Written for the person using it, not for the person who wrote it. Where something was fixed, this says what it used to do — that is usually the more useful half.
30 August 2026
The first screen now says who unPaper is for
A clearer promise for people using company-approved AI, an honest PowerPoint handoff, and a local-processing claim visitors can test themselves.
The work AI is now the starting point
The entrance now leads with the real constraint: many employees cannot choose another model or slide service. unPaper gives the assistant their workplace already approved a professional slide system instead of asking them to add a second AI.
A strong start, not a claim of perfection
The product now describes its outcome as a strong, editable starting deck. Converting to PowerPoint and finishing the work with human judgment is an intended part of the workflow, not a fallback hidden after generation.
The data boundary is drawn and reproducible
On your device now separates what the chosen AI provider sees from what unPaper processes locally, then provides a harmless sample deck and an offline conversion check that a reviewer can repeat without trusting a badge or slogan.
Present view fits the first time
A newly opened deck could measure hidden slides too early, making a dense agenda look too tall until the viewer switched modes. Present view now fits each slide when it becomes active, and every bundled template carries the regression check.
17 August 2026
Every template, reviewed slide by slide
A full pass over the five deck templates, rendering every layout in every look and fixing what only shows up when you actually look at it.
Diagrams that were quietly out by a few pixels
Arrowheads that did not meet their lines, cycle arrows that floated away from their circles, markers heavier than the shape they marked, and totals that did not equal the sum of their parts. All six were arithmetic, all six rendered without error, and none would have failed a test that checked the page loaded.
The chart legend became a real element
Legends had been rebuilt by hand on each slide that needed one, so they drifted apart and one measure rendered inverted. A repeated pattern is a class: it is now one element, styled once, and the inverted measure is fixed.
A placeholder that two models shipped as a title
One example slide's heading said nothing about what the slide was for, so two different AI models copied the placeholder text straight into their finished decks. The example now states its own purpose.
The waterfall reaches every look
The waterfall layout was styled in some looks and not others, so choosing a different look could silently cost you a layout the prompt had promised. Every element is now checked against every look at build time, and four slides that had lost their footer got it back.
A gate against the whole class of problem
Most of the above rendered rather than failed, which is the hardest kind of defect to catch. There is now a check that fails the build when a template's own markup has no matching style rule in a look, when an example slide overflows its frame, or when a page number goes missing with scripts off.
10 August 2026
Four conversion details that were losing fidelity
Rounded corners, stacking order, picture opacity and first-line position in the PowerPoint converter.
Rounded corners survive conversion
Boxes with rounded corners converted as square ones. They now carry their corner radius into PowerPoint as a real shape property, so it is still adjustable there.
Stacking order is preserved
Elements deliberately layered over one another could come out in the wrong order, which is invisible until something important ends up behind something else.
Partly transparent images stay partly transparent
Picture opacity was dropped during conversion, turning a soft background image into a solid one.
Text starts where it started
The first line of a text box sat slightly higher after conversion than it did in the browser. Small, and visible the moment a converted slide is put beside the original.
8 August 2026
The whole flow works on a phone
Making a deck on a phone meant scrolling past a preview to reach the buttons that finish it. That is now the shape of the product on touch devices.
Actions come to you
While a finished deck exists on the page, its own actions sit in a bar at the bottom of the screen instead of underneath a full-height preview. It carries only that deck's actions — never navigation — and disappears when there is nothing to act on.
The create loop closes on one screen
You can now paste your AI's reply back on the same screen you copied the prompt from, rather than being sent to the tools hub and back.
Saved files go through the share sheet
On phones, a finished PDF or deck is handed to the operating system's share sheet, so it can go straight to Files, Mail or a chat instead of into a downloads folder you then have to find.
Pasting a deck no longer zoomed the page out
Pasting a full-width deck into the page caused the whole layout to shrink to fit it. The preview is now contained.
Building a PDF says so
PDF generation on a phone can take a few seconds with no visible sign of work. It now reports progress.
6 August 2026
A verifiable offline session, an inspectable prompt, and a licence
Three things that are all the same argument: you should be able to check what unPaper does rather than believe it.
Offline session
The header switch now saves unPaper to your device and then tells it to refuse its own network requests entirely, including anonymous page counts. The same switch puts it back. A genuine loss of connection is detected and shown separately, because one prepared browser is not a private deployment and should not pretend to be.
Prompt integrity
Any screen that copies a prompt now offers a report over the exact string on your clipboard: what was assembled locally, plus deterministic scans for URLs, remote resources, invisible characters and instruction overrides, with a SHA-256 fingerprint. It is a report, not a certification, and it says so.
Organization licensing, written down
Individual use is free, including at work. Organizations licence annually when they adopt it together; private deployments on your own infrastructure are separate. There is no usage metering and no phone-home check.
Command line and agent tooling
The command-line tool and MCP server are published, so an agent or a build pipeline can create, check and convert decks without a browser.
3 August 2026
A deck starts with two clicks
Instead of choosing a template and hoping it fits, you say what the deck has to do, and the example set is assembled around that.
Say what it has to do
The create screen asks the outcome — decide, persuade, align, instruct, account for, explain — and builds the prompt from slides that serve it. The matching runs entirely in your browser and works offline; there is no AI in this step, which is the whole point.
See why each slide is there
Every slide in the assembled set can be inspected: why it was chosen, what it cost, and what was considered and declined.
Browse and override
A catalogue drawer shows the whole library for the look you chose, so you can pin a slide in or strike one out, and your choices survive rebuilding the deck.
More example, not less
We had assumed a smaller prompt was a better one and measured the opposite: a generous library plus explicit permission to ignore most of it produced shorter, more varied decks that used more of the layout catalogue. The default example set grew as a result.
A deck remembers what it was for
A finished deck carries what it was built to do, so re-opening it in the finishing tool picks up where you left off.
1 August 2026
The Lab opens, and the prompt stops overreaching
Published measurements, including two that went against us, plus the prompt changes those measurements forced.
The Lab is public
What we measure about how AI systems build slides, and what we changed because of it. Every number is generated from the runs themselves rather than typed, and one of the launch entries is a correction to an earlier claim of ours.
“You may stop early” was heard as “do less”
A sentence added to protect long decks turned into permission to write shorter ones, and the escape hatch it offered was never once used. Measured, then softened.
A slides-only reply is a valid answer
Some assistants cannot attach files and some run out of room. A reply containing only slide blocks is now treated as the efficient outcome it is: the finishing tool wraps it back into a complete deck locally, instead of reporting it as broken.
Fewer slides is good, fewer kinds of slide is not
A rule intended to keep decks short was also flattening their variety. It now distinguishes the two.
Every Lab entry can be listened to
Each entry ships with a narration, synthesized locally and served from unPaper itself — the Lab embeds nothing from anywhere else.
31 July 2026
Eleven new layouts, taken from what models already build
We asked several AI models for structures without telling them how, watched what they reached for, and added the ones they kept inventing — each only after it survived conversion to PowerPoint as real shapes.
Span bars, waterfalls, tiers and plots
Four structures that models built by hand, over and over, because the format had no name for them. Now named, so a model picks one instead of assembling it from boxes.
Node graphs with real connectors
Boxes joined by lines convert into PowerPoint connectors that stay attached when you move the boxes, rather than into loose lines that do not.
Roadmaps, shares of a total, and two journeys on one scale
A roadmap allocates named periods; a span bar measures a continuous scale. Conflating them had been collapsing the shared axis that is the entire point of a journey comparison.
The arc statistic our own rule was suppressing
A rule told models to avoid curves because we believed they converted badly. They do not — they become native PowerPoint arcs — so the rule was wrong and is gone, and a ring statistic exists now because of it.
Dashed lines keep their shape
A dashed stroke in a diagram converted to a solid one. It now carries the dash pattern into PowerPoint.
Two layouts retired
Two elements had never been used by any model in any run, and were moved to a bench rather than deleted. Only one of those zeroes was evidence that nobody wanted it; the other was evidence about how we had described it.
30 July 2026
Documentation, and shapes that survive the round trip
A real documentation tree, and an importer that brings PowerPoint drawings back as editable shapes.
A documentation section
Guides and reference behind the Learn destination, with search and a real page tree, so the tools themselves do not have to carry the explanations.
Shapes cross in both directions
Importing a PowerPoint deck now rebuilds its drawn shapes — including curved decoration and callout tails — as vector outlines, so they convert back to editable PowerPoint geometry instead of arriving as flat pictures.
Your deck, as it was, comes first
The importer produces several versions of an imported deck. The geometry-faithful copy now leads, because the first question anyone has is whether it still looks like their deck.
29 July 2026
Borrow another AI model's visual taste
Different models design differently when left alone. Those native styles are now archived, and the strongest of them can be worn by any other model.
The style archive
Each probed model's own deck, beside the same taste rendered as an unPaper deck, with what actually distinguishes it. Covers are where dialects look alike; content is where they part, so both are shown.
Style capsules in the prompt
A capsule carries chrome, surfaces, shapes and rhythm — not just a palette — so a model that is available to you can produce a deck in the visual language of one that is not.
Capsules stopped fighting your colours
Early capsules pinned their own palette, so choosing a theme or an accent afterwards did nothing. They are now structure-only, and your colour choices stay live.
Measured, not asserted
Twelve models were probed and their traits measured rather than described. Only two of the twelve build a fixed-geometry deck when left to themselves, which is most of the argument for having a format at all.
24 July 2026
The local boundary, made visible
A permanent mark in the header that says where your work is happening, and a page that lets you prove it.
Local processing, always shown
A quiet mark in the header states that processing is local. It speaks at the moments that matter — a file opens, a conversion runs, an artifact is created — and never shows an invented counter.
A page for checking rather than believing
The device page explains the boundary and hands you the check: turn off the connection and watch the tools keep working.
Honest about the asymmetry
While online, the app itself and anonymous page counts use the connection. Your document content never does. Saying so plainly is more useful than a claim that would not survive somebody opening the network tab.
15 July 2026
The model scoreboard gets its first real data
Eight models, the same briefs, scored on the same rubric.
Edition one
Eight models run against fixed briefs and scored on identical dimensions, so “which AI should I use for this” has an answer that is not a vibe. The conformance dimension is scored automatically from the deck itself.
The prompt every model receives is byte-identical
A scoreboard is only meaningful if every model got exactly the same instructions. That is now a rule the harness enforces rather than a convention it follows.
10 July 2026
The pattern library, and PowerPoint coming back in
A browsable catalogue of every layout the format knows, and an importer for decks that already exist.
A public pattern catalogue
Search and filter every layout across all deck families, each with a live example, an honest note on how well it converts, and the prompt fragment that asks for it.
Patterns you pick become the prompt
Choose layouts in the catalogue and the create screen compiles exactly those into your prompt.
Elements have one source
A layout used in four templates used to be four copies that drifted apart. It is now defined once, with a build-time check that fails when a copy reappears.
PowerPoint → unPaper
An importer that takes an existing .pptx and rebuilds it as unPaper HTML, including charts as data rather than as pictures.
8 July 2026
Change one slide without regenerating the deck
The revision loop, a validator, and a home for AI replies that arrive as code.
Comment, revise, splice
Open a deck, comment on the slides you want changed, and get a short prompt covering only those slides. The replacement is spliced back into your file locally, so you are not paying for a whole deck to fix one number.
A route for code-in-the-chat
Assistants that cannot attach files paste the deck into the conversation instead. There is now a dedicated place to paste that, preview it, and save or convert it.
Header and footer survive conversion
Deck chrome is filled in by a small script, and the converter runs with scripts off — so headers, footers and page numbers were disappearing during conversion. The converter now fills them itself.
Analytics that cannot see your work
Aggregate, cookie-free page counts, with the boundary written down: measure reach, never the work.
6 July 2026
The emptiness is the design
A full redesign built on one idea: unPaper earns trust by what it does not do, and the interface should say that by what it does not show.
One question per screen
Flat surfaces, one accent colour, one action per screen, and hard word budgets. A screen with one question and two choices is the privacy claim, rendered visually.
Light by default
Paper-white with ink text is the brand — the product is called unPaper and the aesthetic is paper. Dark mode is equally flat and equally supported.
Trust stated, never sold
No badge walls, no shield rows, no exclamation marks. The strongest trust statement this product has is its own stillness.
4 July 2026
The format goes public
The prompt, the validator, the gallery and the machine-readable contract all get permanent addresses.
A permanent link for the prompt
The canonical slide prompt has its own page, so it can be shared, bookmarked and pasted without going through the whole create flow.
A public validator
Check any deck — including one unPaper did not make — for intact styles and page rules, self-containment, notes and truncation.
A contract for agents
Machine-readable descriptions of the format at llms.txt and llms-full.txt, so an AI agent can read what unPaper HTML is without a person in the loop.
A gallery of real decks
Twelve finished decks, one file each, built by different models — the format demonstrated rather than described.