A Web production team with zero employees. The story of how I rebuilt my company's website in 3 days with 2 AI employees and ChatGPT

I renewed my company's website in 3 days.

I didn't ask a production company to do it.

The ones I worked with this time were two AI employees running on Claude Code and ChatGPT.

I didn't write any code.

I'm not exactly a master of design tools either.

What I did was

"I want this to be like this,"

"This page is missing,"

"Something feels off here,"

"I want to make this lead to more inquiries,"

and just talk to the AIs.

As a result, in three days, not only did I renew the website, but I was also able to create:

• An environment where I can update the site myself using WordPress
• An owned media site
• Case studies
• Inquiry flow
• Access analysis using GA4
• Conversion tracking
• Weekly and monthly reports
• Data analysis and improvement suggestions by AI

What I felt after trying this was

not just that "you can build a website quickly using AI."

Rather,

it was an experience that made me realize the very way we outsource work might change.

That was the kind of experience it was.

To begin with, why did I want to renew the website?

Until now, the I.K.E. Planning website was basically produced by outsourcing.

Outsourcing itself isn't bad.

However, small problems arise during operation.

"I want to change this text,"

"I want to add a new case study,"

"I want to fix this because the service details have changed,"

even if I think these things, I have to make a request every single time.

Then, it inevitably becomes "Well, I'll do it later since it's just this much.".

Because of that, minor corrections were neglected.

And there was an even bigger problem. The fact that there had been almost zero response from the website until now.

The company has a website.

But it's barely being used for marketing.

So, I decided that this time, I wouldn't just do a renewal that only changes the look.

The 5 things I wanted to achieve this time

First, I decided on five objectives.

1. Enable in-house website updates at any timeCreate a state where we can make improvements immediately upon noticing them, without relying on outsourcing.

2. Launch an owned media platformContinuously update columns and disseminate information with SEO and AEO in mind. This will also contribute to the branding of I.K.E. Planning.

3. Use the website for marketingEstablish a state where we can properly implement measures to increase inquiries from the website, which had been almost zero until now.

4. Analyze GA4 data and make improvementsImprove the website while looking at actual access data, rather than just relying on a 'gut feeling' that things seem okay.

5. Automate everything up to GA4 reports and insightsAcquire data on a weekly and monthly basis, and automatically generate not just the numbers, but also insights and the next improvement proposals.

In other words, what I wanted to create this time was not a new website. A system that allows us to keep growing the website ourselves instead.

So, let's try doing it with just AI.

That's when I thought,

'Can't I do this with AI now?'

I use AI for work on a daily basis.

I have also connected Discord to Claude Code, and it works for me like an AI employee.

In that case, this time, let's try having an AI team handle the website production too. That's what I thought.

However, I didn't have one AI do everything.

What I was very conscious of this time was division of roles.

Me

President and ProducerDecide what to create and what the company should communicate. Look at the finished product and decide whether to adopt it or revise it.

ChatGPT

Advisor and Web DirectorBrainstorming with me, organizing site structure, reviewing copy and text, providing third-party perspective reviews, and creating production instruction documents to pass to Claude Code.

Claude Code + 2 AI Employees

Production, Implementation, and Operations TeamResponsible for hands-on work such as design, coding, WordPress implementation, bug fixes, content creation, GA4 settings, and post-launch analysis.

This is the kind of team it is.

I didn't give perfect instructions from the start.

This is the part that was interesting when I actually tried it.

I wasn't writing incredibly precise prompts for the AI.

Rather, the actual instructions were more like,

'There is no customer support/BPO page.'

'Two case studies are fine for the top page. Let's send the others to a separate page.'

'Let's include columns too.'

It's quite a normal conversation.

Of course, not everything can be implemented just like that.

For example,

"The entrance to the service on the top page feels a bit weak."

When I feel this way,

it cannot be solved by saying,

"Change the CSS of such-and-such."

In the first place, what is weak? I need to think from there.

That's why I made ChatGPT my "Web Director."

This is when I used ChatGPT.

I show it the website under construction and bounce ideas off it, asking,

"What do you think?"

"Doesn't something feel off here?"

Then, it organizes the vague discomfort I was feeling, saying things like,

"I.K.E.'s philosophy and the strength of 'division of labor' are being conveyed. On the other hand, for a first-time user, the gateway to 'what kind of company is this specifically' is weak."

So, I talk further, asking,

"Then, how should I fix it?"

Ultimately, I decided on a flow of: "This is about us" ↓ "I can rely on I.K.E." ↓ "Let's look at the services in detail"

to guide the user.

Once that was decided, I asked ChatGPT, "Make this into a production instruction document for Claude Code."

Then, it becomes an instruction document organized by:

Objective,

Sections to add,

Text to post,

Layout,

CTA,

Smartphone display,

Points to note during implementation.

I pass that to the AI employees on Discord.

The flow is:

Me → Bouncing ideas off ChatGPT → Production instruction document → Discord → AI employees implement it This worked quite well.

Turning "something feels off" into work

When a business owner works with a production company, I think what is surprisingly difficult is conveying what is in one's head to the production side.

"I want to bring out more of the company's character."

"It feels somewhat weak."

"I want to lead to more inquiries."

I understand these as sensations.

But turning that into web production specifications is a different job.

This is exactly what ChatGPT took on for me this time.

For example, even for a single career entry in a company profile,

whether to write "Debuted as a marriage matchmaker in 2024"

or "Inherited and started operating a marriage agency business in 2024."

Even if the facts are the same, the impression conveyed as a company representative's profile is different.

Building a website is, in the end, the work of putting into words "who we are." I used AI quite a bit for that as well.

On the Claude Code side, this is what was actually done.

Meanwhile, I had Claude Code and my AI employees handle the actual production.

For the site design,

Top Page

2 types of services

Company Profile

Case Studies

Recruitment

Contact

Columns

and all other pages were created.

I designed the layouts for both the PC and smartphone versions.

I implemented that design as a WordPress theme and handled the coding.

I also proceeded with the necessary tasks for launch, such as integrating Contact Form 7 and setting up the menus.

As for the content,

FC Headquarters

Regional Media

Healthcare

DM Company

EC/Physical Stores

and other actual client cases were used as a base to create 5 case studies.

I also created service pages for sales support and customer support/BPO.

I prepared columns for our owned media as well.

Of course, it wasn't finished in one go.

Writing "made in 3 days with AI"

might make it sound like it was finished after entering a prompt once.

That was absolutely not the case (laughs).

I fixed it many times.

The display was broken when viewed on a smartphone.

Fix it.

The image was cut off.

Fix it.

The logo was stretched.

Fix it.

The page URL settings were wrong.

Fix it.

Looking at the top page,

"This flow isn't right after all"

and I changed the structure itself.

After creating and inserting an image,

"The size feels off"

and I remade it.

Create → View → Find something that feels off → Fix it. I repeated this many times.

However, what I felt was important this time was **not that "I didn't have to make corrections," but that "the cost of making corrections was extremely low."**With outsourcing, I would think,

"I shouldn't contact them for something this small,"

"I hate to ask for another revision,"

"I wonder if this will incur additional costs."

With AI, I can say,

"That's not it. One more time."

I can say it as many times as I want.

I can do it at night.

I can try it the moment I think of it.

Therefore, as a result, the number of improvements themselves increased.

Finally, "full page proofreading before publication" with ChatGPT

Once the site was almost complete, I converted all pages into PDF format.

I handed that to ChatGPT and requested, “Please check everything as a final proofread before going live.” I asked.

It wasn't just a simple check for typos and omissions.

I checked the volume of text, headings, priority of information, clarity of services, flow to CTAs, balance with images, how it looks on PC and smartphones, and consistency of notation.

Furthermore, I categorized the corrections into “Must fix,” “Fix if possible,” and “OK as is.”

Then, I turned that into an instruction manual for the creators again.

The AI employees made corrections based on those instructions.

I kept this cycle going until just before the public release.

And the AI employees' work doesn't end when the website is published

This was actually what I personally wanted to do the most. “Don't just build it and be done.” I implemented GA4 and made it possible to measure inquiry completions as conversions.

And furthermore, I automated weekly and monthly GA4 reports.

For example,

session count,

number of users,

page views,

average session duration,

bounce rate,

traffic channels,

and most viewed pages

are delivered automatically.

But what I wanted wasn't just the numbers.

Below that, there is an insight section explaining “what can be understood from these numbers.”

And it even provides countermeasures, saying “So, what should we do next?”

In other words, the work continues seamlessly from the AI employee who builds to the AI employee who operates and analyzes.

This might be why the term “AI employee” felt the most appropriate this time.

It proceeded almost entirely through conversations on Discord

Looking back, I hardly wrote any code for this website production.

Basically, I was just talking on Discord, saying things like

“Change this,”

“Add this page,”

“Isn't this weird?”

When I needed to organize my thoughts, I bounced ideas off ChatGPT.

Once it was organized, I turned it into a production instruction document.

I took that to Discord.

The AI employees implemented it.

I checked the finished product.

If it was wrong, I had them fix it again.

It was just this repetition.

Before I knew it, the way I was working had become quite similar to running things with a human team.

What AI changes might be 'improvement costs' rather than 'production costs'

This was my biggest discovery after trying this out.

AI makes website production cheaper.

It shortens the production period.

Of course, those are significant too.

But I believe what really changes is the cost of improving the website.

Websites until now have tended to be:

Build

Publish

Update occasionally

Redesign a few years later.

With the system I built this time, I can run a cycle of:

Publish

Check GA4

AI analyzes it

Propose improvements

Fix it on the spot

Check the numbers again.

The website is no longer a 'finished product.' **It becomes a sales and marketing tool that improves every week.**This is a huge difference.

I finally understood the meaning of the term 'AI employee'

Recently, I've had more opportunities to hear the term 'AI employee.'

Honestly, at first, I thought it was a bit of an exaggeration.

But after trying it this time, my thinking has changed a bit.

An AI employee might not be about creating one AI to replace a human.

It's about breaking down the work:

This task goes to ChatGPT.

This task goes to Claude Code.

This task goes to AI Employee A.

This task goes to AI Employee B.

And the final decision-making is done by a human.

By doing that, you give each AI a role and design the workflow itself.

In this website production, the team was: Me: Decision-making and direction. ChatGPT: Brainstorming, planning, reviews, and instructions. Claude Code + AI Employees: Design, production, implementation, fixes, operation, and analysis.

My own work changes too.

From 'doing the work myself'

to 'deciding what to do' and 'looking at the finished product and making judgments.'

This is quite similar to how a manager works with their employees.

What I built in 3 days wasn't just a 'website'

This time, I was able to realize almost everything I initially wanted to do:

Modify the company website in-house.

Update the owned media.

Disseminate information with SEO/AEO in mind.

Measure inquiries from the website.

Analyze GA4 data.

Receive weekly and monthly reports and insights.

And then, look at those results to improve the website again.

So, looking back,

what I built in 3 days might not have been a website.

What I built was a system that allows us to think for ourselves, fix things ourselves, and keep improving while looking at the data.

Until now, it was

"I want to change the website, so let's contact a production company."

From now on, it will be "Looking at this week's numbers, let's change this part.".

And all I have to do is tell my AI employees,

"Fix this."

When people talk about website production using AI, it tends to be about

"how many days it took"

or

"how much cheaper it was."

However, that was not the value of AI I felt after actually doing it this time.

Perhaps the greatest value of AI, more than reducing production costs, is reducing the "cost of improvement."

The website has been published.

But for this time specifically,

the day it was published is not the completion date, but the start date. From here on, will inquiries actually come from the website?

What kind of articles will be searched for?

How will SEO and AEO take effect?

Which pages should be improved to lead to inquiries?

I will continue to make improvements with my AI employees while watching the GA4 numbers.

I plan to write about those results on Note as well.

▼ This is the website I created this time. Please take a look.