Compare
unPaper vs asking your AI directly
The alternative almost everyone tries first
Before any tool, there is the obvious move: open ChatGPT or Claude and ask for a deck. It costs nothing extra, nothing leaves the assistant you already trust, and modern models will happily produce slide HTML. For a short, simple deck this often works, and if it works for you, you should keep doing it.
What we kept running into is what happens past that point, and we measured it rather than argued about it. Asked freely, a model writes a deck it invents from scratch each time: the layouts vary between slides, structure collapses toward bullet lists as the deck gets longer, the geometry is not fixed so printing drifts, and the HTML usually cannot be turned into an editable PowerPoint because it was never built to be. Ask twice and you get two different design systems.
unPaper is that same request with a contract attached: a named catalogue of layouts to choose from, budgets for words and bullets, a fixed 1280×720 geometry, and a worked example. Same assistant, same chat, same privacy — a longer first message. The measurements are published, including the parts that did not go our way.
Where they differ
| Feature | unPaper | asking your AI directly |
|---|---|---|
| Where your content goesIdentical, and this is the point — unPaper adds no recipient. | Your AI, and nowhere else | Your AI, and nowhere else |
| Cost | Free; you already pay for the assistant | Free; you already pay for the assistant |
| Layout variety across a deck | A named catalogue the model picks from, used in four or five briefs out of five | Invented per deck; drifts toward bullet lists as length grows |
| Consistency between two runs | The same layout language every time | A new design system each time you ask |
| Prints to an exact PDF | Fixed 1280×720 geometry with a page rule; prints pixel-exact | Usually reflows — page breaks land mid-slide |
| Converts to editable PowerPoint | Yes — the format was designed backwards from the converter | Rarely, and not reliably |
| Works with a cheap or restricted model | Yes — choosing from named layouts is easier than inventing one | Weaker models degrade fastest without a structure to lean on |
| First message lengthThe honest cost. A worked example is most of the payload. | Long — a rules block plus a worked example | One sentence |
| Revising one slide | Comment on the slide, splice the replacement in locally | Regenerate and re-paste, or edit HTML by hand |
| Needs anything installed | No | No |
Behaviour described here is measured, not assumed — the runs are published in the Lab. Last reviewed 17 August 2026.
Stick with asking your AI directly if…
- The deck is short, informal, and only you will see it.
- You want a rough structure to think with, not a document to send.
- You are happy in raw HTML and would rather shape it yourself.
- You do not need it to print exactly or become a PowerPoint.
Use unPaper if…
- The deck goes to someone who will judge it, and you make this kind of deck often.
- It has to print to an exact PDF, or arrive as a real editable PowerPoint.
- The model available to you is not the strongest one on the market.
- You want the same visual language across a deck and across months.
Check it yourself
The rows above that describe behaviour rather than architecture come from published runs.
Or read the common questions first.