1 August 2026

We deleted two of our own layouts, and the two zeroes meant different things

unPaper’s template hands your AI a catalogue of layouts to build with. We have spent most of a week adding to it, because models kept constructing structures by hand that had no name. Counting which layouts get used answers the other question too, the one nobody enjoys asking: which of ours never do.

Two had never been used. Not rarely — never, by any model, in any run. Both left the template. A catalogue that only ever grows is a catalogue nobody has checked.

Only one of those zeroes means what it looks like

The icon-led rows layout had a worked example in the template. Every model that received the template saw a slide built with it, and every model chose something else — cards, a plain list, a header band. That is a real finding about preference. Models were shown this and did not want it.

The org tree had no worked example. It was a name in a list and some styling behind it, nothing more. Its zero is weak evidence about what models prefer and strong evidence of something else entirely: a name without a worked example does not get used. Nobody refused it. Nobody could see it.

Which makes this a correction of ours

We had already written both of these off in the same sentence, as dead furniture. That reading was half wrong, and the half that was wrong is the more useful half. One layout failed on merit. The other failed because we described it instead of demonstrating it — a defect in how we published it, not in the layout.

The same conclusion arrived independently from the prompt-depth experiment, which found that naming a capability buys much less than showing one. Two different measurements, the same lesson, and together they are why the prompt we publish now demonstrates every layout it mentions rather than listing names. If your AI is ignoring something you believe you asked for, the likeliest reason is that it was told and not shown.

What we changed

Both layouts came out of the template, and out of the rules-only version of the prompt — the variant that ships without an example file, which must never advertise something the template no longer contains. Hierarchies are served instead by the connected graph, which carries real PowerPoint connectors that re-route when someone drags a box in the converted file. For a tree of boxes that is a better answer than the fixed one we had.

Retired, not deleted. Both still live in the test corpus as a real deck, so they stay measured on every conversion run — 99.26% and 98.95% fidelity, and either is one paste from returning if the evidence changes. 10 layouts survived the same count and stayed.

How honest this is