Generating Designs

Two ways to generate

There are two places to start a generation:

• The floating prompt bar. Toggle it with the magic-wand button in the bottom toolbar, type what you want, and press Enter. Results appear directly on the canvas.
• The AI chat panel. Type your prompt and submit it the same way. Results show up in the conversation: generated images appear as previews you click to place on the canvas, while a generated design or artwork is added to the canvas automatically and gets a result chip — click the chip to scroll the canvas straight to it.

Both run the same generation under the hood, so pick whichever fits your flow.

While it's generating

The prompt field carries a soft iridescent glow, and a beam travels its border while a request is running — so you can tell at a glance whether the AI is still working. On the canvas, whatever is being worked on glows the same way: the layer you're editing, or the placeholder frames standing in for results on the way. The glow fades out once the result lands.

In the chat, the work itself is shown as a checklist that fills in live: picking the model, drafting the design (with a running count of the elements it has laid out), rendering the images it asked for. Each line spins while it's happening and settles with a check when it's done, so a long generation reads as progress rather than a wait. Once the reply lands, the same list is tucked into the collapsible Thought process block above it, together with the assistant's plan — open it any time to see how a result was made.

You don't have to sit and watch. Closing the chat, or switching to another left panel, doesn't stop anything: the generation keeps running and its results still land on the canvas. While a run is in flight behind a closed panel, the AI chat tab shows a small pulsing dot; reopen it and you're back exactly where the work got to.

Generating from a reference

To generate using something already on the canvas as context, select one or more layers first, then write your prompt. Craftwork Make reads the selection's real structure — its layers, text, and layout, not just a picture of it — and works out what you're asking for:

• Edit the selection. "Add a headline here", "добавь сюда текст", "change the title", "replace the photo", "подвинь ниже", "убери бейдж" — the AI adds layers into your design, rewrites its text, swaps its images, moves or removes elements in place, matching the fonts, colors, and spacing already there. One undo reverts the whole change.
• Duplicate it. "Make 5 more like this", "расклонируй и вставь картинки" — the copies are real duplicates of your layers: exact geometry, exact styling, and any photo you didn't ask to change stays the original. Only the content changes.
• Create something new beside it. Anything else uses the selection as a visual reference for a fresh result.

What you ask for decides which happens — in any language. The Style, Ratio, and Count options describe how generated imagery should look; they never turn an edit request into a picture.

In the desktop app with Connect AI, the connected assistant goes further: it works on your document step by step — reading the frames on the page, applying a change, looking at the rendered result, and adjusting — so multi-step requests like "assemble a landing page from these three blocks and fix the spacing" work across several frames at once. Each applied batch is still one undo step.

This is useful for filling in a design ("add the event details to this poster"), variations ("same card but in dark mode"), extensions ("add a matching footer"), or restyling an existing element.

Your references show up as a small stack of thumbnails at the top edge of the prompt. Click the stack to fan the cards out, then drop any you don't want with the × on a card — the deck stays open while you type. With several references attached, your prompt can point to them by number — "combine image 1 and image 2." Removing a reference only lasts for that selection: select the layer again and it comes back.

A result you just generated doesn't become a reference by itself: after its highlight fades, the selection clears, so your next prompt starts fresh. To iterate on a result — "same thing but darker" — click it first so it's attached as the reference.

In the AI chat, every prompt you sent with references keeps a record of them: small badges above the message name the layers it was grounded in. A couple of references show as individual badges; more collapse into one — hover it to see the full list. Click a badge (or a row in the list) and the canvas glides back to that layer. If the layer has been deleted since, a notification says so instead.

Pasting and dropping images

A reference doesn't have to come from the canvas. In the AI chat, paste a screenshot straight from your clipboard (⌘V / Ctrl+V) into the prompt, or drag image files from your desktop anywhere onto the chat panel — each one joins the reference deck like a selected layer would. The AI treats an attached image as the source of truth for what you're asking: paste a screenshot of a design and ask for it to be rebuilt, matched, or extended, and it works from exactly what the image shows instead of inventing its own version.

Attachments belong to the chat you paste them into, so each conversation keeps its own. Remove one with the × on its card; sending the prompt uses them up, and if the request fails they come back along with your prompt text.

When your selection contains text — a written document, a card with copy, a whole page — Craftwork Make reads that text as context, not just a picture of it. So you can select a design and ask for a follow-up that builds on what it actually says, like "turn this research into a 6-card carousel" or "rewrite this page's copy to be punchier."

Using a selection as a template

Select a finished card and ask for a set built on it — "make a 10-card carousel from this template", "расклонируй этот шаблон и вставь картинки" — and the AI duplicates your design instead of redrawing it: every copy is a real clone of your layers with the exact sizes, spacing, fonts, and colors, and only the content changes. Image areas are refilled per card where the request calls for it — and any photo you didn't ask to change stays the template's original. This works the same everywhere — in the prompt bar, in the chat, with or without a connected agent.

A selection that's too complex to reuse exactly says so with a note instead of silently coming back redrawn — select a single card or artboard for the most faithful copies.

When the request calls for real-world content — "the 10 best AI products", "top design tools" — the AI researches the actual items and fills the cards with their real names, prices, and descriptions, and pulls each product's real preview imagery from its page rather than generating a lookalike.

Real icons and Craftwork artwork

Designs built with a connected agent draw their small parts from real libraries instead of improvising them:

• Icons come from the same searchable icon library as the Icons panel — interface glyphs and brand marks land as clean, editable vectors in your design's colors, not as approximated shapes.
• Illustrations and 3D art can come from the Craftwork catalog: when a design calls for a character, illustration, or decorative element, the AI searches Craftwork and places a real asset instead of generating an imitation. Free assets insert as originals; Pro-gated ones fall back to their preview image.

Researching and writing documents

You can ask the AI to research a topic or write something up — "research the @craftwork.design Instagram and outline carousel ideas," "draft a one-page creative brief," "summarize these notes." Instead of a throwaway chat reply, it lays the result out as an editable document on the canvas: a titled page with headings and paragraphs as real text layers you can rearrange and restyle.

Because the document lives on the canvas as text, you can then select it and use it as context for the next step — for example generate the carousel, slides, or page directly from the research you just produced.

Choosing type, model & style

Before you submit, you can set a few options that shape the result.

| Option | What it does | | ------- | ----------------------------------------------------------------------------------------------------------------------------------------- | | Type | Tells the AI what you're making: Graphics, Website, Presentation, or UI mockup. This guides the structure and look of the output. | | Model | Picks which model runs, or leave it on Auto. | | Style | Sets the visual look, e.g. photorealistic, flat illustration, or 3D render. |

Model families

Leave Model on Auto to let Craftwork Make choose the best fit for your prompt and type, or pick a specific one. Models come in a few families:

• Raster image models, which produce a flat image.
• Vector (SVG) models, which produce scalable shapes.
• Editable-layout models, which produce real, editable layers you can rearrange and restyle rather than a single flat image.

Some styles only apply to image models. If you pick a vector or layout model, image-only styles may not take effect.

Aspect ratio & count

• Aspect ratio sets the shape of the result. Choose Original to keep your current proportions, or a preset from landscape to portrait (16:9, 4:3, 1:1, 3:4, 9:16, and others).
• Count is how many results to generate at once, from 1 to 4. More results give you more options to choose from in a single run.

Designing several pages in one style

Ask for more than one page or screen in a single prompt — for example "design 5 pages for a SaaS site" or "a few screens for an onboarding flow in the same style" — and Craftwork Make plans the whole set first: it picks the pages, locks in a shared palette, type, and tone, then designs each page at the same time. A row of placeholder frames appears up front, and each one fills in with its finished, editable page as it's ready, so you can watch the set come together. Every page lands as real layers you can rearrange and restyle.

This works with editable-layout generation (pick a Website, Presentation, or UI mockup Type). For a single design, just describe one page as usual.

Tips

• Name the kind of screen you want — "sign-in screen", "mobile app screen", "dashboard", "pricing page", "product card", "dark mode". Craftwork Make knows how each of those is built (tap-target sizes, real brand icons on social buttons, how a divider or a KPI row should look) and designs to that spec instead of to a generic guess. It understands these in Russian as well as English, and the chat's Thought process lists which ones it applied.
• A small ✦ figure shows the approximate credit cost before you generate. When Model is on Auto, it shows an approximate amount (~) because the model that actually runs may differ. Editing a selected design (adding or changing a few elements in place) costs less than generating a full new design — the charge scales with how much actually changed.
• Press Up and Down in the prompt field to step through your previous prompts and reuse one.
• Want an editable result instead of a flat image? Choose an editable-layout model (or a Type that favors one, like Website or Presentation) so you get real layers back.

For editing an image after you've generated it, see Editing Images with AI.