Episode 1｜June 3: I started building a company inside AI

I'm building a school in Nepal. The first thing I built was an "empty box" of a company.

This is not a story where a successful person teaches you, "Use AI like this and you'll succeed."

I'm not a programmer. I'm not an AI engineer. I can't really write code.

And business hasn't always gone well for me either. I started businesses, failed, got betrayed, lost money. And still, I start the next one.

Even when I make a plan, if another idea comes up halfway, I go chasing it. My girlfriend has told me,

"You can't even make a plan." "You don't do what you say you'll do."

I know it myself.

But my big goal hasn't changed much.

I want to build a company that connects Japan and Nepal.

Teach Japanese. Grow people. Prepare them to work in Japan. And connect them all the way to a job. At the center of the ABAS GROUP I want to build is this work in education and people.

But on June 3, 2026, I had no one to run it with. Not much money, either. No one to look at the finances, no one to organize sales, no engineer. Thinking, deciding, moving — it was all me.

That day, I started something strange.

I tried to build, inside my computer, the company roles I couldn't fill in real life.

This is where my record with AI begins.

About the record of this day

This article looks back at June 3 using screenshots of my remaining ChatGPT history, files left on my computer, photos of my screen taken that day, and my diary.

The actual conversations I had with Claude Code back then are gone. So I keep separate what remains in words, what the file timestamps show, and what I think now looking back. Times are Nepal time.

A file's creation time tells you when the file was made. But it doesn't tell you what I typed just before. This article isn't a replay of every action — it's a day rebuilt from the records that remain.

12:38. "What should I do?"

One of the first questions left in my ChatGPT history that day was this.

ChatGPT's answer started by listing my businesses.

Cars, construction, a poultry farm, a Japanese language school, recruitment, registered support, sending workers, boxing, farming, exports.

And it said:

That's AI's view, not a fact. But every business on that list was something I was seriously thinking about at the time.

After an exchange at 12:49, I asked one more thing.

I wanted to run a Japanese school. But I didn't know where to start.

13:06. Already thinking about another business

Right after, I sent the URL of another website. A site called Political Money DB.

I had just asked "What should I do?" and I was already talking about a new service.

From there, an idea for a talent database came up too. Japanese level, work history, qualifications, desired job, interview results. Could I organize the information of Nepalis who want to work in Japan and connect them with Japanese companies? Could I build a LinkedIn for jobs in Japan? Around then, I was even thinking about a housekeeper matching service.

Looking at it now, it's completely scattered.

I think I'm good at finding interesting things. But in business, that's also a weakness.

Every time you spread out on a new idea, you have less time and money for the work you're already doing.

I've failed because of that many times. The "getting the order wrong" that AI pointed out at 12:38 — I was already doing it 30 minutes later.

14:45. Installing Claude Code

My computer shows that a settings folder for Claude Code was created at 14:45.

Claude Code is a tool that lets the AI Claude read and create files inside your own computer. But back then, I didn't even know the difference between Claude and Claude Code.

"Where is Claude Code?"

That's where I started. Where do you use it from? How do you run it on a Mac? I looked things up one by one. In the end, a photo from that night shows me using Claude Code on the right side of a screen made for writing programs (VS Code).

Not all the setup steps remain. So I won't write today's steps as if they were "the steps I took that day".

But I remember how it felt. Not just getting answers — I could have AI handle files inside my own computer.

A "someone to talk to" becomes "someone who works with their hands next to you".

That hit me hard.

15:12. A folder to put the company in

At 15:12, a folder was created on my Mac's desktop. Today it's named "桜花 (Ouka)", but in a photo from that night it shows up named "ABAS". For about an hour and a half after this, nothing was added inside it.

But I was trying to put the structure of a company in it.

If there's no one to look at the finances, have AI think from a finance point of view. If there's no salesperson, have it organize the sales work.

"If it's just me, I'll get AI to help."

The idea was about that simple.

"Building a company" here doesn't mean registering a corporation. It means organizing the company's information, decision rules, roles and work procedures inside my computer. I didn't develop AI myself either. I was trying to assemble existing AI into a shape my company could use.

The YouTube video I was watching

This idea wasn't something I came up with from scratch. There was a video I was watching.

It was a video from the YouTube channel "かいちのAI大学 / Saix (サイクス)" (@kaichi_ai), titled "【知らないと損】Claude Codeで会社経営を行う方法とは？｜使い方・活用方法を徹底解説！" ("Don't miss out: How to run a company with Claude Code — a full guide to using it"). It's 38 minutes 52 seconds long.

https://youtu.be/rqdG7YnCK54

The screen shows instructions given to Claude Code.

First, make a folder called context. Put everything in it — your diary, your everyday thoughts, first-hand information. Then create staff divided into 5 roles: CMO (marketing), CPO (product), CFO (finance), CSO (sales), CTO (technology). The CTO's role is to gather all the technical information, such as prompts and how to use Claude Code.

I copied this screen almost exactly.

16:07. Please write the company's rules

I asked ChatGPT this.

"Colord" means Claude. Typo and all.

CLAUDE.md is the instruction file Claude Code reads when it starts working. Back then, I didn't fully understand it. But I thought: if I give it a file with the company's rules, I won't have to explain everything from the start every time.

ChatGPT replied that it would be better to make CLAUDE.md a "personality setting for Anish's personal management-advisor AI". It listed my problems — "too many ideas", "running too many things at once", "money gets scattered" and so on — and wrote:

That day, ChatGPT said the same thing twice in a row. Narrow it down.

16:43. Pasting the video screen: "Turn this into a document"

After that, I took a picture of the video screen and pasted it into ChatGPT. The only words I added:

ChatGPT said it would "organize this for your ABAS, based on the image", and made a document called "ABAS AI Management System Design". Its purpose section said:

Watch the video. Take a picture of the screen. Paste it into ChatGPT. ChatGPT turns it into a document for my company. Hand that document to Claude Code.

That was how I worked that day.

16:49–16:52. The shape of a company appears

Looking at the creation times of the remaining files, the boxes of the company multiplied in about three minutes.

At 16:49, folders appeared: context for records, Projects for projects, SOP for procedures, Reports for reports. Inside context: meeting notes, transcripts, diary, customer conversations. Then role files for CFO and CMO. At 16:50, CPO, CSO, CTO.

The same 5 roles as the video.

The CTO's job said "manage Claude Code" and "manage ChatGPT". When I first read it, I laughed a little, thinking I'd made AI create a role to manage AI. But tracing it back, it came straight from the video's description of the CTO.

At 16:51, the project files: registered support organization, paid job placement, construction license. A place for the projects I wanted to move forward.

At 16:52, a procedure manual for residence status applications and a template for weekly reports. SOP means "standard operating procedure" — a document that lets anyone do the work the same way. The weekly report tracked sales, cash flow, sales activity, number of students and number of job seekers.

The shape of a company was lined up on my screen.

But at this point, it's not as if a bunch of automatically working employees was finished. What existed were files defining roles and templates for organizing work.

Still, for me back then, it felt huge.

"This is amazing."

But when I opened them, there was a problem.

There's nothing inside

I open the SOP for skilled-worker placement. I open the SOP for residence status applications. They say:

Steps: 1. (fill in)

These two files have not been added to even once, from the day they were made until today.

The box is there. But there's nothing inside.

When I think about it, of course.

Where I had failed. What I check when I talk to Japanese companies. Where Nepali students get stuck. What I want to value as I build the company.

There's no way AI could know what I hadn't told it. You can copy a video's template. But what goes inside the template can only come from your own company.

I thought I had built a company. I had built an empty box.

17:25. Turning what's in my head into records

At this time, an SOP for "voice transcription" remains. Of the procedure manuals made that day, this was the only one with actual content.

Record. Transcribe. Clean it up. Save it in context. Sort it into meetings, customer conversations, the president's notes. If something was decided, add it to the project file.

The video also said to put your diary and thoughts into context. I'm faster at talking than at writing clean text from scratch. What I said in a meeting. What I talked over with family. The school. Worries. What I want to do next. If I turn that into text, I can give it to AI.

Turn the information that exists only in my head into records the company can use.

That direction appeared here.

But making a transcription procedure and actually recording and running every conversation from that day on are different things. June 3 was the day I built the entrance.

18:08. I want to talk about numbers too

I asked ChatGPT:

One of ChatGPT's replies was this.

This screen shows no time, so I can't say for sure it was the reply to the 18:08 question.

But this reply begins "As a management advisor, here's how I'd reply," and continues as text starting with "Thank you". It's written to be sent back to someone.

That day, I was going back and forth between two AIs. Another screen shows me pasting a question Claude had asked me — "8. Next steps (★ just tell me 3)" — straight into ChatGPT, and adding my own answer.

Claude asks. I paste it into ChatGPT. ChatGPT writes an answer. I take that back to Claude.

I was the messenger between two AIs.

19:56. "So how do I use this?"

The most "me" question of the day was this one.

At that moment, I took a picture of my computer screen with my phone and sent it along.

In the photo, the folder is named "ABAS". The roles had grown from the video's 5 to 7, with CEO and HR added. In SOP, a manual for corporate sales had appeared too.

The right side of the photo shows the Claude Code screen. There's a line I had sent:

Claude offered options A and B, and I told it to do both. On the same night I'd been told twice to narrow things down.

ChatGPT's reply began:

And it continued, "The skeleton is complete."

The skeleton was there, for sure. But open it, and there was a row of manuals saying "(fill in)".

I built it. It looked like a company. I was happy. But I didn't know how to use it.

That day, I wasn't using AI well.

I built it because I could. Only afterwards did I think about how to use it for work.

Even AI called it "playing AI company". This was my first stumble.

20:09. What remained in the diary

The diary for this day remains as a file. But it isn't text I wrote from scratch. The headings and bold text are so neatly arranged that I think Claude formatted and saved what I told it.

Here's what it says:

明日やること：値段・塗装・設備の見積もりを取る (Tomorrow: get estimates for price, painting, equipment)

相談した人：家族 (Who I consulted: family)

不安なこと：一緒にやる人がいない (Worry: there's no one to do this with me)

And at the end:

I think this is the most important part of the day.

On screen, I was building the boxes of a company. In real life, I had decided the location of the Japanese school. But the problem of having no one to run it with remained.

What AI can do, and what needs people. I didn't yet understand where that line was.

Four days later, on the night of June 7, I asked ChatGPT:

The answer was: don't "rent an impressive school" — "start with a contract you can exit". 0 students, 0 yen in sales, not yet open. That was my school at the time.

What I'll give you from here

Up to here, I wrote what happened on June 3.

From here, looking back at that order, I'll write concretely about:

• When you copy a video's template, what you have to fill in yourself

• How to connect what you built to actual work

• What to write in CLAUDE.md

• How to give your diary and meeting notes to AI

• What to prepare before adding more roles

At the end, I'll also include instructions you can adapt to your own company.

I keep the records from back then separate from the improvements I'd make now.

This is not a success story of automating a company with AI.

It's about where I got the order wrong, and how not to get stuck in the same place.

1. The order I went in, and the order I'd go in now

On June 3, I went like this.

Ask ChatGPT "What should I do?" Get told "you're getting the order wrong". Think of another business. Install Claude Code. Watch a video. Ask ChatGPT for the company's rules. Get told "narrow it down". Paste the video screen and have it made into a document. Make roles and manuals with Claude Code. Notice there's nothing inside. Think about how to put records in. Be the messenger between two AIs. And finally ask, "So how do I use this?"

Looking back at this order, the problem is visible.

I built the shape of an organization before deciding which work I wanted to make easier.

So now, I'd decide one thing first. Do I want to manage school payments? Organize sales to companies? Decide tomorrow's work?

One is enough to start. Gather the information that work needs, and then give AI a role.

2. Copy the video's template. Fill in the contents yourself

I copied the template from the かいちのAI大学 video almost exactly. Make context. Make 5 roles. I don't think that itself was wrong. It's much faster than thinking from nothing.

What I got wrong was thinking "the company is built" once the template was in place.

The video maker's company and my company are different. Even with a CFO file, my school's rent and payment schedule weren't written there. Even with an SOP file, what I check when I talk to Japanese companies wasn't written there.

So now, on the day I copy a template, I'd put in at least one piece of content.

After making 5 roles, write the numbers you know into one of them — the finance file, for example. Write "unknown" where you don't know. After making a procedure manual, write at least one procedure you actually do.

What you can copy is the shape of the box. What goes in the box can only come from your own company's records.

3. If you use two AIs, don't become the messenger

That day, I pasted Claude's questions into ChatGPT and took ChatGPT's answers back to Claude.

ChatGPT had heard much more of my story. So it felt like asking ChatGPT would give answers that understood me better.

But if you keep doing this, the company's information gets scattered in two places. What ChatGPT knows, Claude doesn't. The files Claude made, ChatGPT hasn't seen. I'm standing in between, copying and pasting every time.

Now, I'd decide on one place to keep the company's information. For me, that's inside the Claude Code folder. Even if I think something through with ChatGPT, what's decided gets written down as a record in that folder.

Use as many AIs as you like. But keep the company's records in one place.

4. Four things are enough to give it first

The company's purpose

For me, it's education and people work connecting Japan and Nepal. If you don't write this, every interesting new business ends up weighing the same. Like the 10 businesses lined up in the 12:38 reply all looked equally heavy.

Current numbers

Sales, fixed costs, cash on hand, number of students, unpaid tuition.

Write "unknown" for numbers you don't know. Write "assumption" when you assume. A document with facts and assumptions separated is easier to make decisions with than one where every number is filled in.

Projects in motion

For the school on June 3, the next tasks you can see from the diary are:

• School property: status in the record: location decided／next to check: contract terms and payment terms (I actually asked ChatGPT about this on June 7)

• Painting: status: estimate needed／next to check: price and schedule

• Equipment: status: estimate needed／next to check: what's needed and priority

• Operating team: status: no one to do it with／next to check: what work is needed and what kind of person

The "next to check" items are what I organized looking back now. It doesn't mean I did them all that day.

Decision rules

I needed someone who would pull my priorities back, not just add more ideas. So look at profit, feasibility, time and cash flow.

Sort new businesses into:

Do now／Later／Don't do

Put work that strengthens the existing school first.

5. What to write in CLAUDE.md

CLAUDE.md is the instruction file Claude Code reads every time it starts working. You can use it to pass on your company's information and decision rules without explaining them again every time.

But writing it doesn't guarantee it will be followed. The official documentation says CLAUDE.md is not an enforced setting and there's no guarantee of strict compliance. It also says that the more specific and short the instructions, the more consistently they're followed. (Claude Code docs: "How Claude remembers your project", checked October 2, 2026)

What I wanted was things like:

"Don't agree with my new ideas right away."

"Look at the school's numbers first."

"Decide concretely what to do today."

Below is not the actual file from back then, but an example of the instructions I'd give on day one now.

I run ABAS GROUP.

My long-term goal is to build an education and talent group that connects Japan and Nepal.

My short-term goal is to make the Japanese school in Nepal profitable and grow it to about 100 students.

Prioritize the profit of existing businesses, cash flow, student recruitment and systematizing work over new ideas.

Sort new businesses into "do now", "later" and "don't do", and explain why.

Separate facts, assumptions and unknowns. Don't treat unknown numbers as actual results.

Organize answers in this order: current situation, problems, priorities, concrete actions, budget, deadline, probability of success, next step.

When the necessary numbers are available, calculate monthly profit, required funds, break-even point and payback period.

If there's no basis for judging the probability of success, don't force a percentage; show the conditions for testing it instead.

Split the school plan into Phase 1 (0–20 students), Phase 2 (20–100) and Phase 3 (100+).

Finally, narrow today's tasks to at most three.

What mattered wasn't the names of the roles, but the standard by which I wanted work done. What I needed wasn't agreement with new ideas, but a standard for choosing what to do now.

6. context doesn't become work just by saving it

I was trying to put my diary, meeting notes, customer conversations and transcripts into context. To gather the company's information in one place.

But if you think that just making the folder and saving things means AI will read all of it every time, you'll get stuck again.

Specify which records to read, and what you want out of them.

For example:

Separate what's decided, what isn't decided yet, and what needs payment.

Narrow tomorrow's tasks to three, and give each a person in charge, a deadline and a completion condition.

Leave anything not written in the records as unknown.

This way, saving records connects to tomorrow's actions. Turning what I said into clean text isn't enough.

Take it all the way to deciding the next piece of work.

7. Before making "AI staff", give it one job

CFO, CMO, CTO. When roles line up, it feels like a company has been built.

But a file with a role name and an AI that works independently are different things. In Claude Code, "subagents" — separate AIs for each role — have their own place and format for definitions, and a way to call them. (Claude Code docs: "Subagents", checked October 2, 2026) The files in the Roles folder I made on June 3 were not that.

I don't think you need to set up a complicated system on day one. First, have it think from the needed point of view using the same documents.

For finance, ask like this:

Separate sales, variable costs and fixed costs, and calculate monthly profit.

Keep unpaid tuition separate from money already received.

Show the missing numbers and what to check next.

For sales, like this:

Summarize what they want, what we promised, unanswered questions and the next contact date.

Don't write anything they haven't decided as a decision.

If even one of these works for real work, add the next.

At first, I did it the other way around. I added roles first. By the night, there were seven.

8. Numbers show what "school first" means

From here, this is a calculation example, not real results from back then. All numbers are made up for the calculation, not my school's actual tuition or costs. The currency is Nepali rupees (NPR).

Suppose:

• Monthly tuition per student: 5,000 NPR

• Monthly variable cost per student: 500 NPR

• Monthly fixed cost: 90,000 NPR

The amount each student contributes toward fixed costs is

5,000 − 500 = 4,500 NPR.

The break-even point is

90,000 ÷ 4,500 = 20 students.

• 10 students: sales 50,000／variable costs 5,000／fixed costs 90,000／monthly profit −45,000

• 20 students: sales 100,000／variable costs 10,000／fixed costs 90,000／monthly profit 0

• 30 students: sales 150,000／variable costs 15,000／fixed costs 90,000／monthly profit +45,000

This example doesn't include taxes, loan interest, initial investment, or my own pay not counted in fixed costs. If more students mean more classrooms or teachers, it has to be recalculated.

And even if the calculation shows profit, you can't pay rent if tuition doesn't come in. So the numbers to watch aren't just the number of students.

Enrolled students, amount billed, amount received, amount unpaid, next payment date.

Before building a new app, make these five visible. That's what I needed back then.

9. Change what you give AI depending on the school's stage

Phase 1｜0–20 students

First, gather students and get to a state where classes can continue. Track inquiries, trial lessons, enrollment and tuition payments.

Have AI help with organizing information, calculating income and expenses, and organizing next actions. At this stage, replying to today's inquiries matters more than an impressive org chart.

Phase 2｜20–100 students

As numbers grow, you can't manage it in your head. Attendance, learning progress, interview prep, teachers' work, contact with companies.

Keep the procedures you've tested on the ground so the same work can be repeated. Have AI tidy those records into manuals and weekly reports.

Phase 3｜100+ students

Make clear who decides what and where it's checked. You need people who check AI's suggestions and people who carry them out on the ground. Because having a role inside your computer is different from having a real person responsible.

"No one to do this with me" in the June 3 diary wasn't solved by adding more role files.

10. If you start today, start in this order

If you're like me and don't know where to begin, I think this is a good way to go. This isn't the order I followed on June 3 — it's the order I think I should have followed, looking back.

1. Pick one thing that's troubling you in your current work.

2. Write the company's purpose, numbers, projects and decision rules.

3. Give it one of today's records.

4. Have AI give you three next actions.

5. Do one, and record the result.

The first instruction for that can be just this:

Below, I'll write the company's purpose, current numbers, ongoing projects and today's events.

First, organize the information.

Next, choose the one problem to prioritize now.

Finally, suggest up to three things to do tomorrow, each with a deadline and a completion condition.

Ask me about anything unclear. Don't invent results or decisions I haven't written.

【Company purpose】

【Current numbers】

【Ongoing projects】

【Today's events】

Do this once, and move one task forward. That moves a company further than making seven roles on day one and feeling satisfied.

What I want to value in this company

A Japanese school, recruitment, support. Just adding each as a separate idea won't make the company I want to build.

From learning Japanese, to preparing to work in Japan, to connecting to a job.

I want to build a road between Japan and Nepal that people can keep using.

I want to use AI for that. Organize numbers. Keep promises on record. Find gaps in the work. Decide what to do next.

But you need people to teach, people to talk with companies, and people who take responsibility and decide. I can't run away from that work either.

If I could go back to myself on June 3, I think I'd say:

"You don't need to think up a bunch of apps."

"It's not A and B. Pick one."

"First, make the school location you decided today real."

"Get the estimates."

"Gather students."

"Manage the payments."

"Use AI to move those forward."

I still need to keep telling myself this. Because I always want to build something new right away.

June 3: the company wasn't finished

There's no record of that day's AI system producing any sales. The school wasn't finished. I didn't find partners to run it with.

But I had decided the school's location. Inside my computer, the boxes of a company existed. And the last diary entry said:

"There's no one to do this with me."

What I want to leave isn't only the screens of success. What I didn't understand. The order I built things in. Where I got stuck. What was missing to use it for real work. That process.

June 3. The day I copied a video's template and started giving AI the roles of a company. In real life, the day I decided the school's location. And the day the distance between building the shape of a company and actually running one started to come into view.

This is where it began.

Next time: June 4. What did I try to do next, after building an empty box? I'll follow the records that remain.