What 11,624 Real Figma AI Prompts Reveal About How Designers Actually Prompt
We analysed 11,624 real prompts and 24,362 generated designs from five months of Fluvara usage. Here is what designers actually ask AI to build in Figma — and.
Between 9 March and 9 August 2026, designers using Fluvara inside Figma wrote 11,624 prompts that produced 24,362 generated designs across 2,576 sessions. Because every design is stored as HTML and CSS alongside the prompt that created it, we can answer a question the industry mostly guesses at: what do designers actually ask an AI to build?
Dashboards dominate — by a wide margin
Of 9,133 prompts specific enough to classify, 15.7% asked for a dashboard, analytics view or admin panel. That is more than e-commerce (7.4%) and landing pages (6.6%) combined. The reason is mundane and important: dashboards are dense, repetitive and slow to lay out by hand. A sidebar, a header, eight stat cards, two charts and a table is 40 minutes of Auto Layout work and almost no creative decision-making. That is precisely the work people want to delegate.
The category breakdown
| What designers asked for | Share of classified prompts |
|---|---|
| Dashboards, analytics, admin panels | 15.7% |
| Mobile app screens (iOS / Android) | 12.4% |
| E-commerce (shop, cart, checkout, product) | 7.4% |
| Landing pages and hero sections | 6.6% |
| Explicit iteration on a previous result | 6.3% |
| Responsive / multi-breakpoint requests | 5.8% |
| Dark mode or dark theme | 4.2% |
| Multi-screen flows (3+ screens in one prompt) | 2.1% |
Mobile is the second workload, and it behaves differently
12.4% of prompts asked for mobile app screens. These prompts are markedly more specific than desktop ones — they name the screens they want ("home, details and checkout"), because designers think in flows on mobile. The prompts that produced the best results in our corpus followed the same shape: platform + app type + the exact screens.
Only 2.1% asked for multiple screens — and that is a mistake
This was the most actionable finding. Just 190 prompts explicitly requested three or more screens, yet those prompts produced the designs users kept and iterated on. Asking for one screen gets you a picture; asking for a flow gets you something you can actually build against. If you take one thing from this data, name your screens.
Specific prompts beat descriptive ones
Comparing prompts that produced kept-and-edited designs against ones abandoned immediately, the difference was never prompt length — it was the presence of concrete nouns. "Modern, beautiful, clean SaaS app" produces generic output. "Pharmacy management dashboard with inventory table, expiry alerts and a prescription queue" produces something usable, because every noun becomes a component the model must place.
A prompt template that works
[Platform] [product type] for [audience].
Screens: [screen 1], [screen 2], [screen 3].
Must include: [specific components].
Style: [theme], [colour direction].6.3% iterate — and they are the users who stay
577 prompts were edits to an existing result ("make the sidebar dark", "add a filter row"). Iteration is where AI design stops being a novelty: the first generation is a starting point, not a deliverable. Users who never sent a second prompt were far more likely to leave without generating anything they kept.
Method
All figures come from Fluvara's production database over 2026-03-09 to 2026-08-09: 11,624 user prompts, 24,362 stored HTML/CSS design artefacts, 1,526 distinct users, 2,576 conversations. Prompts were classified by keyword pattern; 9,133 of 11,624 matched at least one category, and percentages are of that classified subset. Prompts shown publicly are anonymised, and every design in this post links to its live render.