From Dismissing AI Agents to Handing Over My Code and Website: It Took Just One Month

I registered this domain more than a year ago.

At the time, I just wanted somewhere to organize my notes. But I never really ran the site, and it contained only a few scattered drafts.

Then, on August 17 this year, things started to change.

It started with a game of Snake

That day, I watched a bilibili video suggesting that Godot might be a good engine for making games with AI assistance.

Around the same time, OpenAI gave me a month of ChatGPT Plus, which included access to Codex.

So I thought: why not give it a try?

I asked Codex to make a Snake game in Godot.

And it actually did.

At that moment, I immediately realized something:

This could work.

Of course, I also quickly discovered Codex’s limitations at the time, so I went on to pay for tools including Cursor and Claude Code.

Cursor was soon out of the running.

The tools that ultimately stayed were ChatGPT/Codex and Claude Code.

I decided not to write any Godot code myself

Next, I started making my own game in Godot in earnest.

But from the beginning, I set myself a rather strange rule:

I would not write any Godot code.

I went further: I would not even operate the Godot editor myself.

I wanted to know how far I could get by handing all of that to AI agents.

A little over a month later, the game’s prototype was complete.

And as ChatGPT continued helping me plan, I started entertaining an idea I would never previously have considered seriously:

Could I build an AI-driven game company?

Looking back, the change feels a little absurd.

Because perhaps less than a year earlier, I had been quite dismissive of AI agents.

When I saw people “raising lobsters” or running Qwen models locally, my reaction was simple:

Isn’t this just a waste of electricity?

At least, that was what I thought then.

What changed my mind was that the models finally became capable enough

Gradually, I realized that the question might not be whether AI agents were useful, but whether their models were strong enough to complete real work.

Once Claude Code was using Opus 5 or later models, and Codex was using GPT-6 or later, the situation was completely different.

They were incredibly useful.

A few days ago, I happened to see David Heinemeier Hansson—usually known as DHH—talking about how he no longer writes code himself.

That really resonated with me.

DHH is the creator of Ruby on Rails and the co-founder and CTO of 37signals.

I thought such a major change in someone with his background was well worth exploring.

So I immediately went looking for his interviews.

Gemini, DHH, and a real purpose for this website

This time, I used Gemini to process YouTube interviews, turn them into transcripts, and translate them into Chinese.

I felt that this was exactly the kind of work Gemini was well suited to.

Once I had organized them, however, I ran into another problem:

The interviews were long, and I did not have that much time to sit at a computer reading them.

Wouldn’t it be great if text-to-speech could read the articles to me while I was driving?

At that point, I finally remembered this website, which had been sitting around for over a year without a real purpose.

Since it was intended for notes anyway, I could put the things I truly wanted to save, read, and listen to repeatedly here.

Initially, ChatGPT helped me plan and build a WordPress plugin with a table of contents, or TOC, and a Reader feature for reading articles aloud.

The website finally had a very concrete purpose.

A failed attempt at local text-to-speech

The first version of the plugin used the browser’s built-in speech synthesis.

In practice, I was not happy with it.

So I considered another approach:

If the browser’s voice did not sound good, I could use a local TTS model to generate audio for the entire article in advance. The website would simply play that audio.

ChatGPT helped me plan an architecture using the WordPress REST API, then handed it to Codex to implement.

Technically, the approach worked.

But I abandoned it in the end.

Generating complete narration for a long article locally was much more expensive than I had expected.

It required a higher-end graphics card, and generation took too long for my actual needs.

So that idea ultimately failed.

Yet the failure unexpectedly produced something more important than TTS.

If Codex can publish articles, why not let it manage the whole website?

While implementing the TTS system, Codex had already become able to publish articles for me through the WordPress REST API.

Then it suddenly occurred to me:

If Codex can publish articles directly for me, why am I still managing all of this myself?

I could hand management of the entire website to Codex.

Just then, GPT-6.1 Sol launched.

When I used GPT-6 Astra before, my biggest problem was not a lack of capability, but how quickly it consumed tokens.

On a Plus subscription, it was difficult to use it freely.

GPT-6.1 Sol improved the situation considerably.

For my own work, I feel its productivity is worth considering alongside Claude Opus 5.5.

That is not a benchmark conclusion—just my experience using it.

The outcome matched my expectations.

Codex took over publishing articles on this website smoothly.

From Google searches in 2019 to AI agents in 2026

I first learned about WordPress in 2019.

I was just starting to learn about creating independent online media.

What did I do when I did not understand something?

Open Google.

Search.

Watch tutorials.

Then slowly figure it out myself.

I could not have imagined that, just a few years later in 2026, so much of running an independent online publication could be handed directly to AI agents.

Finding and organizing information, translating, modifying code, developing WordPress plugins, planning website features, preparing articles, and finally even publishing them could all be done by agents.

The pace of this change is almost hard to believe.

Why save transcripts rather than AI summaries?

Since DHH had such a major influence on the direction of this website, I felt it was important to preserve his current discussions of AI agents here as transcripts.

These posts are not primarily about attracting traffic.

They are first and foremost for me.

I want to return to them later, or have the website read them aloud while I am driving.

Some people might ask:

With AI summaries so convenient now, why keep such long transcripts?

The reason is simple.

Summaries leave things out.

And what gets left out is not necessarily without value to me.

If DHH’s AI agent interviews were edited in the usual mainstream-media style, the result might contain little more than:

He no longer writes code himself, which agents he uses, and how he sees the future of programming.

Those things matter, of course.

But an interview might also include his comments on a laptop brand, his work environment, which network KVM device he uses, or something else not directly related to AI agents.

For a news summary, all of that can be cut.

For me, that is not necessarily true.

Any small detail might lead me to another topic worth studying in depth.

Transcripts also make it easier to see someone’s way of thinking, habits, and preferences.

Those are often the first things sacrificed in a condensed summary.

Preparing transcripts used to be difficult; now it is much easier

Preparing transcripts of long interviews used to take considerable time and effort.

With help from Gemini and ChatGPT, it has become much easier.

Of course, I am talking about my own use case.

If you really want to use AI as a production tool, I think a paid subscription is still necessary.

At least for my current workflow, the productivity available from free versions is very limited.

Gemini is a somewhat unusual example.

I do not rate its everyday reliability particularly highly, but Google owns YouTube, giving it a natural advantage when processing YouTube videos and retrieving their content.

I also received a free period of Google AI Pro membership with my Samsung phone, so of course I make good use of it.

And sometimes Gemini does suggest solutions to unusual problems that other models have not considered.

For example, while making my game, I ran into a performance problem caused by collisions between large numbers of objects.

Gemini ultimately suggested an approach that solved it very well.

So my attitude toward AI tools is now quite simple:

AI is a tool. If it is useful, put it to good use.

There is no need to believe in one company or choose sides over tools.

Use whatever solves the problem.

This article itself is a product of that way of working

There is one more thing I find interesting.

The journal entry you are reading was itself created using the process I just described.

I did not sit at a computer and write it from beginning to end in one sitting.

Instead, I wrote it in fragments in ChatGPT.

Sometimes I thought of something on my phone and added a little.

Sometimes I sat down at the computer and added another paragraph.

I could even use abbreviations freely and not worry too much about typos.

ChatGPT then reorganized the fragments, supplied the necessary full names, corrected typos, and applied minimal polishing to turn them into a complete article.

Next, it generated a publishing work order.

Finally, Codex published the article to WordPress.

So if an article on this website lists “Codex” as its author, it means:

Codex carried out the publishing.

It does not mean that Codex invented the entire piece on its own.

For this article, the ideas, experiences, and opinions were all written down by me, bit by bit.

AI organized those fragments and completed the tasks I no longer particularly wanted to do myself.

In fact, it has only been a month

Looking back, what interests me most may not be which model is stronger or which agent is the easiest to use.

It is the time involved.

It was only a little over a month ago that I started seriously exploring AI agents. Since then, I have fully embraced this way of working and even decided that, most of the time, I would no longer write code myself.

A little over a month ago, I just wanted to see whether Codex could make a Snake game.

A little over a month later, I was redesigning game development, website management, and much of my daily work as processes that AI agents could participate in or carry out directly.

What will things look like another year from now?

I honestly would not dare predict that today.