This site was built by asking, and nothing used a back door

The pages you are reading were not hand-coded. They were written by an assistant working through PageWebPro's own tool interface: create a page, set a brand, push a block tree, take a screenshot at phone width, look at it, fix it.

That sounds like a party trick until you look at the constraint we put on it.

The rule was no private access

Nothing here used an internal endpoint, a database write, or a developer-only tool. Everything went through /mcp/editor, which is the same surface any customer's agent token reaches, with the same permission checks, the same module gates and the same validation the visual editor enforces.

That matters because it makes the demonstration honest. If we had reached past the product to build the product's own marketing site, the site would prove nothing except that we have root.

What it actually felt like

Slower than you would think, in a useful way. The interesting part is not generating a page. It is the loop: push, render, look, and notice that the picture is too small or that the second column collapses badly at 390 pixels.

Five to eight rounds a page is normal. Nobody skips the looking.

The parts that were awkward

A published page serves a static snapshot, and writing new blocks does not rewrite it, so for an hour we were staring at correct content that was not being served. A content type's base path silently shadows a page with the same slug, which is why the index here is at /updates and the articles live under /dispatch.

Both of those are now written down. That is the other half of building with your own tools: every sharp edge you find is one a customer would have found.