictate what the thing is that we're tracking. I'm gonna give you context about what that thing is. You're going to ask me clarification questions, and you're gonna follow along with the instructions on the left hand side.
08:40You're also going to go and create a Claude dot m d file. And once we've done that, we're gonna create a unique process for each of the things that we're going to track. The actual thing we're changing, the metric we wanna get to, and how what the process is of us actually improving that over time.
08:53So familiarize yourself with the items left hand side, and let me know when you are ready to proceed. Awesome. So put this one in here like so, and then let Claude initialize itself with the information.
09:01Now this is wonderful. It's fully initialized. And remember, when you're deciding what you wanna change, I will come on to that in this video.
09:07You're not gonna wanna miss that part of the video because it's really important, is we typically wanna pick the biggest constraining factor. What is the one number that exists in our world that if we were to actually change, it would have the biggest impact on our business? I've chosen a random example, but you'll get what I'm talking about.
09:20So now it's ready. We're gonna basically decide in the first step what the thing is and give it, crucially in this step, context. Because without context, it's gonna take way longer.
09:29Now, obviously, if you and I sat down and optimized this every single day together, we'd probably get there just as quick. But the idea is that we can automate more of it by bringing in Claude code here. So here's what I want you do.
09:37I want you to optimize the copy on my school landing page. Your objective is to increase one number, which is the conversion rate number. The details that you're going to get fed into are gonna be number of clicks and actually the conversion on those clicks.
09:49Your job is to actually change the copy on the homepage with the view of improving the conversions. Let me know all the specific data that you would need in order to get that. So for example, this might be you know, I might be online, it could be Google, it could be YouTube, it be several places that will be mentioning me.
10:04Your job is to understand what is required to move that number up. And now you've given it this context, what we have to give Claude for this to work properly is going to be our business brand. And we call this business.nd.
10:14Now there'll be certain things that you can optimize with little context. But in reality, the more information that Claude has, the better it can be optimizing that. For example, if it's doing emails, it'll be great if it knew a lot about your customer.
10:24Right? So we wanted to basically grab a big beautiful white pad, a great coffee, and fill this up. It goes through things like your business DNA, the machine who to target customer, how old are they, what are the problems that you solve, key metrics.
10:34I've spent a long time going through building this. I've experienced of helping thousands of entrepreneurs. You don't need to do this from scratch.
10:40Just take the things I'm giving you out. All three free down below links. Just grab them and have fun with them.
10:45And then part two is about you, your story, who you are, etcetera. So fill this out. Take the time, invest in yourself, get this done.
10:51And when it's done, what I want you to do is head back over to anti gravity. And when you're anti gravity on the left hand side, come over and create a new file. And this file is gonna be called basically business dot m d.
11:00Okay? Beautiful. Come down.
11:01We'll finish that one out. Great. And then I just want you to drop all of this information into it.
11:05Okay? About who you are, what it's about. Now when that's done, you're gonna come over here and say, I'm now gonna give you additional context about who I am and my community.
11:12I'll paste this down below. And in addition to that, and I have business.md context on the left hand side, you must reference this when creating a strategy and actually building out everything so you've got context to improve.
11:23Now once you've done that, that's wonderful. And then we're just gonna drop some interesting context based on the length. That could be on the left hand side, that could be on the right hand side.
11:30So if you do surveys, for example, this is a wonderful place to do it. So we often survey people in our community like, hey, what do you love? What do you dislike?
11:38How can we help you? How can we be better for you? I work for you type of thing.
11:41Right? And that's great data. We have hundreds a 100 people currently participate in that.
11:45And so I can use that then to actually feed what the new copy might look like. So now I'm gonna give you some context about the group. And so we wanna do is create a new folder per project.
11:53This one's gonna be school. Okay? And within school, we're gonna put all of our core data in there.
11:56Okay? So let's click on school. Let's click on a new file.
11:59And this is going to be about page. Okay? Dot m d.
12:02So here, we are going to basically put in all the information. And so the key part of this auto research thing working is gonna be data. So this is an example of a survey that we did.
12:10We asked lots of questions. How can we help you out? What's great and blessing?
12:13And we've had hundreds of responses on this, which is cool. And, essentially, now what I can do is take this data and work with Claude to understand, like, what are the great things that we can do more of that we should probably shout more about inside the community? Like, is it meetups?
12:26Is it the fact that we see you know, what is the thing that they're that you really love? So what we can do is download this data and give it to Claude as contacts to help us improve and make sure that we talk about the things that people really care about. So in this instance, again, this applies for anything.
12:40This could be website data. We would just be doing the same thing, but extracting it from WordPress or GoHighLevel. Honestly, the thing doesn't matter.
12:46It's just about the process and following the three steps that I'm showing here. So in this case, we will just simply download the data. And as I bring that in, we've been prompted to install Rainbow CSV.
12:55Honestly, I have no idea what Rainbow CSV does, but apparently it's great. I'm on a new Mac, by the way. Hence why I'm having to download all this stuff.
13:01But here's the key thing. We've got the data now in there. So then we've given this.
13:04I'm just gonna say hey. We've given it a particular folder. For this particular project, I have added a folder on the left hand side called school.
13:10Please use that and review it as basic context. And then the next step is we're gonna agree on the frequency and actually build up the automation. Let me know if you have any further questions in terms of what the expectations of you are and how I can help you do the best job possible if the objective is to increase the conversion rate.
13:26And guys, check this out. So let's come back here and give us some interesting insights. Let's look at this.
13:30The data I need from you is current metrics. So what are the actual numbers from school right now? What are number of clicks?
13:35What is the current conversion rate? Where's the traffic coming from? By the way, you get all this data from school.
13:39Sam Evans is doing a wonderful job. If you're watching Sam, keep up the work. I know he subscribed to GI channel.
13:44This guy's on it. Traffic sources. Where will people be coming from when they land on this page?
13:47YouTube videos he take. He was asking for the context. Isn't that fascinating?
13:51Then we've got the page URL. Can I access it? Now I'm gonna come to that in a second because, guys, you better be buckled then because I'm gonna show you some really cool stuff in a minute.
13:57Business.md, can you fill this out? Because I gave it a sample one for the purpose of the video, but you get the idea.
14:01Then conversion tracking setup. How many currently measure how many how are you currently measuring conversions? School built in analytics, blah blah blah, and then price.
14:08Is it generally a free trial? Yes. It is.
14:10It's only in a few hours. What's the paid price? Blah blah blah.
14:12So this is cool. And then what's really key, I just wanna double click and show you this actually, is that the 184 responses give me clear signal on the top reasons people join, their primary pain points that I can use as a voice of the customer foundation for copy testing.
14:25When she gave me the above, I proposed the optimization loop structure, how frequently to test, and how to change iteration, etcetera. So to finish off step one, we now need to give it the data. Bear in mind, we're not saying here that we're 900 IQ.
14:36We're asking questions. Remember, it the question you ask is more important than the answer sometimes. In fact, most times, that's true.
14:41So what wanna do now is actually give it the data. So then here, we'd head back over to our platform, and we'd get the data, and we would drop in the historical data. And school gives you that over time.
14:49And so if I come over to a community just as an example, this is a test one that we created. We're having a bit of fun for Glider. And you can see it basically breaks down everything.
14:56You can scroll down, and you can see page visits over time. So you get this level of data. But there's one problem.
15:01Skool does not expose this data to an API yet. So I'll tell you exactly how you solve that in a second and how you can solve any issue you've got regardless of what the platform is that you're using. Obviously, if you've got an API, you can just make the API call.
15:13It's super easy. But this is when it's not that easy. So we have all this historical data.
15:16So what we would do is come down with a control five, control f five. Actually, somebody in the comments said, if you control f four, you get to do this. Took about a tip of the day.
15:24That's cool. But the point is we then would screenshot the data. We would get the historical data so it has what was known as a baseline, then we can provide that back over to ClawCode.
15:33So then to complete the section out, just give it all the contacts, answer all those questions, and then you're gonna be ready to move on to one of the most exciting steps in the entire process. This very nicely takes us on to the second step, which is all about measuring. Now the question is, what should we measure?
15:45Now ideally, this process works best when we can do high volume and incredibly high levels of iteration cycles. If I could do a thousand a day, do the next day, do a thousand.
15:55That is where this is in its golden. But in reality, in business, not everything is 10,000 tests a day and the next day we can do it. So we need to use some decision making heuristics to make our lives easier.
16:05Now as a general rule of thumb, around 250 conversion events would be considered stable.
16:12If you're looking at things like your brand value, that is something that you can't you track that over months, not days or weeks. So for something like, I say a school page, if you're getting thousands and thousands of visits and lots of conversions, I'd say roughly 250 ish sales is a good enough time frame window that you'd be comfortable.
16:29500, you'd be more comfortable. A 100 would be a little bit more unstable. So with that in mind, we can now start to look at the frequency and how we track this to make sure that you're actually getting the right data.
16:39Now in this step, what we're also going to do with the number is we're actually gonna have a look and do some research. So if I show you something in my last video, I showed you exactly how you can vectorize anything you want to. The example we actually used was Alex Homosy.
16:51Okay? So this is one of my apps. Let me come over and show you this.
16:53If I come up to YouTube chat, and you can see here we got Homosy about. Right? So you can ask Homosy about anything like, I don't know, something like, hey.
16:59What is the value equation? Remind me what that is, please. Whatever it is.
17:02Right? When you ask us a question, what this effectively does is this scrapes every one of Alex Hormozi's YouTube videos every day, and it has all of his books already preloaded into context, which is why he knows stuff like this. It also sounds like Alex as well.
17:14So what's interesting here is I've got this. Now if I'm doing anything copy related, I wanna feed my AI the very best examples of this. If you want to know how to do this, I'll put a link for a video on screen.
17:24I recorded it like two days ago. It's just like literally the video before this, Shows you how to do that step by step below. I get resources, I get bugs I can check out with information.
17:32Really cool. So I want my AI to have the ability and the specific domain knowledge to do the thing. So how do we accomplish that?
17:39Well, you can build a custom knowledge base like I have, and the way that we did it here was on Pinecone. Again, I'll click the link so you can check that out. And then we wanna feed that to it so it has the right level of knowledge to just not be generic.
17:50We want it to be super powered. So I've just signed now into pinecone. Which is wonderful.
17:54And all we need to know is the index name, which is going to be this one here. Great. We'll copy this guy.
17:59As you can see, you can even ask it questions here to test it, by the way. And then you're just gonna wanna come on the left hand side, API keys, create a new API key, and then provide that to anti gravity, which is dead straightforward. And if you feel sensitive about putting your API key into a chat, you can just say, hey, list it as an environmental variable, and then just paste it in there if you want to after the fact.
18:14And then what I'm gonna do is ask you some research as well. In addition to this, I'd like to do some deep think research on the whole topic to actually understand what it is that is the best thing for conversions and what that looks like, and if you need any further context for anything with inside the community. Again, this is more about how you approach a problem than the problem itself.
18:31And because I miscalculated that, I'm just gonna go ahead and grab the transcription from the app and then come back over and drop that in. Guys, and you know what's really cool about this? It's also identified that I've got indexes there for my school community, which has things like posts and clashing materials.
18:42So now it knows that. It's even found better ways to help me improve and articulate what we're doing in an even better extent, which I think is wonderful. Remember, context is king in this game.
18:53And so at this point, it's not texting the queries, asking it questions, you can get images, multimodal inputs and outputs, all that wonderful thing. So you can see it can pull directly from anything it wants to. Now it's gonna pull together the scripts, so we built in the process.
19:04So we're not just letting it loose and say, hey, make me better copy. What we're doing is giving it context to help improve it, which I think is a really wonderful thing. And then the strategy we have, if you think about it like this, is we essentially build out these folders for all of our mini tests.
19:16And you can be the one if you want to to update it and change it and track it. But ultimately, we can let AI just actually automate that process for us. But again, the shorter the testing window, generally speaking, the faster we're gonna get results.
19:27But in a perfect world, which we don't live in at the moment, and we're progressing that way very quickly, is that we actually have this doing it regularly. But in reality, it can be at whatever cadence we want it to be, which might be weekly or fortnightly in this case.
19:38And as you can see, now it's gotten all the information and data that I have. So it's got Homersi books to give a sales framework scarcity urgency offer framing. It's got my YouTube comments, so all you beautiful people and you beautiful person that drop a comment below, I read them all.
19:50Thank you so much. I appreciate you. We've But got your language on, your your beta language.
19:53We've got my transcripts, so this very thing that you're hearing right now. We've got transcripts from basically online. In addition to that, we've all got everything else inside the community, which is the voices of the people inside the group.
20:04And then based on that stuff, it's now doing deep research to supplement it, to give it a little bit more context, and then we're ready to rock and roll. And so now we've locked down what we're measuring, and we've got all the data. The last part of the process is to actually decide exactly how we're going to automate it.
20:17And if you don't know how to automate something without an API, your options are gonna be so limited. But with this strategy, we can work around that effortlessly. And guys, to do this, we're actually gonna use Claude.
20:26Now here's an interesting distinction. If you can do it with an API key, just in the same way that I did it with Pinecone when we was talking to anti gravity, you'll literally come over and say, hey, this is the endpoint that you're gonna query to grab the thing. You're gonna say, look, all the data is in here.
20:39Can you just go and grab it if I give you the API keys? And most of time clause is gonna be like, But it can't do that all the time because school does not expose that as it's just one of many examples. And this is so cool.
20:49Now to do this, we're gonna have Claude CoWork do autonomous tasks for us in the browser. So the first thing we have to do is download Claude CoWork to our laptop, which you can do by downloading the Claude app at claude.com/download. And when you do that and you come over, you're gonna have access to all the Claude features, which are chats, co work, and code, which is awesome.
21:06So let's head over to Chrome and download the whole thing. So I'm in Chrome. I'm gonna come over to extensions on the right side.
21:11It's actually great that I've got a new, like, MacBook. So I can actually download everything with you as we go through this thing. So we're gonna full screen this one, and we are looking these are current extensions.
21:19So we wanna come over to the extension store. So we're going to visit Chrome Web Store, and then we're just gonna type in Claude at the top, and then download it. Okay.
21:26There we go. Claude, don't be put off by the rating. I think they did an early release at some point and their ratings got trashed, and this was a long time ago.
21:33Now it's really, really cool. So now Claude is alive and well in our browser environment, which is fantastic. Yes.
21:38We're human Cloudflare. Thank you for checking us out. And then once that's installed, guys, come over here, come to the dots, click on extensions, click on manage extensions, and then I want you to click on details on Claude, and then I want you to come out and pin it to the toolbar.
21:51So it'll appear there, and then you can basically tag Claude in and ask her questions. First thing we need to do though is sign in to log in with your Claude account. And then once you've installed it and validated it, you will see it now on the right hand side.
22:01We can click on next. Claude has group type access. That's fine.
22:04Awesome. So then if we go over to a website, so for example, if we come over to school or whatever website that you're using, let's just load that up real quick, and you'll see I should be on the Aspinall way, Tom Aspinall's school community, which is cool. So if I click on Claude, I can actually now ask him things.
22:19Right? So I could say like, hey, what is the top rated post on this page? Okay.
22:22Just as if for instance. Right? Now, basically, what we can do here is interact with Claude on a web page.
22:27Yeah. We're fine without gonna approve the plan. That's absolutely no problem.
22:29And we don't wanna ask for acting. Let's just act without asking for this case. Then it will assess the page and give us some information.
22:34And as you can see, it's identified the top page. Cool. So now we have Claude in the browser.
22:38Now the cool thing is that we can activate this via Cowork in our own actual desktop to do browser automations for us, which is really cool. So then I'm gonna come over to Claude Cowork and give it the following task. Like, hey there, what I would like you to do is set up a schedule task to go over to school.com to my own community at the URL I'm gonna give you down below.
22:55Once you're on that page, you're going to do the following steps. Alright. And then what I'm going to do basically is just list out everything you want Claude CoWork to do.
23:02So let's go over to school and actually pick up the specifics, which in this case is just coming to settings and then clicking on dashboard within the settings. So now we know that, let's copy the URL and head back over to Claude, then we're gonna down here. I'm gonna say, I want you to click on settings, and then click on dashboard, and then I want you to copy and paste all the information that you see and bring it back to me within this chat call.
23:21So this step one is basically making sure that you can get the data. Now as you can see, it's come down asking Jack, how often do you want this data? So for this purposes, I'm gonna say I want it weekly, which is great.
23:30And before setting up the schedule, let me try and attach them right now. Yes. Go ahead and test that for me, please.
23:34Wonderful. And just like that, without me touching a single thing, it's gone ahead and done it. I've got overview members, we've got financial metrics, growth, where our assignments come from.
23:42And again, this is all done automatically, and now Claude is basically saying, hey, what time of day do want me to run this weekly since New Year time zone? Wonderful. Goodbye.
23:50I'm thinking of something like Sunday morning to kick off the week. So we're gonna say, great. Do it at say a specific time.
23:55So it may just say, hey, let's do 9AM Sunday morning. And then here's the key thing. How do we get the data from this over to anti gravity?
24:02Well, this is where the connector comes in. Because what we're gonna do now is connect this to, like, a Google Drive. We're gonna have a sheet where this updates that data into a sheet, and then anti gravity every day can reach over super latest metrics are, and now it's fully connected together.
24:15So cool. So now let's actually connect it with Google Drive so we can con schedule. That is awesome.
24:19Wonderful. So now we just need to connect it to Google Sheets. And then it doesn't have to be Sheets.
24:23It could be Notion. So for example, what I could do here is come down and give us information. Hey, dude.
24:27What I'd like you to do is to create a brand new Notion page. And all of the data that you get from my weekly school dashboard, I want you to put into a Notion page. Give me the URL and the name of that page so I can pass it on to a different system.
24:39And then I want you to go ahead and update your scheduled tasks, please. And then we're coming back over to Claude with an anti gravity, and I'm gonna say, hey there. You're gonna get your weekly dump of data every Sunday morning at 8AM UAE time.
24:52It is all gonna be available and updated in a Notion page. You are on a regular basis every Sunday to go over to that Notion page and extract the information and build a running log on the left hand side within the school folder for data trending. Let me know everything you need to connect to Notion so that you are able to do that.
25:08And again, I know how to do that, but the reason I ask is because you're gonna have connection intricacies of every single API point. Right? But when we can just ask Claude, hey, this is what I wanna connect to.
25:17How do we do that? And then it'll literally tell us the specifics. And then we'll have a full pipeline where Claude will go and do the scraping, and then this will connect.
25:23The only watch out though with this is that your laptop does need to be open. But the really cool thing is that if it is closed, once it opens, it will try again. So some of the email automations that I have in terms of creating drafts work really well because I have multiple set throughout the day.
25:38But in addition to that, if it fails for some reason, it just tries again. So we wanna build that laxity into the system so that it doesn't fail. Actually, one of the reasons why I bought a specifically a Mac desktop is that it can be running twenty four seven in the background, whether I'm at the beautiful coffee shop or And again, Cloud Code Work itself will ask follow-up questions.
25:56So which schedule task do want me to update and which should I change to make? So basically, we're gonna specify now. So the task that you need to update is the only one that you've created, which is the scraping of the data on a weekly basis.
26:06Okay. It's already doing it right now, which is fantastic, but we'll queue that up. Little side hack as well is there are some circumstances where you don't even need called co work.
26:13You can get a cookie for the particular browser, and that's just one of things that we we do in the community. But if you get a cookie for something, you don't even need called co work to work in the background. And the way that we set that up is it basically just hits on a regular basis, and it's like a job that exists in anti gravity.
26:26And whenever you open your laptop, it will just go on repeat, which means that you don't even need called co work for certain things. If you know the magic buttons to press, but this is a great way to do it and unlock so much capability. So now it's done.
26:38We're gonna actually go ahead and grab the page URL because that's the only thing that anti gravity needs. Let's come down and click on update. Cool.
26:43So it's all done. It's just recommended that we do a run now. So let's go ahead to our schedule tasks and do that.
26:48To come over, let's click on, it's running every Sunday night before. Let's come down and just do this one real quick and click on run now. Okay.
26:54So that's doing it in the background. Also, did you notice, by the way, one thing here is keep awake is really cool, which and I explained this to us. When enabled, call will prevent your computer from going to sleep.
27:01So that's if it has to run. So if you don't want to commit to go to sleep, we can get Clawd to keep it awake, which is wonderful. So let's come back over to anti gravity and give it that Notion link.
27:09So we come over. Let's just drop that one in real quick. Cool.
27:11So I've come into anti gravity. I've given it the Notion link, so now it can go ahead and access the information, and it's gonna test to see if they can get the data. Now we can see the Cloud Cowork is doing its magic in the background.
27:20So this is landed on the Aspergill Way. No idea why it keeps doing that. Props out to Tom, heavyweight champ, and now it's heading up to my community.
27:26It's gonna grab the data, get it, throw it into Notion, and then anti gravity can access it and begin its copy changes. Now, one key thing that I want to really double click on here and explain is that generally speaking, if you have something that already works, we want to test and change one thing only and do micro changes at a time.
27:43You can do radical tests. Obviously, it depends on the nature of the thing that you're actually assessing. So guys, Alannan, as you can see, it's got all the information.
27:51It's got MRR, engagement, visitors, growth, everything we need. So the question now is, has antigravity accessed it? So let's go and check out.
27:57So we're gonna come over and say, hey there, all the data is updated. Go ahead and check it. And as you can see, anti gravity confirms the data is live and it shows everything down below, which now means it's gonna go ahead and do those changes.
28:07So the only thing left to do is get it to do its very first iteration. And we're gonna actually couch it by saying we want a small change initially, not to change the entire thing if we do the conversion of the crates. Now I've just asked Claude to go ahead and generate our first copy iteration, which is gonna be interesting to see.
28:21And whilst it's doing that, let's go ahead and have a look at how you actually know what you should be focusing in on first, because this is gonna be super handy for you. So your guide to choosing. Essentially, what we're looking for, the the highest performance here are things where we have high volumes and very clearly defined metrics.
28:36Think of emails, think of sign up pages, stuff like that. Your kind of bronze circle is gonna be medium volume, good metrics, and kind of get data quite quickly. But as you can see, I've detailed some things on the left.
28:44Right? Medium volume, good metric, daily feedback. The the kind of last hits are gonna be lower volume and slower feedback.
28:49So you wanna start center where you have a lot of volume. Now just to give you a slightly different view of where you should actually start with this. So what should I put first?
28:56Does it have a measurable metric? If we can't measure it, we don't do it. Can you get data via API or scraping?
29:02So that's using the strategies and techniques that I've showed in this video. If you can, awesome. If not, we're gonna use a bit of manual flow.
29:07And bear in mind, this one should almost always be yes. Almost always. And if you can't think of how to do it, ask me and we can figure out.
29:14I'm pretty sure. Like, the thing is guys, we're essentially only really limited by imagination at the moment. And finally, does it run frequently enough for fast feedback?
29:21In other words, do I get an update once a year? So again, this really is like your quality metric, and then obviously that will lead to quite a perfect candidate. Now with that in mind, one thing just to think about what actually makes a great candidate for this.
29:31So good candidate is, you know, it's something like emails. You get high volume coming through.
29:36It's something like landing page CTR. You can split taste on headlines. Bad candidates, again, are quite subjective things, things with their metric.
29:43SEO rankings are fine, but again, that's quite slow to do. Although technically, you could physically do that. That's really just a performance indicator at the end of the day.
29:52And then finally, the point here is can a computer read the result automatically? If Claude can't read it and then make an informed decision on it, it won't be able to do it. But the level of things that you can do with this are so impressive.
30:02It literally should be game changing. Wonderful. And so now we've got this pain block induction.
30:06V zero, v one is basically creating what's called a pain block, which is identifying the pain points of the user, which is interesting. This would actually take us over the total number of characters, so I need to feed that into the system so it's aware of that. But this is incredible.
30:18And if you come down to hypothesis here, it's really interesting to look. Adding a three line pain block between the headline and the offer will increase conversion rate because warm YouTube viewers already know Jack and believe in AI automations. Their objection isn't trust, it's self identification.
30:32They need to see their exact situation. Interesting. Per Hermoji, visitors who see themselves in the pain block read the rest of the page, visitors who don't bounce.
30:41And then we've also got a testing road map here with basically the element and then the priority. So this is the changes it plans to make based on the current data, pain block, urgency block, social proof, headline, etcetera. And the way we can take this to a whole new level is actually feeding it additional data points.
30:54For example, we could feed it image data, and it could start to make suggestions based on that. But you can see exactly how you can basically apply this strategy for anything that we're doing. Now knowing how to leverage the auto research skill without knowing how to build this database of knowledge from your chosen experts is going to severely limit your skills and capability with this, which is why the next thing we need to do is build out a knowledge base in a system so you can use this and supercharge results, which we're gonna do by watching this video right here.
The Hook
The bait, then the rug-pull.
Andrej Karpathy named his iteration framework after WD-40 — a product called WD-40 because WD-39 failed. In this tutorial, Jack Roberts picks up that logic and wires it into a live Claude Code pipeline, turning a Skool landing page's conversion rate into the first test subject of a loop that, theoretically, never has to stop.
Named ideas worth stealing.
01:29model
AutoResearch Loop
- Generate variant
- Deploy
- Harvest data
- Log result
- Next experiment
A closed iteration loop where each run's result is logged and informs the next variant. The log compounds as the primary AI context.
Steal forAny repeating process with a measurable output — email sequences, ad creative, onboarding flows
03:46list
Three Requirements for AutoResearch
- One thing to change
- One objective metric
- A way to read the result automatically
Minimum viable criteria before starting an AutoResearch project. All three must be satisfied; if any is missing, the loop stalls.
Steal forProject scoping / deciding what NOT to automate
26:26model
Business Guide to Choosing
- Does it have a measurable metric?
- Can you get data via API or scraping?
- Does it run frequently enough for fast feedback?
Decision flowchart for selecting the first (or next) AutoResearch project. High volume + fast feedback + numeric metric = ideal candidate.
Steal forPrioritizing which business metric to automate first
How they asked for the click.
30:40next-video
“Knowing how to leverage the auto research skill without knowing how to build this database of knowledge from your chosen experts is going to severely limit your skills and capability — which is why the next thing we need to do is build out a knowledge base.”
Soft CTA woven into the final summary, not a hard ask. Points to a prior video on vectorizing expert content. No subscribe pitch, no sponsor.
Visual structure at a glance.
