screenroveEmpty state UX: AI prompts and starter content replace the blank slate
Empty state UX across 109 products: 20 screens swap the Create button for an AI prompt, and 71 ship templates, sample data or runnable starter code.
In 20 of the 342 empty states we captured across 109 products, the first thing a new account sees is not a Create button but a text box asking what you want. That is the clearest shift in empty state UX on our empty states wall: the blank slate is being filled before you arrive, either with an AI prompt or with starter content, and 71 screens (21%) ship templates, samples or example data.
The two moves travel together. 12 of the 20 prompt-box screens also carry starter pills, templates or sample content, against 59 of the other 322 screens (18%). A prompt box on its own is still a blank slate with a cursor in it.
This piece covers lists, libraries, editors and canvases; the dashboard version is in AI dashboard design: the prompt box is the new home screen.
The prompt box takes the Create button's place
Only 3 of the 20 prompt-box screens belong to products we file under AI tools. The rest are monitoring, forms, analytics, video, automation and 3D design products, so this is no longer a pattern for AI-native apps alone.
Checkly is the purest example. An empty account has no monitor list and no New button, only "What do you want to monitor?", a line that says "Describe your site, app, or API in your own words and we'll set up your first checks together", and four pills underneath. The pills are specific: "Checkout with Stripe" names a real flow and a real vendor, which tells a developer what level of detail the box expects.
Descript does the same on a new drive: "What can I help you with?" over a box that reads "Upload a file or describe what you want to make, and I'll help you plan it." Seven task pills, from "Clean up video recording" to "Generate an animated story", plus "Browse more…", turn a blank home into a menu of jobs.
Others keep both paths but put the prompt first. Netlify's empty Projects page has no "no projects" message at all: it opens on "Start building with an AI agent", six ready-made starter prompts from "Marketing site" to "Habit tracker" and only then "Bringing your own code?" with Git providers and a drop zone. SurveyMonkey stacks "Start from an idea", "Start with a template" and "Start from scratch" above a box that says "Describe what you're trying to learn, or paste your survey questions". Zapier's new Zap puts a Copilot box ("Describe your workflow") above an "or" and dashed Trigger and Action steps.
On canvases, the prompt moves into a side panel and the canvas stays clean. Spline's blank scene is a silent grey viewport; the Agent panel says "Let's build!" over five chips such as "A glossy bouncing sphere" and "20 cubes in a grid". Recraft leaves a bare dot grid while its chat panel offers "Start with a prompt — or use a starting point below" and six chips. Jitter opens a design assistant unasked next to a blank scene, with four tasks such as "Clean up my timeline" and a "Something else…" field.
Starter content is the other half of empty state UX
The templates or samples filter matches 71 screens across 50 of the 109 products. Where it shows up is telling: 21 of 63 blank canvases (33%) ship starters, against 43 of 214 first-use lists (20%) and 7 of 31 no-data screens. None of the 19 no-results screens or the 15 all-done screens do. Starters cluster where a user has to make something from nothing, which is also where an AI prompt now lands; see blank canvases with starters.
Runnable samples beat explanations
Neon never shows an empty SQL editor. A new project opens on 11 lines: a comment that explains branches and says "Try generating sample data and querying it by running the example statements below, or click New Query to clear the editor", then a CREATE TABLE playing_with_neon, an insert built on generate_series and a SELECT.
The same idea appears elsewhere in smaller doses. Checkly's empty Heartbeats page puts a ready-to-paste HeartbeatMonitor from checkly/constructs ("Run backup job", period 7 days, grace 2 hours) under the New button. Hygraph sets an outlined "Add demo Model" next to "+ Add Model". Typeform's empty results page offers "Generate test response" next to "Share your form", and ToolJet's blank app offers "Connect to sample data source" and a Sample database.
Example data that previews the payoff
PostHog splits its empty feature pages in two: hedgehog, copy and "Create your first experiment" on the left, and on the right a panel headed "Your results, once running" with an "example data" tag. The fake result is concrete: control at 3.1%, "one-click" at 4.6%, "Chance one-click wins 96%".
ElevenLabs takes a quieter route on Dubbing: five ghost rows ("Product Launch Video" in Spanish, "Customer Testimonial" in French) fade to white under the uploader, so the empty table already has its final shape. AI image tools fill the space with other people's results. Krea shows a wall of captioned examples such as "Two cats added to a living-room photo"; Leonardo asks "LOOKING FOR INSPIRATION?" over three images, each with its prompt and a "Use Prompt" button. Compare Ideogram, whose results pane says "No generations yet" and offers a "Create my first image" button that repeats the Generate button beside it.
Templates as the main action
Shortcut's empty Docs page has a small "Create a Doc" button and then 12 cards under "Start from a template": 11 templates, from "Product Requirements Doc" to "Incident Postmortem", plus "Blank Doc" as the last card. Glide shows six workflow cards under "Start from an existing one".
The empty moment is now a sales surface
Prompts cost money to answer, and it shows. In at least 6 of the 20 prompt-box screens, a credit balance or upgrade button sits in the same view as the prompt. Runway is the starkest: on the Agent page a strip reading "You've run out of credits. Upgrade to keep creating with Runway." is docked onto the top of the prompt card, so the paywall and the input read as one unit.
The others are softer. Spline's blank scene puts "Build with the AI agent" and an Upgrade button directly above its composer. Recraft shows a 60-credit balance and Upgrade above the canvas, Jitter's toolbar carries "Upgrade now" under the assistant, Retool's AI builder has Upgrade next to Publish, and Netlify tags its agent prompt "400 Extra Credits". Outside AI, the empty moment sells too: SurveyMonkey answers "Your survey has no responses" with a green "Collect responses" or a yellow "Buy responses", and VEED marks "Create Brand Kit" with a yellow paid-feature bolt.
What to steal
- Never ship a bare prompt. Put 4 to 7 starter pills under it, written as real tasks in your users' words. Checkly's "Checkout with Stripe" teaches the expected level of detail better than any helper text.
- Keep the manual path in view. Zapier and Netlify put a divider ("or", "Bringing your own code?") between the prompt and the old way in; SurveyMonkey keeps template and scratch buttons right above its prompt.
- Make samples runnable and disposable. Neon's comment tells you how to clear the sample; Hygraph and Typeform keep demo content behind its own button.
- Label example data, and show the payoff. PostHog's "Your results, once running" with an "example data" tag shows the goal without pretending it is theirs.
- On canvases, keep the canvas clean. Spline and Recraft put starters in the panel and leave the work area empty, so the first object a user sees is one they made.
- Decide on purpose what the first prompt costs. If a new account cannot run its first request, say so before they type, as Runway does, rather than after.
How we counted
Templates and samples use the wall's own facet, set when an empty state offers templates, examples, sample data or starter items: 71 of 342 annotated screens (18 unannotated screens are left out). For prompt boxes there is no facet, so we opened every candidate and counted a screen when a free-text field asking for a natural-language request to an AI sits inside the empty state as a way to start. That gives 20 screens from 19 products. We excluded screens that offer AI only as a button, chip or slash-command hint, and image generators whose prompt bar is the tool itself (Krea, Leonardo, Ideogram), treated here as example walls.
All screens were captured in desktop browsers with research accounts in September and October 2026. A vision model recorded each screen's attributes; they are checkable on each screen's page. We record what we saw, not A/B test results, so none of this says which approach converts better; it says where products have moved.
To pull these references into your own build, connect the Screenrove MCP server and ask your agent for empty states with starter templates, or for prompt-first empty states on blank canvases.