Why text is what the AI can use

Why the viewer only shows text. The viewer — the middle zone of VS Code — is, underneath, Notepad on steroids: it shows characters.

Open a .md, a .csv or an .html and you see the contents. Open a .pdf and you see %PDF-1.7 %âãÏÓ, or simply a notice that it cannot display it.

With images there is a convenient exception: VS Code ships with a viewer for .png, .jpg, .gif and .svg. For PDFs you have to install an extension — the VS Code kind, not the file-name kind.

But the extension lets you look, not work.

With the PDF viewer installed, you can read the file. That changes nothing about what the AI can do with it, does not make it editable, and does not make it comparable between versions. It is still a block.

Install the extension if it helps you review at a glance. But if you are going to work on that content, what follows is the important part.

The AI reads text. When you hand it a binary, somebody has to extract the text first — and in that extraction part of the structure is lost and noise creeps in: columns that merge, headers that repeat, tables turned into a tangle.

And even if the extraction comes out perfect, the underlying problem remains: on a binary it cannot make a surgical change. It can redo the whole thing, which is not the same — every pass risks touching things that were fine.

The comparison that makes it clear: asking it to fix one word in a text file is like asking for it in an open Word document. Asking for it on a PDF is like asking for it on a screenshot of that Word document. It can read the screenshot, yes. But to change the word it has to draw the screenshot again.

On text, by contrast:

• It edits one line and the rest stays intact, byte for byte.
• You can see what changed, line by line, before accepting it.
• You can go back to an earlier version without losing the rest.
• It weighs almost nothing, so many files at once fit in one conversation.

And the same applies to visual work. A banner is also built in text — an .html with its .css — and exported to an image only at the end.

Changing a word or a color costs two seconds. If you start by drawing the image, every change costs the same as the first one.