How to build a site without programming, step by step

This site was put together without a single line of code written by hand. That is not an advertisement, it is the starting point: if it worked for me, it will work for you. What follows is everything I wish somebody had told me on day one.

It will not say which AI tool we used - we explained why in the other article. Everything here holds for any AI tool that can do two things: read and write files in a folder on your computer, and run commands. An AI tool that only answers in a chat window is not enough. That is the one technical thing you need to know before you start.

What you need before day one

Four things, and three of them are free.

1. A GitHub account. This is where your site will live. Free.
2. A Cloudflare account. It takes what is on GitHub and shows it to the world. Free, and it stays free at this size.
3. An AI tool with access to a folder. This is the only account you pay for.
4. A domain name, when you are ready. A few euros a year. You do not need it on day one - the site works on a free address until you decide.

You do not need a server. You do not need a database. If somebody tells you that you do, they are doing something other than what you are doing.

The first rule, and the most important one

Say what you want to make, not how it is done.

This is the difference between a good day and a bad one. “I want the home page to show the last five news items, newest at the top, and an old one should disappear by itself after a week” is a task. “Write me a loop that filters an array by date” is an attempt to program in a language you do not speak.

You are the editor, not the programmer. The AI tool is the one that knows how.

What to build with

This is the choice we made on the first evening and never once regretted. You do not have to copy it, but here is why it is what it is.

A static generator. The site is not assembled every time somebody opens it, but once, in advance, and after that it is just files. No database, no server for anyone to maintain, nothing to break into. For a zine, a blog, a portfolio or an association’s page, that is the right tool.

Git. It sounds frightening and it only means this: every change is recorded and every one can be undone. When you break something - and you will - that is the difference between five minutes and a lost day. The AI tool will use git on your behalf. You only need to know it exists and that it is protecting you.

An editor in the browser (a CMS). Do not leave this for later. If the site has anybody on it besides you, that person has to be able to publish without you. For us it was a question of survival: my brother writes, and I do not want to be the bottleneck everything has to pass through.

Markdown. Every article is a plain text file. It will open in twenty years, in anything.

The first five evenings

Do not try to do it all at once. In this order, and every evening has an end.

First: have the AI tool set up an empty site and publish it. Nothing on it has to be pretty. The goal for the evening is an address you can open on your phone and show to somebody. That feeling carries the next three weeks.

Second: one article, start to finish. Write it, publish it, look at it on a phone. Now you know the whole circuit.

Third: the layout. Home page, sections, menu. This is where you will say “no, I do not like this” for the first time - and that is work, not time wasted.

Fourth: the editor in the browser, so somebody else can publish.

Fifth: the checks. A tool that reads all your texts before publishing and flags what is wrong. More on that below.

Everything that comes after - photographs, maps, an archive, a second language - is easier than these five.

Where you will lose a day

This is the most useful part of this guide. Every item on it cost us something real.

1. You will believe a beautifully written text. The AI tool sounds exactly as convincing when it is right as when it is wrong. A person who does not know usually sounds like a person who does not know - the AI tool does not. The most beautifully written text in our project was also the most wrong.

2. You will ask for a story and give no data. We got an article about a mountain walk containing two people who were not there, including a dog, and a trail nobody walked. The rule that came out of it: no data sheet, no article. If a number or a name does not exist in what you gave it, the AI tool will fill the gap - not because it is lying, but because that is its job.

3. It told you “the build passed” and you believed it. That is not a check. A build can pass perfectly while the page looks like a car crash - our middle column once repeated itself nine times and the build was green. Open the page and look at it.

4. You will copy a number out of a video or somebody else’s page. We took a claim from one recording and wrote it down as fact. It was wrong, and the entire paragraph that rested on it fell with it. The rule: a claim from somebody else’s source is not a fact but that source’s claim. Either you attribute it in the sentence, or you verify it in the primary source, or you drop it.

5. The editor will shift your data. One afternoon ours wrote 122 wrong values into 47 news items, because our records had fifty different shapes and it lays fields out in order. If you use a CMS with lists, give every record every field, empty ones included. That was not its mistake, it was ours.

6. You translated only one side. If you work in two languages, every change has to go into both in the same step. Leave it for tomorrow and tomorrow you will have two different sites and a link leading nowhere.

7. You took somebody else’s photograph. No. Your own photographs, the official logo of whoever you are writing about, a YouTube thumbnail with a link to that video, Wikimedia with the author’s name and the licence. Everything else belongs to somebody else.

8. The disk filled up. It sounds stupid until it happens in the middle of a job. Twice, for us. If the AI tool reports an odd error for no reason, check whether there is room first.

9. The link looked perfect and led somewhere else. Added 21 September 2026. This is the nastiest item here. We were given five addresses of the form pubmed.ncbi.nlm.nih.gov/9724645/, a proper domain with a number after it, and three of them led to entirely different papers. The claims attached to them were mostly correct. What was invented was the evidence, not the claim. The rule: nobody copies a number in an address from memory, only from an open page.

10. You set a check that can be passed from memory. Our rule read: look at whether anything follows the domain besides a slash. An invented identifier passes that perfectly, so the tool honestly reported having carried the check out. If your check can be satisfied without opening a single page, it checks nothing.

11. You tightened the instructions and made things worse. After one bad batch we added three required fields and five self-check questions. The next batch contained five invented news items out of five. An empty field gets filled. Every new required field is a new place where, when the fact is missing, something that looks like a fact will appear. Write rules that take something away - remove it, bring it down to the level of the source, attribute it to somebody else’s text - rather than rules that ask for more to be written.

Write down the house rules

This is the thing I would do on day one if I could start again.

Make one file of rules in the project and write into it everything you agree on. Not instructions for use, but decisions: whose text is not to be touched, how a dash is written, where a link may point and where it may not, what happens when a piece of data is missing.

Ours has seven sacred rules and every one of them came out of something that first went wrong. The AI tool reads it at the start of every conversation, so it does not make the same mistake twice. You read it when you forget why you decided something, and that will be more often than you think.

How you check

Three things, every time, before publishing:

• A tool that reads all the texts and flags what sticks out: a missing space after a comma, a post without a picture, two identical links, a date in the future. Written once, useful forever.
• A test build with one command. Always the same one. If it is different every time, somebody will eventually skip a step.
• Your own eyes on the finished page. Especially when you have changed the layout.

And one question worth more than all of the above: “where did you get that from?” Ask it every time you see a number you did not type yourself.

How much this costs

The site: one domain name a year. Everything else on the list above is free and stays free at this size.

The AI tool: a subscription. That is the only real cost, and it grows with how much you do, because you are not paying for code, you are paying for conversation. If that is a problem - and it will be - work in shorter, clearer steps. A long discussion in which you do not know what you want is the most expensive thing you can do.

Tokens: where it all goes

This is the part nobody mentions until you hit the wall, so here it is up front.

An AI tool is not billed by the task but by tokens - roughly, by the pieces of text that pass through the conversation, in both directions. That means something counter-intuitive: you are not paying for the result, you are paying for the road to it. Ten minutes of wandering costs more than a well-stated task solved on the first try.

And something more important: a conversation fills up over time. Every question carries with it everything said before. The same task at the end of a long session costs several times what it costs at the start. A good day therefore does not look like one long conversation but like five short ones.

For us this became a real problem, not a theoretical one. Here is what has helped so far.

Know what you want before you open the conversation. This is by far the biggest saving and it costs nothing. One sentence written down in advance replaces ten messages spent working out what we are even doing.

One subject per conversation. When the job changes, close it and open a new one. Dragging the home page layout through a conversation that started as a typo fix is paid for with every message after that.

Whatever repeats - make it stop repeating. Every tool of ours came out of work we first did by hand: the spellcheck pass, the data normaliser, the summary of the watch data, the list of subjects already covered. You write it once and from then on it runs for free, however many times you run it. This is the only saving that grows over time.

Prepare the material outside the conversation. If the AI tool has to hunt for numbers through your files, you pay for the hunting too. Give it a finished data sheet and you pay only for the writing. On a hike that is the difference between half a page and a GPX file with ten thousand points.

Write the rules down instead of repeating them. That rules file from the section above is not only tidiness - it is a saving. Without it every rule gets explained again, every time.

Say no early. “I don’t like this” after two sentences costs a tenth of the same words after a finished article. The most expensive texts in this project were the ones we let run too far before saying no.

Send the easier work elsewhere. Not everything has to be done by the same AI tool. Finding subjects, first drafts and rough translations can go to something simpler and cheaper, and what comes here is what needs the project’s memory and accuracy.

What we have not done yet

There are plenty of ideas still to come: shorter project briefings at the start of a conversation, a tool that prepares its own summary of the day instead of it being retold, a better split between what has to be read and what does not.

You cannot do it all at once. We are building this and learning as we go, mostly from our own mistakes - which is how it usually goes in life as well.

When to stop

Once the site starts working, there will come a moment when you can add things endlessly. Do not.

Our site today is bigger than it needed to be, and that is the one thing I would do differently. Every new section is an obligation somebody has to keep filling. Ask yourself before each one: who is going to be writing this in three months?

If the answer is not the name of somebody who has agreed - do not open it.