Skip to content

04 · The workspace, step by step ​

The window itself is simple — progress on the left, preview on the right, type selector up top. What's worth understanding is the shape of the pipeline: which steps wait on you, which wait on the agent, and which one you'll pay double for skipping.

One pipeline ​

The site describes it in four beats: choose a starting point, set the visual direction, generate the artifact, deliver or make a video. In practice that's six steps:

StepWho movesWhat you owe it
1 · Pick a typeYouWeb page, prototype, deck, dashboard, image, or video
2 · Pick a design systemYouA direction. Palette, type, and spacing lock here.
3 · Write the promptYouOne clear line about the goal
4 · Answer the formYouThe generated questionnaire — ratio, audience, length, tone
5 · GenerateThe agent—
6 · Fix and exportYouComments, marks, direct edits, then export

Steps 2 and 4 are the ones people skim, and they're the answer to "why does this look so generic".

Artifact types ​

The entry points, roughly (names drift a little between versions, and desktop and web don't match exactly):

EntryOutputTypical use
PrototypeSingle-file HTML interfaceLanding pages, SaaS screens, mobile screens
SlideshowHTML deckTalks, reviews, proposals. The PPT tutorial calls it Deck Mode.
DashboardCharted control surfaceData panels, admin consoles
ImageBitmapPosters, covers, icons, avatars. The tutorial calls it Image Mode.
VideoMotion and videoProduct shorts, data animation, via hyperframes
Design SystemA DESIGN.mdReverse-extract a brand from a screenshot or Figma

Deck Mode and Slideshow are the same thing

The official PPT tutorial writes "Image Mode" and "Deck Mode"; the buttons say "Image" and "Slideshow". One produces a single image, the other an editable HTML deck. Both namings appear in this book because you'll meet both.

Left pane, right pane ​

Once generation starts the window splits:

  • Left is process. What the agent is reading, which file it's writing, how far along. Worth watching the first time — you can see what it understood your prompt to mean.
  • Right is result. A live preview in a sandbox. The artifact is real runnable files, not a screenshot, so what's on the right is the finished form.

Where artifacts land ​

In your project directory. Worth stating once plainly, because it decides everything about how you collaborate afterwards:

  • Git manages it. You can diff it and revert it.
  • Engineering can pick it up directly — no export-and-rebuild step.
  • It isn't in anyone's database.

Three ways to refine ​

After generation there are three ways to change something, each with its own occasion. Step 9 of the deck workshop covers them properly; here's the shape:

ToolWhenExample
CommentThe problem is attached to a specific elementThis logo is too big; this label needs rewording
MarkThe problem is a visual region, easier to show than sayThis group sits too high; this should align with the title
EditYou already know the exact changeOne word, one type size

The rule: don't regenerate what Edit can fix, and don't describe in prose what you can circle. Regenerating a whole page is the most expensive edit — save it for directional problems.

The button people miss ​

Next to the generation controls there's usually Critique — a five-dimension self-review of the artifact (philosophy, visual hierarchy, detail, function, innovation), each scored 0–10, producing an evidence-backed report split into Keep / Fix / Quick wins.

Running it after a first version beats staring at the thing yourself. See Go deeper / Anti-slop.

Next ​

05 · The discovery form — why the questionnaire before generation is the most important step in the pipeline.

An unofficial community guide · compiled from open-design.ai public material and hands-on notes · the official docs win where they differ