How Ridvay Actually Works (Prompt to Editable Design)

Type this into Ridvay: "Poster for a weekly open mic night, Thursdays 8pm, free entry, bold and modern."

About twenty seconds later you get a poster. Big OPEN MIC headline, the date underneath, a photo panel, the venue name in the corner. That part isn't surprising — plenty of tools produce a picture from a sentence.

Here's the part that is: click the word OPEN and a text cursor appears. Not a selection box around a region of pixels. A cursor, in a text layer, with a font name and a weight and a colour sitting in a panel on the right. Change the font and nothing regenerates. The photo doesn't move. The rest of the poster doesn't quietly redraw itself into something you didn't ask for.

That difference is the whole product, and almost nobody who tries Ridvay for the first time realises it's there. So here's what actually happens between your sentence and your poster.

The model doesn't paint a picture — it writes a document

When you send a prompt, the AI isn't generating an image. It's writing a structured document that describes a design. Internally we call it the IR. Stripped down, one element of it looks about like this:

{ "type": "text", "x": 240, "y": 570, "width": 620,
  "lines": [{ "text": "OPEN", "fontFamily": "Poppins",
              "fontSize": 130, "fontWeight": 800, "color": "#ffffff" }] }

A page has a width, a height, a background, and a list of elements. Each element is a text block, a shape, or an image, and each one knows where it sits and what it looks like. The model's job is to make good decisions about those numbers — that OPEN should be 130px and white and near the top, that the date should be a fifth of that size, that there should be a rule between them.

So the AI is doing design work — hierarchy, spacing, colour, type pairing — and then writing it down. It is not painting.

Everything else follows from that.

Which is why you can edit it

An image model gives you a grid of pixels. If you want the headline bigger, the only lever is to ask again and hope. You've all done this: you get something almost right, you ask for one change, and the next version has a different photo, a different mood, and a new spelling mistake. You lose the version you liked. There's no way back because there was never a structure to change — only a picture to re-roll.

Because Ridvay's output is a document, "make the headline bigger" is one number changing from 130 to 180. Nothing else in the file moves. The photo you liked is the same photo. The colour you picked is the same colour.

In practice that means the useful loop is:

1. Generate once to get a starting layout.
2. Click the thing that's wrong.
3. Change it directly.

Step 3 is the one people skip, and it's the one that turns a mediocre first output into something you'd actually put in a window. If you take one thing from this post: the first generation is a draft, not the deliverable. There's more on that in how to edit an AI-generated design without starting over.

Where the pictures come from

Designs need images, and there are three different sources depending on what the design asks for.

Real photos come first. If a design needs a photographic background — a plate of food, a crowd, a street — a generated image is usually the wrong tool. Generated food looks like generated food. So Ridvay looks for a real stock photo (Pexels) before it reaches for anything else. Free, and it looks like a photograph because it is one.

Icons, logos and shapes come from a catalog. When the design needs a coffee cup, a wifi glyph, a country outline, or a brand logo, it doesn't get drawn from scratch — the model writes a canonical key like icon:coffee or geo:us-texas and the renderer resolves it to a real vector. This matters because AI-drawn icons are where designs go to look homemade. A rectangle labelled "cup" reads as a blue box.

Generated images are the fallback, for when you need something specific that doesn't exist in stock — a particular illustration style, an invented product, a surreal background.

How a document becomes a PNG

At some point you want a file. The renderer reads the document and draws it — text, shapes, images, painted straight onto a bitmap at whatever scale you asked for. No browser, no screenshot. The same drawing code produces a 1080×1350 social post and a print-scale poster; only the numbers change. (A few older cases — diagrams, some exotic motion — still route to the previous browser-based renderer.)

The genuinely hard part is fonts, and for a reason that's the opposite of what you'd guess. A browser is forgiving: hand it a font with no Cyrillic and it quietly substitutes another one, so a Russian headline comes out looking like a slightly different design rather than a bug. Drawing directly forgives nothing — a missing glyph is a tofu box, right there in your poster.

So before a single character is drawn, the renderer checks the font's actual glyph table against the actual characters in your text, and swaps to a face that can render them. Archivo Black is a superb poster font. It carries 423 glyphs. It looks fantastic right up until someone types in Russian, and then it has opinions.

Motion is the same document plus a timeline

A video isn't a separate feature with its own editor. Animating a design attaches timing information to the elements you already have — this headline slides up, that badge fades in 200ms later, these rows stagger.

Rendering it works the way a browser animates a CSS transform: each element is drawn once, and then every frame just moves, fades or scales that already-drawn layer. Frames stream straight into the video encoder as they're produced, so nothing hits the disk in between.

Which means a poster and a 25-second vertical video are the same file with different instructions about time. Change the headline text and both update. The video on our own social accounts is made this way; so is this post's cover image.

The parts people never find

Four things sit one click away and get used by almost nobody:

• Brand kit. Save your palette, fonts and logo once, apply them to any design after. It's the difference between "an AI poster" and "your poster."
• Resize. One design reflows to story, feed, and print. Reflows — the layout recalculates for the new shape rather than stretching. See how to resize one design for every social platform.
• Live image links. Share a design and you get a URL that renders the current version as a PNG. Embed it in a site or an email and it updates when the design does. Bind a countdown to a date and the number on the image goes down by itself.
• The MCP server. If you use Claude, Copilot, or another MCP client, it can compose and export designs directly — no browser involved. Roughly half the designs made through Ridvay in the last month came in this way.

What it's bad at

An honest list, because you'll find these anyway:

• Long copy. Anything past a headline, a subhead and a few lines gets cramped. Posters are not brochures, and the model will try to fit what you give it.
• Exact brand reproduction. It can use your colours and fonts precisely. It cannot redraw your specific logo from a description — upload it.
• Fine typographic control. Kerning pairs, optical alignment, hanging punctuation. If you care about those at that level, you're a designer, and you'll want to export and finish it elsewhere.
• Reading your mind. "Make it pop" produces nothing useful. "Make the headline twice as big and the background a solid dark navy" produces exactly that.

Do it now, in about a minute

Open Studio and paste this:

Poster for a weekly open mic night at a bar, Thursdays 8pm, free entry, huge bold headline, dark background with one accent colour, small venue logo bottom-left.

What comes back is an editable design — text layers, shapes, a photo panel, all separately selectable. Then make these three edits, in this order:

1. Click the headline and drag its size up until it's slightly uncomfortable. Almost every AI first draft is too timid with scale, and scale is what makes a poster readable across a room.
2. Swap the accent colour to your actual brand colour. One click on the colour swatch; the photo and type stay exactly where they are.
3. Delete something. There's usually one line doing no work. A poster with three pieces of information beats one with six.

That's the whole loop, and it's the thing this post exists to convince you of: generate once, then edit — rather than regenerate five times hoping for a better roll.

Open this prompt in Studio →

The one-sentence version

You type a sentence. A model writes a design document. A renderer turns it into a picture, a print file, or a video — and because the document is still there underneath, every single piece of it stays yours to change.

Try it with your own poster → — free to start, and every layer stays editable.