How I am building this blog with AI

I have some experience with websites. For better or worse, I have been involved in the process — sometimes a rather painful one — of creating them for various companies.

The journey will sound familiar.

You hire a designer. You have one or two meetings to explain what you want. They present a proposal. You can request certain changes, within the agreed budget and scope, until the assignment ends.

And then you start living with the website.

Either you keep an external person to look after it, or someone in the organisation has to learn how to modify it: add articles, update information, manage translations, prepare images and wrestle with formatting.

I have used translation tools whose user experience seemed to have been designed in the Middle Ages. And I have seen an apparently insignificant change turn into a small expedition through menus, blocks and settings.

The final design is not always what you imagined, either. It is the result of negotiating between your idea, the designer's interpretation, the budget and what the chosen platform allows: Squarespace, Wix, WordPress or any other.

This can be done. That cannot. This requires an extension. That needs additional development.

Then come the fees: hosting, platform, maintenance, add-ons or services you keep adding. And when you want to change something important, you discover that explaining what you need is not enough. You have to negotiate again with the tool, whoever manages it, or both.

This is not a criticism of designers. Their work has value. It describes an experience I know: you want a website that adapts to your business, and you end up adapting part of the way you work to the website you have managed to build.

But I think this is changing.

With this blog, I wanted to try a different way of working: explain what I want to achieve and use artificial intelligence to build it, maintain it and help it evolve.

It was not about adding a chatbot to a website. It was about organising the work around AI so that I could concentrate on thinking, writing, deciding what deserves to be published and reviewing the result.

And there was a second, equally important idea: that the work needed to solve something once should allow us to automate it afterwards.

The aim is not to remove me from the process.

It is to remove from the process the things that do not need me there.

The recipe: an idea, a conversation and a website

I mainly use ChatGPT as a workspace and a place to coordinate the work. There I discuss ideas, provide texts, ask for reviews and give instructions to change the blog. Behind it all, there is still code, files and a publishing system. The difference is that I do not have to handle every piece personally.

When I finish an article in Spanish, I can ask for English and Norwegian versions to be prepared, preserving the structure and adding it to the appropriate section.

I do not want three different articles about the same subject. I want the same editorial piece in three languages, preserving the ideas, the important nuances and my position.

AI can also suggest where the text fits within the blog's organisation. If I disagree, I change it. Once the rules are established, I do not have to explain from scratch how I want to work with every publication.

That does not mean it remembers everything perfectly or never makes mistakes. It means we can write down decisions and turn them into reusable instructions and procedures.

Everything we solve can become part of the system.

A real example: from article to audio

One of the blog's articles explains how to commission a complex report from ChatGPT: which account we have, which model we choose, how we prepare the materials and how we review the result.

The text was prepared in Spanish, and the English and Norwegian versions were added afterwards. Later, we decided to add the option of listening to it.

That required more than pressing a button and calling the job done.

The text had to be prepared for narration: preventing the voice from reading out web addresses, turning a table into an explanation that could be followed by ear, and preserving the relevant examples and warnings.

The recordings were then generated using the synthetic voices we had chosen for Spanish and English, with a local system on my Mac. Finally, a player was added to the article, with playback, speed and download controls.

Up to this point, we might think we had simply added audio to a page.

But the important part came afterwards.

Solve it once so you can repeat it afterwards

Once the process works, I do not need to solve it again for every article.

We already know how to prepare the text, which voice to use in each language, how to generate the recording and how to add it to the player. That work becomes a recipe the system can execute again.

Then I can say:

“Apply this same process to every article on the blog that does not yet have audio.”

And AI can work through them, prepare the narrations, generate the files and add the players according to the rules we have established. If it encounters a problem, it can flag it for review rather than improvise a solution and carry on as though nothing had happened.

The same applies to translations, formatting or classification.

First we build and test the process. Then we automate its repetition.

We are extending audio gradually, but not because every article requires starting from scratch. A gradual rollout lets us check the recipe before applying it on a larger scale.

That is the difference between asking AI to help me with a task and building a way of working with it: the effort I put into solving something once can spare me having to do it again dozens or hundreds of times.

Execution time and resource consumption do not disappear. What disappears is the need to repeat the same operations personally.

AI prepares; I approve

Automating execution does not mean handing over control.

I can commission the preparation of ten articles, their translations and their audio without supervising every mechanical operation. But that does not require them to be published automatically when the system finishes.

The division of work is simple: AI prepares; I review, correct and decide when it is ready.

I do not want the system to confuse “I have finished” with “I can make it public”.

I can accept a proposal, reject it or ask it to keep an earlier version. I can also disagree with its editorial judgement: a sentence can be more elegant while conveying what I mean less accurately.

My role is not to rubber-stamp a result because it looks good. It is to decide whether I consider it mine and am willing to publish it.

The machine produces and organises. Editorial responsibility remains mine.

I can also change the tool while using it

The most interesting part is not just publishing articles. It is being able to change the website itself through the same process.

I can say:

“I want this section to be called Learn to use AI.”

“The practical guides should appear first here.”

“Do not suggest articles from other sections in this area.”

“This introduction does not explain clearly what I want to build.”

Based on those instructions, AI can prepare the necessary changes. I see the result and continue making corrections.

The conversation starts with the intention, not with the menu where I am supposed to find an option.

And when a decision needs to be maintained, we can incorporate it into the procedure: subsequent publications should respect that organisation, use the same format or include particular checks.

We are not just changing a page. We are improving the way we produce the next ones.

This is very similar to the vibe working I discussed in another article: produce an initial solution, test it, keep what works and change what does not.

I do not need to imagine a perfect website before starting. I can discover how I want it to work while using it.

It is not magic, and it does not always work first time

It would be dishonest to describe this experience as though we had eliminated every difficulty.

We have had to correct texts, review translations, adjust elements that did not look right on a phone and resolve publishing permission problems. Sometimes a change was ready but had not yet reached the public website.

The technical work does not disappear. What changes is who performs it and how I direct it.

I have also found that an apparently clear instruction can be interpreted differently. And that it is worth checking the published result rather than settling for the system saying it is done.

That is why it makes sense to turn the errors we detect into checks for the next run. If a translation is incomplete or an audio recording no longer matches the updated text, the process should detect it before publication.

The aim is not to keep correcting the same thing forever through conversation. It is to use each correction to make the system work better next time.

What matters is not that the website “has AI”

It is worth distinguishing two things.

Artificial intelligence helps me build and maintain the blog. That does not mean a model has to run every time someone opens an article.

The novelty lies mainly in the way we work.

That is also why I do not think the question is about choosing between ChatGPT and a traditional platform. What interests me is finding out how much work I can delegate and which decisions I should retain, using the tools that best handle each part.

The same logic can apply to a presentation, a report, a knowledge library or a small management system.

Not adding AI as decoration, but asking how we would organise the process if we could count on its production capacity from the outset.

And how to retain what we have learned so that we do not have to start again with every assignment.

Digital DIY, one thing at a time

That is why I want to share these experiences in “Learn to use AI”: to show what I have tried, what I have achieved, where I have gone wrong and what is needed to take the next step.

First, publish an article. Then work in several languages. Later, add audio. Then improve the organisation or add a genuinely useful feature.

One thing at a time. But everything that works remains available for what comes next.

In my case, the most important change is simple: I can devote more attention to the content and how I want the publication to work, and less to performing every operation needed to achieve it.

I write, review and decide.

AI helps me build the process and repeat what we have already learned to do.

The technology is still there. It has simply stopped being the centre of my attention.

And perhaps that is one of the best ways to know it is working.