He was building an engineering team to build Hermes agents for his clients.

At first, he asked questions to me by looking at his screen.
I understand he generated these questions with AI; the tone and the sentences are obvious.
But as I spoke about the agents I built, I could see the spark in his eyes.
He started asking questions by directly looking at me, and the AI tone in his questions suddenly vanished.
I understand now that he is interested.
But at that point, he lost me.

Because I was already thinking about writing an article about the Agents I built over time.
Some of you guys found this disrespectful, but I've worked thousands of hours, and if the one who'll hire you doesn't know about the job, things always go bad.
I answered his remaining questions and did not even ask further questions after his AI questions were over.
The meeting was finished; I opened my Substack and started writing this one.

I am going to show you 5 AI agents that I created through Hermes, but first, let me explain to you what a Hermes agent means.

What is a Hermes Agent?

Hermes agents are very similar to OpenClaw agents.

They run continuously on your computer unless your PC is turned off or your internet connection goes down.

You can talk to them through Telegram, and they can access the applications you allow.

In the background, they use either LLM APIs or OAuth.

OAuth means using your existing LLM subscription, such as ChatGPT or Claude.

With OpenClaw/Hermes(third-party tools), however, you cannot use Claude because Anthropic does not allow it.

You can install them by hiring a VPS (Virtual Private Server) or on a Mac mini.

You can read this one I showed you how to install one.

Now, let me introduce you to my favorite Hermes agents.

1. One Idea Can Change Your Life

This one is special to me because it made me believe in something simple: every great thing starts with just one idea.

I’ve seen this happen in my own life, and I keep seeing it on Substack too.

For example, look at the numbers behind one simple Substack Note I published.

One of my Substack posts brought in 25% of all my subscribers. Can you believe that?

That made me think: if I want to keep finding ideas that can inspire people, I need a Claude Skill built for exactly that.

It needs to search the web, collect the right signals, and surface ideas worth turning into content.

And the final report should look like a dashboard.

Because plain text bored me.

So I built it as a Claude Skill and explained everything.

See How I Built It

Note: I love it so much, so I turned it into hermes agent. (Upload the skill to Hermes and tell it to create an agent using this skill.)

Subscribe now

2. Trading Bot Copies Millionaires

I love math; in middle school, they called me Gencalculus because I am good at it.

So my wife always told me that I should use this skill in Finance.

But this is entire different discipline.

That’s why I sometimes stop calculating and just copy the millionaires.

It’s actually interesting that this is legal. I was surprised at first too. :)

Unfortunately, in the country where I live, we can’t open short or leveraged positions.

So I couldn’t use it with real money, but I tested everything through paper trading and shared the full process.

See How I Built It

3. The Security Hermes Agent Watches Me Over Camera

I’ve always been fascinated by security systems.

Maybe I simply like feeling safe, or maybe I like knowing what’s happening around me, but either way, I have two security cameras watching my house.

After building systems with the Hermes agent, I started wondering: could I make AI watch my garden 24/7 and report anything to me?

And then I wondered: could it actually call me when a stranger enters my garden?

I built both features.

Subscribe now

Now the AI watches my garden and alerts me when something happens, and somehow, that makes me feel a little safer.

See How I Built

4. Second Brain Who Can Think and Never Forgets

The idea of having a second brain felt incredibly exciting when I first discovered it.

So I built one. It worked, but it was far from perfect, and over time I kept reshaping it around every new tool I discovered: Obsidian, Telegram, and the Hermes agent.

Subscribe now

Eventually, I built a second brain that knows an almost uncomfortable amount about me, even patterns in how I think.

How? I let it analyze my previous conversations with Claude, gave it a diary, connected it to an agent, and structured its memory using Karpathy’s wiki-style method.

See How I Built It

5. Chief of Staff Hermes Agent on a Local Model

By now, you might be thinking, “You have too many agents. How do you even keep track of them?”

And you’re right.

Subscribe now

I actually have at least five more that I haven’t shared yet, and sometimes one of those conversations stops halfway through without me noticing.

That’s why I built my own Chief of Staff.

It can even run on local models, quietly checking the conversations between my agents and me and reporting back whenever something gets left unfinished.

See How I Built

What is next?

You see, I am using AI agents daily, so I want you to read about them from a person who spends thousands of hours with them, builds them for companies, and even interviews people regularly to learn more.

Subscribe now

Also, to share my experience with you guys, I built AI Academy. And the community is starting to shape here.

Here, we’re building in public.

I built a curriculum for NotebookLM, prompting, and Claude using my Substack articles.

When you read the article and finish the build, you need to share it inside the AI Academy to get my approval.

I’ll read your build, add my comments, and you’re done.

In a couple of weeks, Claude Cowork and Claude Code courses will be uploaded here. I am writing these courses for big Substack accounts and getting their approval to upload here.

To reach here, you need to be in the inner circle; read the details from here if you’re interested in it.

Thanks for reading, and final note:

I needed more freedom.

Because I don’t like to be caged.

So I quit my job.

BTW, the company I was working for was the third biggest company in my country, Turkish Airlines.

I was working as an Engineer.

This shocked everyone near me.

Everyone was expecting me to land another job.

But my plan was different.

That plan eventually became the AI freelancing system I am about to show you.

Step 1: Data Science and AI Education

Believe it or not, I even searched for universities offering psychology programs.

I emailed some students at Sigmund Freud Private University in Vienna.

At that time, my best friend was preparing to become a data scientist.

When he told me about it, I loved the idea and started taking courses on Coursera.

I collected every certificate I could.

But there was one problem.

I couldn’t work for corporate companies because of visa restrictions, as I was living in Turkey.

I either had to move to the US or find another way.

I found another way.

Step 2: AI Freelancing on Upwork

I created accounts on Fiverr, Upwork, and several other freelancing platforms.

I loved Upwork and dreamed of working with multiple companies overseas.

So I did.

I have worked more than 7,400 hours.

I started as a data science writer and evolved into an AI expert.

I achieved my dream of working with companies around the world.

I worked with some of the largest companies in their industries, as well as smaller companies trying to stand out.

I watched job titles and expectations change constantly.

So I had to adapt to those changes too.

I had to build projects that clients would love.

And my own title changed many times along the way.

During that time, I realized that I always follow the same pattern.

The Pattern of Success

During this period, I worked on countless projects.

After a while, I decided to start writing about them.

There was demand from you, and that is how LearnAIWithMe became what it is today.

But how do I decide which projects to build?

How do I learn what I need?

And how do I turn that new knowledge into real experience?

I am going to show you.

But first, I want to introduce the improved version of my RADAR system.

Following AI Freelancing Trends

To become a successful freelancer, you have to follow the trends.

So I did.

But then I realized there is a better way now.

Guess what I used?

Claude Skills, of course.

Claude Skills are among the most powerful tools available right now, yet they are also incredibly easy to install.

So I built my own radar using the Upwork API through Apify.

Let me explain to you how it works.

The Upwork Trend Radar (Built with Claude Skills)

This radar runs on a schedule.

It checks new trends and sends me a report like this.

First, the demand: how many jobs are listed, what the average hourly rate and fixed-price budget are, and how many proposals each listing receives on average.

Next, the skills in demand. Even when I filter for AI jobs, many listings also ask for Python, machine learning, APIs, and similar skills.

The next two sections are experience level and what moved, showing which skills companies are chasing

Next is the most important part: the action points. Based on these, I tend to build similar projects.

And if you think you’ve already done enough trending projects, check the rising list. These may not be in the top five yet, but they could be early signals.

I want you to earn money with AI.

Not just write prompts.

Or even build projects.

I want you to truly benefit from it.

Now, let me give you the skill file.

You can build your own radar and start chasing the job you want.

Or even better, make them chase you.

Read more

I wanted to know what the 3% have that the rest do not.

My first thought was better tools.

So I went hunting and found one.

The Bloomberg Terminal looked good.

Then I saw the price.

I said, "No way!“

So I find free version of it.

And I built a Claude skill that makes me a smarter trader, let me show you what I’ve found as a replacement and what I built.

The Free Bloomberg Terminal Alternative

An hour of digging later, I had the free version.

A free, open-source terminal.
Live indices, portfolio, news, options, 100+ data feeds, all for zero.

The screen was solved. One problem left.

The Claude Skill That Tests Every Trade Idea

This screen cannot tell me if my idea works.

So, I found the second tool.

A tool that tests any strategy.

You also don’t need to be technical to do it.

This Claude Skill has two modules, let me show you.

Module 1: Backtesting Any Strategy

For instance, here is my question:

Bitcoin dropped 10%. Should I buy the dip?

So I tested it.

Every time BTC fell 10% from its high, put in $100 and hold it.

Then I ran the same $100 into the S&P 500 on the same days.

Now I could test anything. But I wanted more.

Module 2: The Morning Report Routine

I did not want to ask the questions all the time.

I wanted an employee who asks them for me, reads the news while I sleep, and only brings me the ideas that survive a backtest.

And each morning, it reports me like this.

Look closer at what the tester is checking

What if you’d bought it 6 months ago

That is the whole system.

So now I have a screen to watch, an engine to test, and an employee to do it for me.

Now, let me give you the skill that makes me a smarter trader at first.

And I’ll show you how to install this Bloomberg Terminal free alternative.

Also, I am going to show you a trick to use it for free.

Here they are.

Read more

I read it in the comments.

And I got a lot of DMs.

I answer your questions all the time.

Over time, I realized that the more AI use cases I publish, the more you like them and the more you ask for.

This is the pattern.

And it needs more attention.

The Pilots Already Ran

I tested this class silently this summer, without naming it.

A Claude skill audits like a $300/hour consultant, a McKinsey-style report, a Deloitte-style audit.

You loved it, so you voted before the series had a name.

Now it has a name and a schedule.

Every Wednesday, one real use case from a real business is rebuilt as a Claude skill.

You can install it in 5 minutes and run it forever.

AI in Business

We had different series before.

Build-it teaches you the reflex of building with AI. We built 30 different builds, using AI.

Build-ship-repeat, or BSR, as you called it, teaches you how to turn raw ideas into real products and ship them.

Now we go one step further.

AI in Business points the same muscle at the money you already make.

You do not need to own a business. AI in Business targets the most expensive work inside the business you work for.

So most of you will earn more from AI than from any app you might ship.

You become the one who replaces the expensive work with AI.

Want Your Business in a Case?

You know my freelancing experience. (7,000+ hours.)

These days I take on jobs with different businesses to collect real-world use cases.

You see, one of the benefits of freelancing is that you can take on different jobs much faster.

For instance, I have an interview today.

It is about Hermes agents. For a while, I’ve been building Hermes agents for LearnAIWithMe, and I’ve been wondering how they are used in different businesses.

But this is not enough.

DM me.

A couple of lines are often enough.

Explain your use case and tell me why it matters and how it affects other business professionals.

I want to hear your solution.

Maybe I’ll build on top of it and share it on LearnAIWithMe if I’m convinced it can solve problems for others.

Case #1: Coming Wednesday, August 19

The first use case will be next Wednesday. I have a shortlist.

The solution will be a Claude skill.

And this skill can be installed with one prompt.

So every use case will give you a reusable skill.

12 weeks later, you will have 12 working skills in your business.

You don’t need to learn how to prompt right now.

You learned how to build.

Now you’ll change your business or become irreplaceable in your work.

Do you want more?

If you want me to review your use cases and audit your work, join the Academy from the Vault.

Note: Claude Code (updated), Claude Cowork, and ChatGPT Work series will be added to the AI Academy soon.

I created them in collaboration with two of the biggest Substack publications, and I’ve received permission to add them to the Academy.

To become a billionaire, you take the risk others won't and hold a relentless belief that you can do it.

This is a crazy move, and it inspired the Claude Skill I'll show you below.

I also am a bit crazy.

I quit 2 jobs that my friends from the university would die for.

These companies were among the top 5 in my country at the time, but neither my friends nor my family could understand me. ( I don’t think they still can.)

But the real thing is, I don’t want to be in the cage.

And I thought freelancing would be the best fit for me.

So, I’ve worked thousands of hours over the years to achieve this.

I’ve worked at every level of business, from small companies to billionaires.

But I’ve realized that no client can give me the feeling I seek.

And I was chasing this feeling and started writing on Subsatck at 2023.

At first couple of years, I wrote about Data Science, and honestly, I did not have that much time.

But over time, I realized that this is very similar to what I chase over the years, a community that can share ideas, give feedback, and build together, as we did on the AI Academy, inside the vault.

Like in the comments, criticising the Claude due to their limits.

So over time, LearnAIWithMe became something I am eager to jump out of bed every day at 5:30.

And next, I want to be the best at what I do, and sometimes I sit in this chair and think, "I love this."

But the crazy thing is, I don’t just want to be the best in one category; I want to be the best in the country.

And I want to be the best in the world. (Don’t have the screenshot to show that, yet :))

But how could I do this?
The pace at which I want to grow is unprecedented.

When I created a project called 90 days on April 7, LearnAIWithMe had 5.3K subscribers.

The projection Claude predicted that LearnAIWithMe will be 10K at the end of the year, in the best prediction, when I’m planning my growth.

Look at it two months later.

That’s why I knew the ceiling couldn’t be it.

So I searched the web for an answer on how to grow even further, to find limits.

P.S. We had reached 17K, but I removed 2.6K inactive readers two days ago to increase engagement in the community.

What does this Claude Skill do?

People who are bending those limits are billionaires.

People like Lucy Guo. They are relentless in chasing an idea by believing this can be true.

So I built a skill that researches billionaires’ public speeches.

And cluster them.

After you initiate this skill, it’ll interview you and put you in a cluster of billionaires.

The billionaires, thinking like you, solving problems like you.

And create your report, here is mine.

Let me show you.

My Report: Which Billionaire Thinks Like Me

Here is my cluster.

No wonder I like Lucy Guo’s mindset, we are wired the same way.

Next, it analyzes my past.

Claude predicted 10K by year-end, and I passed 16K in two months.

And what should I do to hit 100K free & 5.000 paid?

And the part explains what I might keep avoiding.

And finally her words.

Every line is sourced from a real interview.

How the Claude Skill Works: Two Phases

The skill asks five questions before it starts.

• Risk appetite

• What drives you

• How you build

• What you do in a crisis

• Your goal, as a number

My answers were;

• I leap when I see what others miss

• Freedom, I don’t want a cage

• Build from zero, ship rough

• I go harder

• 100K free subscribers, 5,000 paid

I built this skill in two phases.

Phase 1

When I built the skill, research agents went out, one for each billionaire on the Forbes top 100 list.

Each agent searched the web for their real words on how they got rich and saved them in a file.

103 people, 1,103 quotes.

Everyone traces back to an interview, a letter, or a talk.

The skill studied all of them once and collapsed them into seven clusters.

How they take risk, how they build, how they act in a crisis.

Each cluster has its own agent file.

Phase 2

It interviews you, matches your answers against the seven clusters, and drops you into the one you belong to.

That cluster’s agent takes your goal and answers it in the cluster’s own voice.

At the end, it turns the whole report into a visual dashboard with GPT Image 2 and Claude Code artifact.

How to Install This Claude Skill

You can install using one prompt and the files that I uploaded to the Google Drive.

Let me give you the Google Drive link:

Read more

First, the prompt you pasted merges with a system prompt, written by Anthropic, your CLAUDE.md, and any Claude Skill that matches the task you described in the prompt.

Next, the model reads it, calls the tools if neccesary and returns with the answer.

What you see as one reply is often ten round-trip.

So the real question is this:

What can you do to get the most out of it?

In this one, we’ll use 6 different features of Claude Code to maximize your harness.

And in the end, I’ll build a Claude skill named “Harness Engineer”.

Harness Engineer Skill

This Claude skill interviews you first.

Based on your answers, it recommends the right mix of CLAUDE.md, Skills, MCPs, Hooks, subagents, and workflows.

If you approve, it’ll build them for you, step by step.

Most of the time, it would not install all six, because it is programmed to install what you actually need.

You are in control all the time; if you don’t want any of them, it’ll skip.

How to install this Claude skill?

Let me show you it is easier for both of us.

Under 6 minutes, I built a project, have a strategy, and the Claude skills are ready to install.

Next, it might offer hook, subagents while creating content or dynamic workflows, if it see fits.

Now, I’ll give you the prompt and the skill, but not just that.

After that, I’ll start from Claude.md, explain to you how to create one, install skill, create skill, mcp, cli, hooks, subagents, and dynamic workflows.

Let’s start.

Here are the files.

Read more

Google first rolled out this to ultra users. Then, on Aug 4, it reached 100% of Pro users.

So, I wrote 10 prompts to test all of it on my own notebook.

But first, let me show you my Notebook, which I’ll use through entire post.

This notebook contains 52 weeks of my income from Upwork, which I gave to my accountant for 2025.

Also, I uploaded all prompts here.

So you can easily save them.

Subscribe to get Claude Skills like this

1- Invoice Test

The first prompt shows how the new cloud computer works. I uploaded 52 PDFs and requested that it create a spreadsheet by analyzing them.

After pasting the first prompt, check what NotebookLM said; it said “Executed code”.

When you click on the “Executed code”, you see step by step what it does.

Here, your personal cloud computer is working because this prompt triggers coding.

After 10 minutes, the task is finished.

Also, I see the file was generated in the Studio.

But this time, we can see this inside the studio, so when I clicked on it, it downloaded, and after that, I opened it.

Here it is.

In 10 minutes, Antigravity inside NotebookLM analyzed 52 PDF invoices and generated a full Excel report.

I wonder how much work this could automate worldwide, especially in accounting and banking.

Subscribe to see more useful builds.

2- The Receipt Test

Here is the second prompt and what it shows.

It shows a new feature, called agentic research.

It was the third item in the Gemini Notebook announcement. This feature basically do a market research and collects your data.

After pasting the prompt, look at the notifications.

It has already started doing market research.

It read my receipts and searched the web for current AI freelance rates.

And at the end, it suggest me to import these findings into the sources.

NotebookLM has changed a lot, so let’s check the next one.

Subscribe to follow the trends

3 - URL Vacuum

After Step 2, I realized I should raise my rate because market prices were higher than I expected.

So I researched the market, read a few articles, and saved everything in a Google Doc. The new URL Vacuum feature can pull sources directly from that document.

After pasting this prompt, here is what NotebookLM did.

And when I check the sources, I see that these were already uploaded.

Share

4- Income Math

Now let’s plot a graph by analyzing these PDF’s.

It did the calculation pretty fast.

If you’ve used NotebookLM for data analysis before, you probably remember even simple analyses sometimes taking 20 to 25 minutes. Now, it finishes in minutes.

But of course, we want an actual plot, not an infographic.

And here it is.

It looks like my December should have been better, but these are the period where I decided to quit freelancing and fully focus on here, so no problem at all :)

Leave a comment

5- Find What I Missed

This is the technique I used all the time when I am analyzing any data using AI.

Because the data analysis should show what’s being hidden.

Here is the prompt I used.

This will also create charts and analyze all the PDF’s. We really make sure our Cloud computer is tired today:)

It found 3 different patterns.

And draw a graph for each of them.

And it did it in a couple of minutes.

Some of the texts of the graphs overlaps and I think they can be adjusted, let’s prompt this.

Fix the overlapping issues inside the graphs.

After a few minutes, here its answer.

Now let’s check one of these plots.

It is better; it needs one more tweak for the numbers, but this version is also enough for now.

Subscribe now

6 - The Single Point of Failure

This is the scariest part for me. Because I might have a clue of what it would find.

Here is the prompt.

Also, do you realize what Google aims? I feel that they are turning NotebookLM into a tool more like ChatGPT or Claude in sense. Because now it can collect data and do data analysis, it definitely becomes more than a research tool.

Here is the full report.

It is good and detailed.

I saved it as a Note by clicking here.

Next, I saved this note as a source by clicking here.

And next, I pasted this prompt;

Turn this into a 16:9, hand-drawn infographic. ( Beyond Platforms: AI Freelancer Income Diversification Strategy)

And here is the image.

I did not click on “Generate infographic” in the studio or anything like this. Just one prompt, and the infographic is ready.

Subscribe now

7 - Three At Once

Every file generated so far took its own prompt. This time I asked for three in one go.

Because with this new upgrade, they said that you can generate multiple at once.

Look at what has happened, just after I pasted this prompt.

A slide deck, an audio overview, and an infographic are being generated at the same time.

Here is the infographic.

You get the idea, that’s why I am not sharing the audio or slide deck.

Let’s continue.

Subscribe now

8 - The Mac Mini Upgrade

I want to upgrade my Mac mini, it is good, but I need a better one because of Kimi K3 and Deepseek, actually :)

But I have to convince myself and my wife to buy it, that’s why I am using this prompt.

The initial report suggest me to by new one.

But let’s ask it to research on the web, for the latest Mac mini models for more complicated LLMs.

And also, I want to see this report as an infographic. I am tired of reading long paragraphs.

Sure, also turn the research report whenever it’ll be ready, into a handwritten, 16x9 infographic.

Here is the image it created.

Here is the full report.

Subscribe now

9- Anomaly Hunt

Let’s check the anomalist.

Here is the prompt.

After pasting the prompt, the xlsx generated inside the studio.

I downloaded by clicking.

There is not any anomally detected, my accountant would love this.

Subscribe now

10 - Chat to Deliverable

It’s time to wrap up.

Using this prompt, we’ll turn everything we just analyzed into an editable slide deck, so we don’t lose any of the insights.

In 1 minute, it generated PowerPoint slides.

After downloading it, I opened it, and here it is.

It includes the latest analysis we did together.

Subscribe now

Thanks for reading!

If you liked what you saw, I build AI systems that actually work every Monday, Wednesday, and Friday, and I share the full process, steps, and files with you.

Share it with your friends so more people can benefit from it.

Share

In just 4 weeks, his portfolio is up 23.8 %.

After digging deeper, I found this Substack.

At the end of week 20, Nathan’s portfolio sat 30% below starting capital.

I love his confidence and this trial, because for a while, I have also been testing AI to trade.

I have built trading bots before. A bot that copies millionaire wallets, a Polymarket bot that returned 2.1x in 18 days.

But these bot acts by following a script and never tell you whether your trades were right or not, you’ll measure by the results.

You know who has receipts?

Warren Buffett. He ran Berkshire for 59 years and turned every $1,000 into $44 million.

Last time, I built a digital Buffett with NotebookLM and Claude that could answer questions but never had to prove anything.

This time, I wanted more than answers. I wanted a Warren Buffett AI agent that gets graded.

He picks, then history judges him, on my screen.

I built it using Claude Code and a bunch of other tools.

Let me show you.

The architecture: a picker, a judge, and a desk

Three layers.

This is the same loop structure I use everywhere now: one agent produces, a second agent audits, and I only see the result after they have argued.

• Layer 1 is the picker. An open-source project ships with a full team of AI investor agents. I kept exactly one: Warren Buffett.

  • It reads ROE, margins, debt, and owner earnings, then returns a signal with a confidence score and its reasoning. Its original data source has no free tier.

    I paid $20 to find out what that buys, reverse engineered everything and packed it for you.

• Layer 2 is the judge. An AI backtesting engine takes the same ticker and runs the last 12 months against SPY.

  • A validator then checks the engine’s math against an independent calculation. If the numbers disagree, the AI’s number gets thrown out.

• Layer 3 is the desk. A local dashboard on my machine.

  • I type a ticker, press one button, and both layers fire.

This dashboard is powered by the skill I’ll send you.

All you need to do is download the skill and the app files, create a new project, and you’re ready to go.

But first, let me show you how the app works.

What will the result look like?

The result will be the dashboard.

Pick your stock, click on “Run Desk”.

Stage 1: Buffett states his opinion.

For Microsoft, he said neutral, 40% confidence: “Strong margins& book value growth, but weak ROE and price far above intrinsic value”.

If you need more, click the scorecard to see why.

Stage 2: The judge runs the tape.

MSFT returned -22.2% over the last 12 months.

SPY returned +19.5%.

Stage 3: The desk turns the backtest into a decision

For MSFT: watch, no edge either way. Revisit if the price drops or fundamentals shift.

What’s next?

After your analysis is finished, it’ll suggest that you can run similar backtests.

Let me show you how to install it in 10 minutes.

10 Minute Setup

I put everything you need to install this system into one folder.

It even has One-prompt.md to install everything in 10 minutes.

Download this folder, or if you have Google Drive MCP installed, or better gws-cli installed, give the link to this folder to your Claude.

And it’ll install everything, because the entire folder has a roadmap, specifically designed for your Claude.

Here is the link to this folder:

AI Academy

One more thing.

Inside the AI Academy, there are more skills.

Also, we are building together.

More information is here.

Read more

As someone who enjoys building things with AI, it immediately caught my attention. But I did not know about Y Combinator.

It turns out Y Combinator (YC) is one of the world’s best-known startup accelerators.

They invest in early-stage startups, typically $500,000 per company.

And they often share lists showing expectations from developers.

So, I picked one of the ideas and built it.

But there is more for you.

I also created a skill for you.

So you can build anything you want.

One sentence enough.

But why should you use this skill instead of telling Claude what to build?

Because this skill first researches the industry, finds gaps, and builds.

So the app will be better than your rivals.

Let me show you the skill first.

The Moat Breaker

This Claude Skill first searches the web and finds your rivals.

Next, it identifies industry best practices.

And trains a notebookLM.

Next, it creates 3 agents, dedicated to building this app.

• Shell Agent: Build the front-end of the page, the part you see.

• Engine Agent: Connects the front-end with the back-end, works like an engine.

• Gap Agent: Find gaps in the market.

And while building an app on Claude Code, these agents talk with NotebookLM.

The reason for this Claude Code and NotebookLM combination is to avoid hallucination.

Because NotebookLM is very good at grounding.

And I’ll show you what I built next.

The Primer

I gamify the entire process.

To join the game and start learning, you will need to answer a few questions.

Once you fill in the information, your profile is ready.

When you click on “For grown-ups”, you see the progress, the exact curriculum standard behind every skill, and what actually tripped your child up.

You get sixty seconds, and every right answer earns a star, in the lightning mode.

Questions come one after another with four big colorful buttons, no typing, just tap and go.

Eva got 28 correct answers in one minute (I did it for her; she is only 2 months old 🙂), setting a new record and earning 28 stars for the bank.

And Eva has badges, which makes the entire process more exciting.

Another mode is racing Pip the fox, where every right answer moves you one step closer to the flag.

After each answer, the tutor explains the why in one kid-sized sentence, like a private teacher sitting next to you.

Win the race and everything lands at once, bonus stars, a new badge, and Pip asking for a rematch.

Now, this is the app I built with it. But you can build any app you want.

I also added this app’s files to the Vault, so you can download them, submit the app to YC, turn it into a product, or let your son or daughter use it. Your choice.

The Installation of The Moat Breaker in 5 Minutes

Here are the everything inside the Google Drive, including the single prompt that will install everything for you.

Download the folder, paste the prompt, and let it install everything on your PC.

Next, describe any app you want in one sentence, build it, and apply to YC. If you get accepted, definitely let me know.

Here is the Google Drive link.

Read more

It outputs 5 paragraphs.

I don’t need the entire story.

I need a short answer.

So I often add this line to my prompts.

Explain in one sentence.

It works every time.

Which shows me one thing.

The model can be short.

Like the sentences in this article.

I feel lost in longer paragraphs.

Maybe that’s because English is my second language.

That’s why I tested 3 Claude skills.

And I built apps using these three.

Let me show you what I built first.

What Did I Build With These Claude Code Skills?

First, I built an app using Osiris.

I use one prompt.

The App is ready in 15 minutes.

Osiris lets you watch the live CCTV.

Check the news, all over the world.

The app name is Fire Detection System.

Next, I built a Conversation Autopsy in 2 minutes.

It analyzes your conversation.

Pinpoint where the issue started.

And the response that would have de-escalated it.

The third app is The Things You Keep Saying.

It’ll analyze all previous conversations.

And create a report, showing you.

It all started with one prompt.

10 minutes later, the app is ready.

What do these Claude Code Skills Actually Change?

The real thing is not the apps.

The way we build apps.

Without overwhelming.

Without hitting limits.

Without reading paragraph-long responses.

Even if you don’t have a limit problem, you’ll like them.

Because they are also practical.

Look at that response.

Concise, short, but enough.

In the next sections, we’ll discover these 3 skills.

And build apps using them.

You know what I built.

I’ll also give you the prompts to build them.

And the app files, so you can install them.

Read more

I asked it about conflicts with my neighbor, my family, and sometimes even shared my own thoughts.

Yet, I never thought about building something about this.

I assumed not many people were doing it.

It turns out I was wrong.

According to a 2026 review in Digital Public Health, roughly one in four AI users do the same.

So you also asked about it.

But one wrong message can cause chaos during a crisis.

That’s why I always follow a pattern.

When I’m angry, I turn to Stoic philosophy.

And I try not to focus on what I cannot control.

Today, I turned it into a system.

I built my own Sigmund Freud with Claude.

Let me show you.

What Does This AI Therapist Actually Do?

I named my Sigmund Freud the Mind Mirror.

Because AI will analyze all my conversations and profile me, using its methods.

(I’ll show you how.)

At the beginning, it asks you to choose your preference.

Because let’s face it, there are many different belief systems.

You don’t have to fit into just one.

That’s why I added a custom option.

You can use it to turn the system into your own digital Sigmund Freud, or anything else you choose.

But to make it accurate and help it understand you better, it needs a lot of data.

So, where can you get that data?

If you keep a diary, that’s perfect.

Use it.

I used my conversations with AI because it already stores every chat you’ve ever had.

Three years of mine were sitting there.

You just need to export them, and I’ll show you how in a moment.

Now, let me show you a real conversation from a day when I was struggling.

Now read the answer again. It directly quotes what I told it months ago when I was feeling anxious:

Sweat-producing exercise calms me.

It knows that I ignore my neck pain until I collapse.

I never said any of that in this chat.

It pulled this information from three years of my history and 8,303 indexed memories.

How the AI Therapist Reads 32 Months of Chat History?

I’m going to show you how to access your memories, but first, let me show you another feature.

This second screen shows how your mind has changed over time.

It creates charts based on every message I’ve sent to AI since August 2023.

That’s 32 months of my mind in one graph.

Also, it summarizes each year with a couple of sentences.

I even forgot what I’ve done in 2023 :)

All of these were written by AI, which analyzes my conversations from my chat history.

How to Build Your Own AI Therapist

First, we’ll export the chat conversation history.

And we’ll profile yours using one prompt and a psychological method you prefer.

Next, we’ll build an app using only one prompt.

If you have built with me before, this is the same one-prompt method from $30M App Clone.

And this prompt is inside the Google Drive.

Here it is.

Read more

But a couple of days ago, a PhD candidate from Columbia University claimed that he solved six Erdős problems in five days using GPT-5.6 Sol.

And he shared the methodology, including his prompts.

If these prompts are powerful enough to solve decades-old unsolved mathematical problems, I wonder if this system can be applied to other problems.

Because essentially, you can turn everything into a math equation.

Let’s be honest: if you found a better-paying job with twice the salary, would not you quit?

We both know the answer.

But the questions ahead of us may not have such clear answers.

That’s why I needed his system to guide me.

So I turned it into a Claude Skill that I could use whenever I had to make a hard call.

I called it the Gencalculus.

(In middle school, my friends called me that. I solved math problems in my head faster than they could write them down. Gen + calculus.

The name stuck for years.)

I’ll show you what I built, but first, let me show you the winning formula of Shouqiao.

The Winning Formula

Shouqiao wrote six prompts to solve these questions, and all six follow the same eight steps.

I read every prompt he wrote; their structure looks almost identical.

That’s why I followed the same eight steps and added an interview in front of them.

Because your problem is probably not a math problem.

• Should I quit my job?

• Should I raise my prices?

• Should I hire someone?

So the interview turns them into options with a deadline and numbers attached.

Let me show you what I built.

Gencalculus Skill: A Graph That Makes Decisions

Now, let me show you how this Claude skill works as a graph.

When you ask a question, it creates three agents and a challenger that tries to break every answer.

The three agents work in parallel and never see each other, so they cannot all land on the same wrong idea.

Whatever survives the challenger gets scored and stress tested. If it fails, the whole thing loops back and runs again.

And the Gencalculus skill builds a Claude Code artifact at the end, so you can see the results, like this.

How to install everything in 5 minutes?

I give you a link to the GDrive, which includes all files.

After downloading this skill, open your Claude App and paste the “INSTALL-PROMPT.md” prompt.

And you will be done in 5 minutes.

After the paywall, I’ll walk through one of my own use cases, show the complete artifact, and explain how it works so you can adapt it to your own use case.

But first, here’s the Google Drive link.

Read more

Before researching Claude Skills for this article, I looked him up to see what he was doing, since he had left the company before I did.

I saw that he now lists a Generative AI certification and had liked a post about Claude Fable 5.

If your colleagues see you using Claude Skills, there’s a good chance they’ll ask questions like, “Can you trust this?”, “Are there any security risks?”, or “Is it safe?”, just like people did with OpenClaw.

And three months later, whether on LinkedIn, in a Trello card, or in discussions, trust me, you will see the same people talking about AI and using it in their workflows.

I have built more than 30 Claude skills for clients and worked with more people than I can count. Sometimes, more than I would have liked, but that is freelancing.

Let me show you a message that one of my former clients sent to a friend I worked with and me:

He was talking about the skills we built for him.

A well-written skill makes the same AI system ten times more useful, and creating one takes nothing more than a folder and a Markdown file.

In this guide, you’ll train your AI like a new employee using Claude Skills. By the end, you’ll be confident enough to build your own.

What Is a Claude Skill? Plain English First

A Claude Skill is a set of rules saved in a markdown file.

And the file that has this skill is often called SKILL.md.

Claude reads it before touching your task, so you stop repeating yourself.

Here is one of my Claude Skills.

This one audits my business. I feed it my newsletter stats, and it returns a ranked report of what is weak, what it costs me, and what to fix first. A consultant who read every number I ever produced. (Full build is here.)

But where do Claude Skills run?

Claude Skills run in Claude, Claude Cowork, and Claude Code. They also work in other agentic environments like Codex or OpenClaw.

Let me show you where to find Claude.

(In Claude, click on profile, settings, and skills to see yours.)

You can browse more skills developed by Anthropic after clicking “Browse”.

Toggle one on, and it is live in your next chat.

How Do Claude Skills Work?

Claude skills sit on disk as folders. Claude does not carry them around.

Every SKILL.md opens with two metadata lines, a name and a description.

When a chat starts, only those two lines get injected into Claude’s system prompt. That costs 30 to 100 tokens per skill, the price of one long sentence. The rest of the file stays on disk.

When your task matches a description, Claude pulls the full file into context and follows it.

Scripts and reference files in the folder load only if the job calls for them.

Anthropic calls this progressive disclosure. This is why 30 installed skills cost you almost nothing until one fires.

Calling Claude Skills Using Commands

A Claude skill fires in two ways.

The first is manual. You call it by name. In Claude Code, you type /skill-name, and it loads on the spot.

It asked for my goal, and check this out, it even asked in my native language, Turkish. It connects to your Claude account and analyzes it automatically.

Calling Skills using Trigger Words

The second is the interesting one, trigger words. Claude compares your message against every description on the shelf, and when the words match, the folder opens on its own.

Here, the same consultant skill was fired twice.

First run, I typed the command.

Second run, I typed a plain sentence, audited my business, and the trigger words caught it.

How to create a Claude skill?

Let me show you the fastest way.

Just paste the prompt I’ll give you in a second.

The link inside the prompt has the page inside the Claude Platform document, explaining how to prompt with Claude Fable 5.

These are the newest models, and the team has changed how they should be prompted.

Instead of learning everything from scratch and wasting tokens, let’s create a Claude Skill that learns directly from the official documentation the team has shared.

Paste this prompt to Claude.

I want to create a Skill from this system card:

https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5

The Skill should optimize my prompts based on this.

With just one prompt, you started training your AI.

You don’t have to learn how to prompt Claude Fable 5 in such depth.

Let Claude learn it for you.

Just skim the document, build a few projects with Fable 5, and before you know it, you’ll have learned how to prompt it effectively.

By the time I finish writing here, the skill is ready.

Now, I have read all of the document, and if you want to learn how to prompt with Claude Fable 5, you can read this one, where I explained.

Let’s test our skill.

I crafted a prompt using the old way.

You are the world's best software architect, senior frontend engineer, UX designer, and product manager.

Think step by step. Be extremely careful. Before writing any code, analyze everything in detail. Do not make assumptions. Consider every edge case. Create a comprehensive plan first, then explain your reasoning, then implement the code.

I want to build a Bloomberg Terminal-like dashboard for stock investors. It should look modern, include watchlists, charts, news, financial statements, earnings, analyst ratings, insider trading, macroeconomic indicators, and AI summaries.

Write production-ready code with best practices.

Let’s call our skill with command and ask it to update this prompt for Claude Fable 5.

Here is the optimized prompt.

Here are the reasons for these optimizations.

Pretty amazing, right?

Also, the new method lets you create a skill by recording your screen. To discover this, read this one.

Where to Find Ready-Made Claude Skills

You do not have to write your first Claude skill. Thousands are ready to steal.

Start inside Claude. (Head to Profile / Settings / Skills first).

The Browse button you saw in settings lists skills built by Anthropic, one toggle and they work.

Here are the skills for you to install.

The bigger shelf is skills.sh. There are 800K+ skills available here.

It is an open directory of community skills; each card shows install counts and a security audit.

Find one you fancy, and let’s say this, caveman.

It’s developed to cut your token usage by 80%, so before reading a post with a 10-item to-do list, download this, and do not hit the limit.

How? The easiest way is this prompt (paste it into Claude Code):

Hey Claude

Download and install this skill: https://www.skills.sh/juliusbrussee/caveman/caveman

Let’s see.

To see the Skill you just downloaded, close Claude and reopen it.

I once built a Wall Street terminal using multiple APIs to track my investments.

Now I want to build something similar again, but this time I want it to feel closer to a Bloomberg Terminal, which costs around $32K per year.

I also do not want to spend too many tokens, so I use the /caveman Skill for this build.

/caveman
I want to build a dashboard similar to the $32K Bloomberg Terminal.

Use this project as the reference, including its architecture, code, and API keys:

/Users/learnai/Desktop/LearnAIWithMe Projects/Wall Street Analyst

I can’t stop building with AI, sorry.

But we are in the era when people can build 99% of their paid subscription with Claude code, without being technical, and most of them don’t even know.

I used to build wall street analyst using Claude, now I want to go one step further, but I don’t want to explain what I want from the beginning, and don’t want to spend a lot of tokens, that’s why I inherited my old project, used one prompt, and the caveman skill.

By the way, the terminal is ready, even before I finished writing these sentences.

If this were a Build-It article, I would explain what every component does: one brain that fetches live market prices, economic data, and news, one screen that paints it all like a Wall Street trading desk, and one AI analyst sitting inside who reads everything and tells you buy, hold, or sell.

But this article is about Skills, so I’ll skip that part.

As you can see, with a single prompt, you can build your own market terminal. What a fascinating time to be alive.

What Not to Do When Creating a Skill

Now I am going to give you eight mistakes that I see most often when advising clients to use their skills.

Next Step

Log your daily routine and define what you do repetitively.

Next step is to turn them into skills, but I always suggest you stay in the loop because human touch is the most important and irreplaceable part, it was, and it always will be.

I’ve built 22 skills in the last 6 months.

And the list keeps growing day by day, because you can do anything using the skill.

Example?

Let me give you 5 of them that I’ve built with one sentence explanations.

AI Second Brain: It turns your past conversations, projects, notes, and results into a living AI memory that learns from your work and suggests what to build, write, or do next.

Job Hunter: It searches Indeed, Upwork, and LinkedIn for matching jobs, researches companies, writes cover letters, triggers by talking only, and sends the best opportunities to Slack.

The Red Team: It stress-tests your ideas, launches, or business plans by attacking them from multiple angles to find critical weaknesses before they become expensive mistakes.

The SEO Optimizer: It scores your article, fixes its SEO section by section without changing your voice, then scores it again to show the improvement.

Reverse Engineer Pipeline Skill: It studies any successful product, breaks down how it works, and turns the research into a build prompt that Claude Code can use to recreate its core engine.

As you can see, there are too many options available.

So start with smaller steps, and start building.

Don’t overthink, build it, and it will get better.

Thanks for reading.

This sounds like a business opportunity, because at that time McKinsey and Deloitte sold these reports in six figures, $500K to $1M per commercial project.

I first got very excited and even bought a domain name, learndatalabs.

But I could never find the time.

A few days ago, I came across a Y Combinator startup.

They are automating the kind of work that normally ends up in a $500K McKinsey consulting report, using AI.

Then I found another company: Operand. Its YC launch title says it all: “An AI to Kill McKinsey.”

One drawback of these AI-generated reports is that they are obviously AI Slop.

And even the big firms like Deloitte can fall into this trap.

They shipped a 237-page report to the Australian government on a $440,000 contract, which carried a fabricated quote from a federal court judgment and citations to academic papers that do not exist.

In the end, Deloitte refunded $97,000.

It looks like even the big companies are using AI in their reports. Why would you not?

You do not need a private equity fund to need a report, because the same thing happens when you dig into a market you are curious about, when someone asks you to look at their business, or when you plan a trip with real money behind it.

So I did what these firms do, then added the step they skip.

But this time, every claim gets verified against the source it cites, and anything that fails verification gets dropped.

I wrapped it into a Claude skill.

Let me show you what I built.

The Market Research Report I Built With a Claude Skill

Here is the cover of my Market research report.

The line under the date is very important.

Thirty-four claims made it into this report, and every one of them was checked against its source.

It started with the answer instead of slowly building toward it. You already know the question because you gave the first instruction to create the market research report.

This section reports the players in your field and their pricing tiers. It can kill the idea or show you where to sit.

Every number in the report carries a superscript, and clicking it lands you here on the source it came from.

It sizes the market two different ways and reports the gap between them.

In this one, published research said $350 million, and the bottom-up build said $33 million.

Not verified sources. I added this to the report to show how the sub-agents I give task for to double check, eliminate false claims.

It creates an 11-page report in just 15 minutes by launching multiple agents, validating every claim, and using several skills to produce the final PDF. Even the cover image is generated with AI.

How does this Claude Skill work?

The skill asks three questions before it starts.

• Which market and geography

• Target audience

• Why it matters

Here are my answers;

• Paid AI Interview prep product

• Just me

• AI interview prep in the US and India between 2024 and 2026

Next, the skill creates five sub-agents: one for market size, one for competitors, one for demand, one for distribution, and one whose only job is to argue that the entire idea is bad.

Once they finish, three more agents reopen every cited source to verify that every claim is actually supported by the source.

How to Install This in 5 Minutes

Everything I built sits in one folder.

And you can install everything, using only one prompt.

Here it is.

Set up the market report skill from this Google Drive folder.

[Paste-the link or Upload the file]

Download every file. 
Read 0-START-HERE.md first, it says which file goes where.

Create skills/market-report/scripts and skills/market-report/templates, then put each
file in its place under the name that file says, dropping the number prefix.

Install what it runs on.

  pip install playwright openai pypdf && playwright install chromium

Then prove the renderer works before we do any research.

  python3 skills/market-report/scripts/render_report.py --fixture sample.pdf

Open that PDF and show me it came out. 
If it did, tell me the skill is ready and stop
there. 
Do not start a report until I give you a topic.

Now here is the link to this skill.

Read more

This week, I was testing another tool and came across this one. Buzz.

It runs agents on Claude Code and Codex, so I gave it a shot.

What makes Buzz different from a Claude bot in Slack?

It looks like Slack. I saw agents in the sidebar and felt nothing. Slack has bots, too. I connected Claude to Slack months ago, and I talk with it every day.

But my gut said test it. I am glad I listened.

This is far superior to agents on Slack or any other platform.

How do Buzz agents run on Claude Code and Codex?

These agents hold a role, sit in a room, and run on my existing Claude Code and Codex subscriptions. You can even write their instructions.

Look at mine, Fizz, Honey, and Bumble come as defaults. The setup asks you to pick a default model.

What is inside my three-agent AI team?

I built three AI agents on top: the Number auditor(Claude Opus 5.0), the quality gate(GPT 5.6 Sol), and The Memory(Opus 5.0).

Then I added them to one room and wrote their instructions so they know they are a team.

The room name is idea.

Because we all know that one idea can change our lives.

Now let me show you how they work together and how you can adapt them to any business.

What did I build?

The team I built is designed to work together, because I wrote their instructions that way.

How the work splits across Claude Code and Codex

One prompt that I pasted initate the entire process, so 3 agents, 2 claude code and one codex, are now working on my data.

The thing I want is to build a dashboard(Claude Code agent), be a referee to this Claude Code agent, and tell it to improve the parts needed(Codex), and remember everything in the next sessions, memory(Claude Code)

Now, I am building a dashboard of my Substack statistics.

You can build your sales dashboard.

Your ad spend report. Your inventory numbers. Your client hours. Your churn by month. Swap my two CSVs for whatever your business exports every week, anything with number works.

Pass 1: The Quality Gate blocks my numbers

Look at how the quality gate assesses the CSV, and check if there is a misleading column.

At the sametime the number auditor started doing the analysis.

Next, the number auditor lists the 4 risks that can distort my numerical analysis.

And the quality gate give score and tells that this is not ready.

The Number auditor solved the issues and sent the data to the quality gate.

It's approved!

What did the AI team find in my Substack data?

Here is one of the findings the Number auditor finds. (I am not writing them to increase my paid subs, but good to hear :)

Also, after 5 minutes, they are exchanging ideas and thoughts, scoring the analysis, while I am writing this one for you, how crazy is that?

This is the closest thing that I built to the AI team.

The dashboard my AI team shipped

After 10 minutes of exchanging messages between the number auditor and the quality gate, everything has been approved.

And my dashboard is built, with the verified numbers.

LearnAIWithMe has 750K views in the last six months. The analysis comes after this traffic is the one I care about the most, but I can not show you that part; it is a bit confidential.

Thanks to our team, the dashboard runs on verified numbers, which is the most important thing in this workflow.

AI can make very basic mistakes in data analysis. It buries the metric that matters and inflates the one that does not. Two models checking each other cut that risk.

You can also invite people into these channels, which is where it gets interesting.

Next, what Buzz is, how to install it, how to create agents, and write their instructions. I am also giving away all three agent instructions.

Read more

And these days, I am testing ChatGPT’s new features like ChatGPT Work, GPT 5.6 Sol and Codex, so as not to stay behind.

And I must admit, some of them are really powerful.

If you recently switched from ChatGPT to Claude, you know Claude doesn’t have a voice feature as good as ChatGPT’s.

It looks like OpenAI realized this could be a point where they could differentiate themselves from Claude, and now ChatGPT Voice can be used within Codex.

Meaning, if you give the Codex app computer access, you can do whatever Codex can do just by talking.

You can imagine what I felt when I read this. I must try immediately, because it looks like an opportunity for me as a new father.

While testing it, I came across a feature most people still don't know about, and I did not expect it to be this powerful.

If you connect your phone the Codex app, using “Remote”.

You can now control your computer from anywhere through your phone.

ANYWHERE.

And you don’t have to type like you have to while controlling your Claude via Dispatch, or control your OpenClaw/Hermes by typing via Telegram.

Now all I have to do is connect the AirPods to my phone, open the ChatGPT app, activate the remote, and convince my wife that I also might talk with Juniper, because I have to finish something.

The only Juniper she wants is the Tesla one, but I convinced her.

I left the PC on at home (haven't tested my Mac mini yet).

While everyone is talking about prompts, loops, and graphs, I’ll be outside, near my newborn girl, my wife, and Juniper, who can access everything.

Amazing, right?

Today’s walk, I built a working Claude Skill this way.

It cost me 20% of my AirPods’ battery, 5% of my iPhone’s battery, and 10% of my Max plan’s limit.

When I got home, this skill was sitting on my computer, finished.

Before showing you, let me explain to you how I decided to build this.

Why I Built the News Truth Barometer With Codex

Yesterday, while walking with my wife, we started talking about an AI news story I had seen on Twitter.

I told her Sam Altman was going to Washington to meet with government officials because GPT-6 was about to be released.

I did not remember the exact X post I read, but she told me this is not the story.

The Washington visit was real, but the GPT-6 part was still speculation.

We talked about it for a while and checked the story when we got home.

She was right.

How much can we trust the news we read online?

Today, after planning to build something to show in my post, I remembered this.

Maybe I can build something that detects the credibility of any news given.

Also, I have to go out for a walk with my wife, baby girl, and my dog, and I love this Codex Voice mode.

So I built a Skill that scores news.

What I Built With Codex

The skill I built can take a headline, a link, or even a rumor. (NEWS CLAIM)

It gives the same claim to three agents.

One checks Primary News, one checks Independent News, and Skeptic tries to find everything that might be wrong with it.

Then, a separate Judge compares their findings and gives the story a truth score.

While I was testing the Skill, Codex also built a small barometer app that turns the result into a visual.

After a few attempts, Codex built the skill and tested it with mock data.

But the issue was that it could not send the SS to me, so I told it to email it to me.

I had already given it computer use, so it did.

It opened Gmail on my computer, attached the image, wrote an email to me, and here it is.

I looked at the results and wanted a small change from Codex.

It then updated the Skill, tested it again, and it worked.

This time, I gave it the real story that my wife and I had discussed:

Sam Altman is visiting Washington before a GPT-6 launch.

The three agents researched it independently.

• Primary News gave it 82%.

• Independent News gave it 72%.

• Skeptic gave it 72%.

Then the Judge agent reviewed all three reports and gave the final result:

72% TRUE.

The Washington visit and the meetings with US officials were supported by credible reports.

But the GPT-6 part was still unverified.

And this is the visual the Skill created:

Then I asked Codex to email this result to me too.

It wrote another English explanation, attached the new screenshot, and sent it.

Now I am going to give you the exact Skill files and show you how to install them in less than five minutes.

But I will also show you how to install the Codex app on your computer, connect your phone using Remote, and enable Computer Use so Codex can actually use your computer while you are away.

Then I will give you one Terminal command that keeps your Mac awake, so you can leave home, connect your AirPods, and continue building from anywhere.

Here are the skill files.

Read more

And it is very similar to Claude Cowork.

So, I compared Claude Cowork and ChatGPT Work by building Cal AI, a $30M app used by millions of people to track their daily calorie intake.

Why build instead of clicking around?

Feature tours without doing anything do not help you. A real build does.

Look at this guy I found on Starter Story.

He copied two apps and made $30M combined.

The market is full of successful copies.

Pick a product you care about, then build a better version.

This sounds controversial, but it works.

To do that, we’ll use ChatGPT Work and Claude Cowork because;

• ChatGPT Work has “sites”.

• Claude Cowork has “artifacts”.

That’s why this will be a fair comparison.

So, I’ll use a reverse-engineering pipeline that I built earlier.

Let me explain how it works first.

What Is the Reverse Engineering Pipeline?

In one sentence: it researches the product you give it, trains a NotebookLM, gathers additional information, and turns everything into a single comprehensive prompt.

And this one big fat prompt will let you build anything you want.

Like I always do, I wrapped this up into a Claude Skill, so you can use it, using one single prompt.

Use the reverse-engineering pipeline on [any product] and write one big, detailed prompt for me.

So we’ll decode this $30M App, Cal AI, using the skill.

In the meantime, you’ll see how Claude Work and Claude Cowork create different apps from the same files, and we’ll compare those two together, but first, let me show you the results as usual.

What ChatGPT Work vs Claude Cowork Built From One Prompt

We built the same calorie tracker twice.

ChatGPT Work: Building Alma in 74 Minutes

First, here is the one ChatGPT Work created. It named the app Alma.

Alma is a desktop style web app with a 4 step onboarding.

You need to fill your age, height weight, target pick and your activity level.

Next, it calculates your calorie intake using the Mifflin-St Jeor Formula.

Then it drops you into a daily dashboard.

Claude Cowork: Building MacroLens in 52 Minutes

Next, we built the same app with Claude Cowork.

It named this one MacroLens.

MacroLens went mobile first.

The whole app lives in a phone shaped frame.

The onboarding runs 7 steps and asks a question ChatGPT skipped, how fast you want to lose weight, from a steady 0.25 kg per week to an aggressive 0.75.

The meal scan is the strong part.

I uploaded a photo of Turkish döner and it split the plate into ingredients, döner meat, pide, pepper, each with a gram slider and macros pulled from a USDA table.

It caps free users at 3 scans a day.

I created both apps using one prompt.

The Reverse Engineering Pipeline wrote this prompt for me by collecting information about Cal.AI and then created a notebooklm and assigned an agent to ask questions about how to build.

By the way, both apps came from one prompt.

The reverse engineering pipeline skill wrote that prompt for me in 5 minutes.

I’ll give you the prompt and the skill below. We’ll dig into ChatGPT Work and Claude Cowork as deep as we can on the way.

Let’s start. Here is the skill and the prompt.

Read more

It is so powerful that Anthropic even had to merge a team before release.

But it has one drawback, the price.

Now Anthropic put it under Max plan, and even if you pay max plan, you’ll have seperate Fable 5 limit.

That’s why this headline got my attention, the GPT 5.6 Sol.

According to the Benchmarks that I checked, it was almost as good as Claude Fable 5.

But who understands the benchmarks right? I need a more plain explanation, like speed, price to performance, which I created for you.

I like OpenAI. It was my first love, but Claude is still my favorite because its ecosystem feels better. OpenAI also keeps adopting ideas Claude introduced first, such as Cowork and scheduled tasks.

Still, I keep a ChatGPT Pro subscription so I can test OpenAI’s latest models.

Last week, while working on an intense project, I was close to hitting my usage limit. That made me wonder: is there a way to use Codex directly inside ChatGPT?

It looks like many people had the same idea I did because there’s already a plugin that lets you use the Codex plugin inside Claude Code.

But it looks a bit complicated to me.

To make it clearer for you, I wrote a Claude skill that lets you use Codex inside Claude.

But my Claude Skill has more features, let me show you.

The Codex-Wingman Skill

The Codex-Wingman skill turns GPT-5.6 and Claude Fable 5 into an amazing team.

When you’re about to hit your Claude usage limit (around 90%), Codex takes the wheel.

This Claude Skill works in three modes.

• Referee: In this mode, GPT-5.6 reviews Claude’s code like Sam Altman. It tells Claude where to agree, where to trust its approach, where to disagree, and what to improve.

• Failover: In this mode, GPT 5.6 takes control when the limit hits 90 percent, Claude stops coding and directs the tasks to the Codex.

• Verdict: In this mode, GPT-5.6 checks every handoff and decides whether the work is ready to ship, needs a few changes, or should be rejected.

With one Claude Code skill, your two AI subscriptions finally work as one team, and the result is better than either one alone.

Why This Works?

If you use Claude before, you know, hitting the limit is like another Sunday morning activity.

And sometimes, you need to keep going, so I see people even suing the Claude team :)

Or paying $500M for the API bill.

Also, I want these two to work together because we all know the Fusion techniques; it surpasses the Fable 5 intelligence, and it lets multiple model runs at the same time.

So using two different models for the same tasks can be really good because both AI use different parameters, and the combination will be better than both, because the training datas are different.

But enough with technicalities, let me show you how to install this skill and how it works in action.

Read more

The changes affect Ultra plan users for now. Only one of them affects you.

If you are on the Pro plan, Google will give you a cloud computer, so you can run your code inside Gemini Notebook, too.

Pretty crazy. When they make it available to Pro users, I’ll test it and report back.

Now I want to show you how to build with Gemini Notebook.

You don’t need to be technical.

I’ll show you everything step by step and give you the prompt.

That’s it.

3 prompts, 3 apps, in less than 30 minutes.

Sounds interesting, right?

Let’s start.

How to Build With Gemini Notebook?

Visit Gemini, from here: https://gemini.google.com/app

Next, click on “+”.

And scroll down to more uploads and select “Notebooks”.

Next, choose the notebook you want.

Next, click on “+” and select “Canvas”.

Next, paste one prompt.

Believe it or not, that’s it!

Next, it’ll start building your app.

After a few minutes, your app will be ready.

Also, let me show you how to train NotebookLM.

How to train NotebookLM?

Watch the video above, or follow the steps below.

1. Visit here.

2. Click on Try Gemini Notebook.

3. Click on “Create a new Notebook.”

4. Click on “Websites”.

5. And paste the YT Video you want(or videos, one by one.)

App 1. The Course You Actually Finish

To build an app, you only need one NotebookLM.

Train it by following the steps in the previous section.

Next, go to Gemini and follow the steps I just showed you:

1. Select NotebookLM,

2. Choose Canvas

3. Paste the build prompt.

Here is the prompt for you to copy/paste.

The link to the YouTube video series: https://www.youtube.com/playlist?list=PLXV9Vh2jYcjbnv67sXNDJiO8MWLA3ZJKR

Here is my screen;

5 minutes later, the app is ready.

Let me show you the app.

As you can see, this app turned the entire MIT Foundation Models and Generative AI course into a 14-day study plan.

Day 2 does not unlock until you successfully answer Day 1’s questions. It also shows you where to improve and what to review after each mistake.

And it takes only five minutes to build.

After creating the app, click on “Share”, copy the link, and share it with your friends.

Also, you can do these things using a free account.

App 2. The Tutorial That Becomes a Checklist

This app turns your YouTube video into a checklist you can tick off. By the end, you’ll know you completed everything covered in the video.

For the setup, we’ll repeat the same steps from the previous section. Check it out first if you haven’t already.

In short:

• First, train a NotebookLM.

• Next, connect with Gemini.

So I did the same.

Here is the prompt for you to copy/paste.

Here is the YT link that I used to train NotebookLM:

youtube.com/watch?v=B_KAEqiC-0Q

Let me show you the app.

Look how it turns a YouTube video into a step-by-step course you can follow.

Let’s say you want to learn how to do [X], but you’re a beginner and feel overwhelmed by the full video. Train a NotebookLM, connect it with Gemini, and it will guide you through the process in the right order, one step at a time.

App 3. The Practice Interviewer

Now we’ll do an interview with Boris Cherny, the creator of Claude Code.

Follow the same steps we followed in App 1.

Here is the prompt for you to copy/paste.

Here is the YouTube video I used to train NotebookLM: youtube.com/watch?v=B_KAEqiC-0Q
In this Youtube Video, Boris Cherny explains how to master claude 30 mins.

Let me show you the app.

As you can see, I answered his questions, reviewed his feedback, and completed the interview.

What’s next? Build your own interviewer using a different video and prepare for your next interview.

What should I do next?

Pick one video you want to work on, and build your own app.

Share it with your friends, let them break it, and improve it day by day.

The next natural step would be connecting NotebookLM with Claude Code to build production-ready apps.

If you don't know how to use Claude Code, no worries. I have a Claude Code MasterClass for everyone, so you can start learning today and build your own app in hours.

Life is practical; don’t stop after learning.

Start building!

Thanks for reading.

Now we have graphs. And they were introduced by the creator of OpenClaw.

And everyone is talking about how to build graph.

But I did something different. I wanted to build a graph that works in a real business, one that people pay for.

So I searched Upwork and found this job.

I read the entire posting and saw that the client listed 10 agents.

In a sense, they were describing a graph. Here it is.

Before going deeper into this graph, decoding it, and building the same thing with Claude, you need to know three terms. Prompts, loops, graphs.

Prompts, loops, graphs in plain words

I have one sentence for each of these terms. Here you go;

• A prompt asks AI to do something once.

• But a loop makes it retry until the work passes a check. (They have goals)

• A graph connects many loops and agents into one map, with a human standing at the gates that matter.

You know prompts. I already showed you loops, including 11 real loops used by different Substack writers.

This time, let me show you the graph I built to handle the kind of $135/hour Upwork work clients pay for.

What we will build: The 135/hour graph

The $135/hour graph we’ll build has three agents and one routine.

Let me explain the three agents(wrapped up as a Claude skill) first.

• Intake agent: Reads every incoming email. The task is sorting them into five categories.

• Proposal Agent: Prepares drafts, but the pricing part is left empty; it needs a human.

• Verify Agent: It scores every draft. This part is very similar to loop, below 85, send it back, and give up after three rounds.

The Claude routine we’ll create makes them work together. Every weekday at 7 AM, or at any time you prefer, it runs all three. If a draft passes all three stages, it is sent for human approval.

The crazy part is that I turned everything into a single prompt, so you can install this team and the routine with it.

Here it is.

AI Academy

Until August 1, Annual subscribers can also access the AI Academy inside the Vault.

This is where we build with AI, share our work, and learn together.

After August 1, access will be limited to Inner Circle members.

Read more