An assembly step is numbers

Take the side panel of our demo bookcase. It is 2 cm thick, 202 cm tall and 28 cm deep, with an unfinished edge that faces the back. The dowels that join it to the shelves are 8 mm across and 34 mm long, and there are 14 of them. The cam locks are 15 mm across. The shelf pins are 5 by 16 mm, and you need 16.

A blueprint stores exactly that. Each part has its size, the place it sits in the finished build and the place it lies on the floor before you start. Each step lists what moves, which fastener goes into which part along which axis, with which tool, and where the camera looks. The player draws the scene from those numbers, so the panel is always 2 cm thick and the hole is where the manual put it.

A generated video is pixels. Nothing in it says the dowel is 8 mm, and nothing checks that a hole it shows exists. It can look right and still be wrong in a way you only find at step 11.

A script checks every file

Before a manual goes in the library, a validator reads the whole file. Every part and tool a step mentions must exist. Every move must start and end inside its step. For each kind of hardware, the pieces used across the steps must add up to what is in the bag: the 24 nails used in step 9 of the bookcase are the 24 nails in the bag, no more and no fewer. Titles and labels must fit on a phone screen.

When we wrote this, all six manuals in the library passed with no warnings. One rule the script cannot check, so we keep it by hand: the words and the animation must agree. If a step's text and its motion disagree, we reorder the steps. We never cover the gap with a tip.

You cannot check a video like that. You can only watch it and hope.

You can stop it, turn it and look inside

  • Scrub and rewind. Drag through a step to the exact moment the cam lock turns, or replay it.
  • Half or double speed. Slow down for the tricky joint, speed through the shelves.
  • Any angle. Turn the model with a finger. A video has the one angle its camera chose.
  • X-ray. Placed panels go see-through, so you can see the dowels inside the joint after the shelf is on.
  • Sizes. The parts drawer lists every piece with its size. Measure a board and you know which board it is.
  • Your pace. Steps play by themselves but never move on by themselves. A video keeps going whether you are ready or not.

The co-pilot uses the same numbers. When you ask "which screw is this?", the answer comes from the manual's hardware list: part number, size and count. It is told never to invent a step, a size or a part number, and to say so when the manual does not.

It costs cents, once

A new manual is compiled once, by Claude, Anthropic's model, from the pages you photograph or the maker's PDF. Our budget is about 9 cents a manual. A closer estimate for a 24-page manual is 19 to 28 cents, because manual pages are mostly drawings and are billed as images. Both are estimates. We will measure the first ten real compiles before we trust either.

After that, the manual is a cached file. The bookcase blueprint is 20,659 bytes, smaller than most photos, and opening it again costs no AI at all. The same file feeds the web player now and the iPhone app next.

Video is priced per second. When we checked in October 2026, Google listed its Veo 3.1 video model at $0.05 to $0.40 per generated second, depending on quality. A five-minute assembly video would cost $15 to $120 before a single retry, and a corrected manual would mean paying for a new video.

What a blueprint cannot do

Parts are simplified shapes at their real size, so a shelf looks like a plain board, not a photo of yours. Rounded edges, cut-out outlines and bendable ribbons for cables help pieces tell themselves apart, but a curved chair back is still rough. A manual compiled from your own scan is AI work and can be wrong, so if a step looks odd, check it against the paper page.

And video still does one thing a blueprint cannot: it shows real hands. That is what the camera is for in BuildMog, pointed the other way, at yours. Here is what the Live co-pilot sees.