29:07How I AI · Tutorial
Loop engineering for beginners
A plain-English field guide to every loop type — heartbeat, cron, hook, and goal — with two live builds in Claude Code and Codex.
June 17thThe argument in one line.
Marketing compounds the same way Anthropic's AI compounds: when your analysis loop automatically updates the knowledge base that feeds your content ideation, output quality improves every cycle without proportionally more human effort.
Read if. Skip if.
READ IF YOU ARE…
- You run a B2B content operation — LinkedIn, ads, newsletters, cold outreach — and currently analyze performance manually, if at all.
- You want to use Claude Code or similar AI automation to close the feedback loop between analytics and ideation so improvements happen without you supervising every step.
- You manage content for clients and need a repeatable system that gets smarter over time rather than resetting with every new brief.
- You are starting from zero and need a structured way to bootstrap a content strategy from competitor research before you have your own data.
SKIP IF…
- You want finished prompt templates or copy — the video is about system architecture, not tactical outputs.
- You are in B2C or consumer e-commerce with fundamentally different content cadences and audiences.
The full version, fast.
Anthropic's internal data shows their engineers now ship 8x as much code per quarter as pre-AI baselines — because Claude builds the next version of Claude. The presenter argues marketing can do the same: set up a three-node loop where a bi-weekly analysis pass (your performance plus competitor scraping plus industry trends plus your own voice notes) automatically updates a persistent knowledge base, which then feeds AI ideation for the next content batch. The key constraint is trend freshness — LLMs go stale, so you must explicitly inject current signals. The key opinion is that AI should handle concepts, not finished copy; the right angle is 80% of the result.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:27
01 · Anthropic's paper and the 8x multiplier
Introduces the When AI Builds Itself paper and the code-contribution chart showing an 8x jump. Frames the question: can marketing compound the same way?
01:27 – 03:15
02 · Context: B2B content agency and client proof
Establishes credibility — runs a LinkedIn content agency, shows client Mike ($100M B2B agency) and Christian (FDA cybersecurity). Mentions early automation experiments and their limits.
03:15 – 04:45
03 · The three-node loop defined
Presents the core framework on a Miro board: Bi-Weekly Improvement Loop to System's Brain to New Batch of Creatives, as a closed cycle.
04:45 – 07:00
04 · The Bi-Weekly Improvement Loop (4 steps)
Drills into the four inputs: own performance analysis, competitor study, industry trend consumption, and operator input. Flags trend freshness as AI's natural weak point.
07:00 – 09:27
05 · New Batch of Creatives (3 inputs)
Covers the three creative inputs: AI ideation (concepts only), authentic personal input (weekly voice interview), and SOPs. Argues strongly against using AI for final copy.
09:27 – 10:27
06 · Cold start and extra ideas
Cold-start protocol: 50 competitor concepts as first test batch. Credits Alex Hormozi's $100M Hooks. Transitions to the Ali Abdaal thumbnail insight on trend decay.
10:27 – 12:00
07 · Trend decay and the thumbnail lesson
Uses Ali Abdaal's YouTube Academy to show that shocked-face thumbnails stopped working after audiences saw 10,000 of them. Trend awareness separates mediocre AI content from great.
12:00 – 14:13
08 · How to find competitors by channel
Practical guide: Apify for organic social, Facebook Ads Library sorted by impressions for paid, Sales Navigator plus Apify for LinkedIn, subscribing to email drip campaigns for newsletters.
Lines worth screenshotting.
- Anthropic's engineers now ship 8x as much code per quarter as they did before AI assistance — the same compounding logic applies to any repeatable knowledge work.
- A marketing system that automatically updates its own knowledge base after each performance cycle improves without proportionally more human input.
- LLMs go stale the moment training ends — if you do not explicitly feed current industry trends into your system, your AI produces angles audiences already tuned out.
- The right concept is 80% of marketing success; manual effort applied to the wrong angle generates zero return no matter how polished the execution.
- AI should write concepts and angles, not finished copy, emails, or ad visuals — the quality gap at the final-output layer is still real in 2026.
- A competitor's email drip campaign is a free swipe file — big companies often spend six figures testing subject lines and copy before you can subscribe and read it for nothing.
- Facebook Ads Library sorted by impressions high-to-low shows you the winning concepts your competitors already paid to discover.
- A 45-minute weekly voice interview with a subject-matter expert produces more authentic content signal than any AI prompt can synthesize from scratch.
- If you have no content history, start with 50 competitor concepts that are already proven to perform — then use your own data to diverge from there.
- The part of the improvement loop that requires human input — industry news, community sentiment, operator perspective — is also the part most people skip entirely.
The feedback loop that makes marketing compound.
WHAT TO LEARN
The gap between AI-assisted content that plateaus and AI-assisted content that keeps improving is one architectural decision: whether the system updates its own knowledge base after each performance cycle.
- An AI content system that does not ingest current industry trends produces angles audiences have already tuned out — the freshness problem is structural, not a prompting fix.
- The right concept is approximately 80% of marketing success; using AI to select proven angles and reserving human effort for execution is a better division of labor than using AI to write finished copy.
- Competitor research is free signal: Facebook Ads Library sorted by impressions shows you what concepts already survived real budget testing, and subscribing to a competitor's email drip gives you a swipe file they paid six figures to produce.
- A bi-weekly analysis cadence — covering your own performance, competitor moves, and industry sentiment — gives the system enough data to learn from without drowning in noise.
- When you have no performance history, seeding your first content batch with 50 top-performing competitor concepts is more reliable than starting from a blank brief.
- Human voice input — a 45-minute weekly interview capturing real opinions and stories — is what separates AI-assisted content from generic AI output; the system produces authentic material only when given authentic raw material.
Terms worth knowing.
- Recursive self-improvement
- A system where the output of one cycle improves the inputs of the next, compounding gains over time without proportionally more external effort. Originally used to describe AI systems that help build better versions of themselves.
- System's Brain
- The persistent knowledge base in the three-node loop: system prompts, brand context, and references to what has and has not worked. Acts as the long-term memory the improvement loop writes to and the creative batch reads from.
- Apify
- A third-party web scraping platform that can pull structured data from social media platforms and is commonly connected to Claude Code for automated competitor analysis.
- Bi-weekly improvement loop
- A recurring two-week analysis cycle covering four inputs: own performance review, competitor study, industry trend consumption, and the operator's own observations. Chosen as the minimum interval to accumulate enough data to learn from.
- Cold start
- The situation where a new content system has no historical performance data to learn from. Solved by scraping 50 high-performing competitor concepts and using those as the first test batch.
Things they pointed at.
10:02book$100M Hooks — Alex Hormozi
10:20productAli Abdaal YouTube Academy
Lines you could clip.
08:08
“The right concept is already, like, 80% of success. If you're just hitting the right point, hitting the right angle, and it is already proven to work in the past, man, like, you're saving so much time.”
05:58
“LLM is trained once, and then you need to directly, explicitly feed it with the latest up-to-date information, so it is aware about the trends. And usually, you're not doing it.”
06:02
“If a person have seen the same idea from the same angle 1,000 times, it is not compelling anymore. It is not novel anymore.”
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.
metaphoranalogystory
00:00So Anthropic just released this paper which is called when AI builds itself, and it is all about self improving AI systems that are helping them kinda speed up their research process even further.
00:12So if in 2024 this was the amount of code that they were pushing and committing using AI, now it is eight x.
00:22So AI is now helping like eight more times to develop itself within anthropic research and development efforts.
00:31And I just thought, okay. How can we apply that to marketing? I am sure that we can do that, make the whole process self improving, so you get, let's say, from three x to five x more output on the same amount of marketing efforts.
00:46Because marketing is all about testing new hypothesis, seeing how it works, and then improving over time. So this video will be about content creation, ads, even email marketing, and cold outreach. It is all applicable.
01:01Right? And just for the context, I'm running a content agency. For those who are new here, I'm helping b to b business owners to grow on LinkedIn with organic content.
01:10And here's one of my clients, Mike. He's running a $100,000,000 b to b agency in The US, and we're handling his LinkedIn content.
01:20So this is why you might want to listen to me. I have some level of experience when it comes to that.
01:27So I prepared this Myro board with some of the ideas and brainstorm around this concept. So my goal is to help you also integrate this kinda recursive self improvement AI in your business.
01:42One of the other reasons why I decided to make this video is that Y Combinator, which is one of the biggest startup accelerators in the world, they started to drop videos about the same topics. How to build a self improving company with AI. And here, the guy is explaining how, like, you can create this closed loop within your company, so you don't need to be supervising what is going on.
02:07That if something goes wrong, it will auto improve itself, and then the next day the problem is fully gone. Right? So I really feel like this is the future.
02:18And we already had some of the these parts of the process integrated inside our agency. So I'll just show you an example.
02:26Here is Christian. He's one of my clients. He's running also a b to b company in The US, helping medical devices manufacturers get FDA approved, and we are handling his content.
02:39So we're publishing multiple posts per week, and we always have to analyze, okay, this post performed that well, this amount of engagement.
02:49Okay. Maybe these are the topics that are resonating a bit more. Let's try to make a bit more content in this direction.
02:55And we had an automated system to build kinda analytics reports, analyze the performance, check out what are the different insights.
03:03For instance, right here, specific technical hacks drive maximum engagement and whatnot. And then we can kinda use this insight to further develop our content strategy.
03:13But it was not the full closed loop. So this video is my attempt to create a new concept, which will be used in the future for us and maybe for you as well.
03:25So here is the loop that I defined. It has three parts. Improvement, then brain, and then new iteration of creatives.
03:35So no matter what you do, if it's, like, organic content or ads or email or even called outreach, you usually have these three things within your process. You have some sort of foundation, like, for instance, your prompts or your SOPs, how you do things, like, what are the regular way for you, let's say, to create content.
03:56Then you have your own context about what is your brand, what are the main topics that you wanna be discussing, and then you have some references about the stuff that is working well, the stuff that is not working well. Let's say, this is kind of the foundation for the system. This is the main knowledge base.
04:10Okay? Then based on this knowledge, you're creating a new batch of creatives, and then you're trying to improve the process.
04:18And so this loop is this kinda three things, but more so from the angle of AI. So let's look at it a bit deeper. Right?
04:28So when I'm looking at the, let's say, biweekly improvement loop, I decided to write biweekly because this is a reasonable amount of time when you're able to collect some data that you can analyze and learn from and then improve the process.
04:43Right? So I see four steps here in the improvement loop. We want to analyze our own performance of our creatives that we've just published.
04:53Second, we wanna study our competitors because we might have 100 competitors, and they are have, like, their own ad budgets or content creation budgets, and they're all constantly testing new hypothesis.
05:07We can have a huge leverage if we are learning from them as well. And then third thing, which is very important, in my opinion, is consuming industry news, comments, understanding where community is headed.
05:21This is kinda the whole trend context about what is going on in the inside the industry. And I feel like this is the part where AI is naturally really weak, because LLM is trained once, and then, like, you need to directly, explicitly feed it with the latest up to date information, so it is aware about the trends.
05:43And usually, you're not doing it, and what you get is very outdated perspective of the industry.
05:50And for that reason, in terms of content and marketing, it will work worse because the angles, they kinda get outdated. If a person have seen the same idea from the same angle 1,000 times, it is not compelling anymore.
06:05It is not novel anymore. So this is a big problem. And the final fourth stop is your own operator's input.
06:14Of course, you have your own perspective. Maybe you have your meetings with clients. You're collecting some intel.
06:20We wanna put it all into the improvement loop. So, ideally, this stuff is all happening almost fully automatically. Here, of course, you need to provide the input, but it can be in a very efficient way, in a very efficient manner.
06:33Maybe your AI is connected to your Slack or you're just feeding some voice notes every single two weeks in the system. So there are different ways to do that.
06:43While the first two steps can be done completely automatically, just have AI system, maybe with Cloud Code set up to fully, like, regularly atomize auto analyze your ad performance or content performance.
06:55And then this thing should automatically update your system's brain, your knowledge base, your foundation.
07:03Hey, we've collected this information about these experiments. These angles are resonating better. In this case, in the example of this analytics report for the LinkedIn content of one of my clients, we are getting some of the insights, and this information should be stored to our knowledge base for this specific client.
07:22And then we will base our new hypothesis specifically based on these insights that we've just collected.
07:32And then it all goes into the next step, which is the new batch of creatives. So AI will handle the ideation. And by the way, I really believe that right now, with the current state of marketing, the best way to use AI is not to generate the final content itself.
07:50Not to write the final email sequence, not to create the final ad visuals, because the quality will be really not good enough. But when it comes to ideas and concepts, it is insanely helpful, and in my opinion, .
08:08If you're just hitting the right point, hitting the right angle, and it is already proven to work in the past, man, like, you're saving so much time. It is very, very easy to put manual effort in creating a well prepared visual, which will generate you zero return on investment just because the concept was selected in the wrong manner.
08:30So this is the first part, and then the second part is your authentic input. Once again, in my case, from for my clients, like, they are doing weekly voice interview.
08:41So this guy Christian, every single week is talking to the voice interviewer for forty five minutes, just, like, sharing his own ideas, sharing his own stories, and then based on this authentic thing, we're creating content.
08:53So I don't believe in the generic kinda AI slope machine approach at all.
09:00And finally, of course, you have some sort of SOPs. So the ways you create content, based on that, we're generating this content. And this is also coming from the system's brain.
09:09So this is kind of the loop that I will be implementing within my content agency throughout the next few weeks. I would love to share my results.
09:19Also, I'm very curious to hear what you think about that, how you're implementing it within your business. So please share it in the comment down below. And also, a few more ideas right here.
09:31First of all, about the cold start. If you don't have, like, the original batch of content, if you're just starting out, you don't have, like, the initial context, maybe you don't know where to start. It's very simple.
09:43You just start by studying the competition, selecting 50 concepts that you like from them that are really well performing, and then this is the first thing that you test. This is where you start, and then you go from there.
09:55And this advice is taken from Alex Ramos' book called $100,000,000 hooks. This is a playbook from his third latest book launch.
10:03Right? And here are a few more kind of side ideas. Like, the whole kind of vision for this video came after I was watching a course about how to grow on YouTube from Ali Abdaal.
10:16Maybe you know this guy. He's like a productivity expert. And, basically, one of the lessons, it was going into the topic of thumbnails.
10:25And here is a thumbnail. But you can tell that this thumbnail is from the past. This is something that stopped working on YouTube a few years ago.
10:35When a person has this kind of shocked face and with, like, crazy emotion, it's not longer working.
10:43And the reason is every single concept in marketing, for a while, it might be popular, trending, working, novel, and this is resonating with people.
10:54But after people have consumed, like, 10,000 thumbnails like that, it's not clicking anymore. They already kinda have this really blurry vision towards this thing.
11:06They will not even recognize that. For that reason, now the thumbnails that are working, they are way more calm, way more kinda less emotional, and this is what is crushing in 2026.
11:18And this got me thinking, oh, this kinda trending component, component of understanding of your niche deeply, like right now what is going on, is absolutely crucial.
11:29And I think this is the component that really separates a mediocre AI system for content with something that is amazing and great. And another side thing is about how to find your competitors.
11:43Maybe you don't know yet who are those companies, those people that you're competing against, and how can you possibly learn from them. So depending on what you do, for instance, if you're doing the organic content, I highly suggest you use Apify.
11:57So this is a third party software. You can easily connect it to, let's say, Cloud Code or just Cloud, and then it will scrape any sort of social media data.
12:06TikTok, Instagram, YouTube, whatever you want. And then you'll be able to pull up the analytics about the engagement, what are the outliers, what are the underperformance, and then ask Cloud Code, for instance, analyze all that.
12:19By the way, I have videos on these subjects on my channel, so many of them. So check them out if this is something that you're into. Then if you're in ads, of course, you can learn from Facebook ads library.
12:31Just filter the ads using the impressions counter from the highest to the lowest. You will see the winning concepts from your competitors.
12:39When it comes to outbound oh, sorry. Not outbound, but personal branding, let's say, LinkedIn, you can use Sales Navigator.
12:47You can filter people within your industry, and then once again scrape this inform their information using APFi. Check out who are those people with, let's say, 15 k plus followers, and then you will analyze them.
12:59They are probably industry leaders. They are probably authorities in your space.
13:04And finally, when it comes to, let's say, emails or newsletters, email sequences, You can find your competitors in some of the previous ways, and then you can just open their website. If it's an e com store, they're usually, like, suggesting you, hey.
13:19Leave your email and get a 10% discount. And then once you leave their your email, you will get an email sequence, drip campaign usually. You will be able to learn what are the approaches that they are using.
13:31Or you can subscribe to their newsletter, and you will also consume the stuff that they are testing right now, that and is probably working for them, especially if it's a big company. Trust me, they probably spent like a 100 k plus testing their copy, their subject lines, their bodies.
13:48So this is a definitely useful piece of data to study. So that was it.
13:54Hopefully, now this concept is a bit more developed in your mind. Please let me know what you think, how you will apply that.
14:02And if you're interested, here's another video where I'm sharing how to use Cloud Code for different sorts of marketing tasks from content creation to outbound and even ads.
The Hook
The bait, then the rug-pull.
Anthropic published a chart that stops you cold: by the time their latest model shipped, engineers were committing eight times as much code per quarter as they were the year before — not because they hired eight times as many engineers, but because AI was writing most of it. One agency founder watched that chart and asked a different question: what would happen if you ran marketing the same way?
Named ideas worth stealing.
03:15model
Three-Node Self-Improving Marketing Loop
- Bi-Weekly Improvement Loop
- System's Brain
- New Batch of Creatives
A closed feedback cycle where performance analysis automatically updates a knowledge base that feeds the next content batch.
Steal forAny repeatable content or ad operation — LinkedIn, email, TikTok, ads
04:45list
Bi-Weekly Improvement Loop (4 steps)
- Analyze own performance
- Study competitors
- Consume industry news and trends
- Operator's input
Four inputs that feed the system's brain every two weeks. Steps 1 and 2 can be fully automated; step 3 requires live data injection; step 4 requires human input.
Steal forContent strategy review cadence, performance marketing retrospectives
09:40concept
Cold Start Protocol
When you have no performance history, scrape 50 top-performing competitor concepts and use those as your first test batch. Derived from Alex Hormozi's $100M Hooks.
Steal forNew clients, new channels, new niches
How they asked for the click.
VERBAL ASK
14:00next-video
“If you're interested, here's another video where I'm sharing how to use Claude Code for different sorts of marketing tasks.”
Soft end-card CTA with no subscribe push. Organic and low-pressure.
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.
