A single prompt wrote this post. It did not make the decisions.

This is the first piece on the site, and it opens with a confession: what you are reading was not typed by a human. A model wrote it, Claude Opus 5, from a single request, in a single pass. No outline, no second round, no "rewrite that bit".

We are saying so up front, because without it half of what follows would be easy to misread. The interesting part is not that a model can write. Nobody calls that news in 2026. The interesting part is why it came out finished in one pass, and why the page did not end up more generic or less human for it.

What it took

The prompt was not the clever bit. The codebase is unusually talkative.

On this site, every number has the reason next to it. That the headline's line height is 1.06, because below 1.029 the capital Ő physically touches the "gy" on the line above. That brand blue is a separate colour when it is text, because on a light background the original measures 2.14:1 and the standard asks for 4.5. That at the bottom of the home page the opening scene does not end but settles under the footer, and that this seam has been measured across 64 combinations of viewport, theme and page.

These are not explanatory notes beside the code. They are the decisions themselves, kept where they apply. When the request arrived to build the first post, none of it had to be invented. It had to be read.

If a value cannot be justified, it is not a decision yet.

That line was written long before today, and it is exactly what made a model useful on this job. An AI does not become a good colleague by knowing a lot about the world. It becomes one by knowing a lot about you. What sits between the two is written-down decisions.

How a post is built

If this is the first post, it is also the template. The structure is not an aesthetic question but a reading one, and every piece after this one gets the same:

• Text on the left, orientation on the right. The article runs in a single column, and the table of contents and the share options stay next to it the whole way. Whoever gets lost in the sixth paragraph should not have to guess their position from the length of the scrollbar.
• The header says when and about what. Date, category, reading time. The date is absolute, never "last week". The category is one of our four services, not a separate blog taxonomy: two category systems living side by side are the first thing to drift apart.
• Section titles are anchors. Every section can be linked to on its own, and the table of contents is built from them. This is not only convenience: increasingly it is not a result list that answers the question but an AI, and the source of that answer is a section, not a whole page.
• A pull quote only from a source. A quote goes in if it was actually said or actually written down, and we say where. Invented quotes do not go on the site.
• A cover image only if there is a real one. This piece does not have one, and you can tell. An illustration with nothing to do with the text does not decorate, it fills.
• At the end there is one thing to do. Not three.

That list was not written for this one post. The second, the tenth and the hundredth get it too, and if it ever changes, they all change at once.

What this means for you

None of this is special to our website. The same difference shows up at every company trying to put AI to work: if the knowledge sits in people's heads, in folders and in email, the model writes generalities. If it is written down, it does your work.

That is why a second brain is not a note-taking fashion but infrastructure. And it is why the choice of model is not what decides the outcome: the same model writes the generic text and the precise one, and the difference is what it was handed.

To be honest in the other direction as well: a model wrote the text, but the decision to publish it was made by a person, and their name is under the title. The sentence in the title is meant literally. What a prompt cannot write is what we take on, what we measure, and what we leave out.