All writing

Essay

Web pages changed how I think.

A document hides its structure. A page has to declare it. Once the format you think in is the same one a model can read and rewrite, the page stops being where you publish an idea and becomes where you have it.

David Galvis 2026 8 minute read

The shape of a thought

<h1> the claim
<h2> the part that carries it
<p>  the evidence
<aside> the thing you are not sure about yet _

For most of my career the thinking happened inside tools that were designed to present it. Slides, documents, a shared folder. You open a deck to work out an argument and within ten minutes you are choosing a font. The medium asks you for a look before it asks you for an idea.

That order is backwards, and I did not notice it for years. What I noticed was the symptom: I would finish a deck, know that something in it was wrong, and have no way to point at it. The structure was invisible. It was implied by position on a slide, by size, by what came after what. Nowhere in the file was there a line that said this is the claim and this is the evidence for it.

The tool was deciding the thinking

A slide holds about forty words. So the argument gets cut to fit forty words, and the parts that did not fit get dropped rather than reconsidered. A document scrolls forever, so everything survives, including the parts that should have died. Both of them shape the thinking before you have finished having it.

I ran into this hardest at Digitech. Marketing research, account intelligence, a case for a renewal. Real material, with sources, contradictions, numbers that came from three different places and did not agree. Every format I had wanted me to flatten that into a linear sequence of pretty rectangles.

The format you think in is the ceiling on what you can think.

Miro, Covid, and the first canvas

The first time I felt a format helping instead of fighting was Miro, at university. My degree ran straight through Covid, so classes were a screen, and my notes were screenshots. Hundreds of them. I put them on an infinite canvas and started arranging.

What that gave me was a spatial hierarchy. Related things sat near each other. A cluster was a topic. Distance meant something. Nothing about the content was better, and everything about my access to it was, because for the first time the shape of the notes matched the shape of the subject.

The limit of a canvas is that it only exists inside the tool that made it. You cannot query it, version it, hand it to anything else, or ask it a question. It is a beautiful dead end. I kept the habit and lost the artifact.

A page cannot hide its structure

HTML is the format that forces the declaration. There is no way to write a page without saying, explicitly, what is a heading and what is a paragraph, what is a list, what is an aside, what is a table and what its columns mean. The hierarchy is not implied by how big you made the text. It is written down.

That constraint is the whole benefit. When the structure has to be declared, the moment you try to write the page you discover which parts of your thinking do not have a shape yet.

What is the claim. If it does not fit in one heading, there is more than one claim and they are fighting.

What holds it up. Evidence sits under the claim it supports, or it belongs somewhere else.

What is an aside. The interesting thing that is not load bearing. Naming it as an aside is how you stop it from taking over.

What is still open. A section you cannot fill is a question you have not answered, made visible instead of skipped.

None of that is a web skill. It is outlining, with a format that refuses to let you skip the outline.

The models read this natively

Then the second half arrived. The models are extraordinarily good at HTML. It is one of the most common things they were trained on, so it is the format where they need the least explanation and make the fewest mistakes. You describe a structure in a sentence and it exists. You say the second section should be a comparison table and it becomes one.

Styling stopped being a cost too. With the current generation of models, the distance between a plain page and a designed one is a paragraph of instruction. That matters more than it sounds: it means design stops being the reason to delay, and the page can look finished on the first pass, which is exactly when someone else has to read it.

So both directions work. I can write structure and the model renders it. The model can write structure and I can read it, because the file is text and the hierarchy is in the file. Nothing is trapped in a proprietary canvas. The same artifact serves the person and the machine.

I stopped writing documents about my work and started writing the surface I work on.

A page plus an archive is a tool

The step that changed the most is the one I did not expect. Once the pages are plain files, they accumulate. Months of them. Research, meeting notes, decisions, the reasoning behind a proposal that got rejected and why.

Point a model at that folder and you no longer have notes. You have a knowledge base, and the pages you generate from it are not summaries. They are instruments: a brief that already knows what you decided in March, a comparison that pulls its own numbers, a map of a topic with every source still attached.

The reference point most people have for this is a hosted notebook that ingests your documents and answers questions about them. This is that, except you decide what it looks like, what it emphasizes, what it refuses to show, and where it lives. Mind maps that render the relationships you care about. A daily brief in the shape of your day rather than somebody's product decision.

You can absolutely build the same thing with APIs and integrations, and at a certain scale you should. I am describing a different path, one that is faster to start and that I can keep entirely in my head. A folder of pages and an agent that reads it.

What changed in practice

I draft in structure now. The headings come first and they are the argument. If the headings do not read as a case on their own, the piece is not ready.

Wrong is visible earlier. A gap in a page is a gap you can see, so the flaw surfaces on day one instead of in the meeting.

Nothing is trapped. Every artifact is a text file that any model, any browser and any future version of me can open.

The archive compounds. Every page makes the next page cheaper, which is not true of a single deck I have ever made.

I am careful about one thing. The machine is very good at producing structure that looks correct, and a well formed page is persuasive whether or not the thinking under it holds. The judgement about what matters is still the part I have to bring, and it is the only part that was ever mine.

Where to start

Three ways to try this on something real.

Take the next thing you would have made as a deck and make it a page.

Write the headings before anything else, and read them alone, top to bottom. If they do not hold as an argument by themselves, the deck would have hidden that from you.

Give the model the structure, not the sentences.

Describe the sections and what each one is for. Let it draft the prose. Correcting a structure you chose is a much shorter job than fixing prose you did not.

Keep the files.

The single page is useful. The folder is what turns into a tool six months later. The value is in the accumulation, and it costs nothing to start accumulating today.

Why this matters to me

I came to this from the creative side, where the shape of the thing was always the point, and then spent years in marketing and sales, where the shape of the thing kept getting in the way of the argument. Working in pages is where those two finally agreed with each other.

Back to home

Keep reading