Most product teams don't have a dedicated designer. A developer opens Figma, stares at an empty frame, and starts dragging rectangles — then spends three hours naming layers nobody will read. An LLM won't draw the interface, but it will do the part that actually burns the time: structuring the flow, naming components, writing the interface copy, and catching the accessibility gaps before handoff.
The prompts below are the ones I keep in a pinned note. Each one names the task, shows a working request, and says what comes back. They work in any chat model with a long context window — the constraint is not the model, it's how precisely you describe the frame you're sitting in.
Prompt 1: From a One-Line Idea to a User Flow
Task: UX planning, before Figma.
Give the model the job, the user, and the success condition. Ask for a flow as a numbered list of screens with the decision points marked.
You are a UX lead. Product: subscription coffee delivery, web app.
User: office manager ordering for 20 people, no time, orders monthly.
Success: order placed in under 3 minutes on repeat purchase.
Output: numbered screen list, each with the single decision the user makes there.
Mark the screen where users most often drop off based on the flow logic.
The value is the drop-off call. The model reasons about where friction concentrates — usually the address and the payment step — and you get a hypothesis to test before you draw anything.
Prompt 2: Wireframe in Words, Then in Figma
Task: Figma, first frame.
Describe the frame in the model's language, not yours: viewport, grid, sections top to bottom.
Viewport 1440x900. 12-column grid, 24px gutters, 80px side margins.
Give me a vertical stack of sections for a pricing page:
hero, plan comparison, FAQ, footer. For each section: purpose in one line,
the exact text hierarchy (H1/H2/body/caption), and what sits in which columns.
You get a layout spec you can build in minutes with auto layout. This is the trade-off: you save the blank-canvas paralysis and pay with a slightly generic structure you must adjust to the brand.
Prompt 3: Component Naming and Variants
Task: Figma component library.
I have a Button used in 4 places: primary CTA, secondary action,
ghost in tables, destructive in modals.
Propose a Figma component structure: variant properties, their values,
and the naming convention. Flag which variants can be merged.
Expect a table of properties like Type, Size, State, and a warning that ghost and secondary often collapse into one variant with an icon toggle. Merging variants early is the cheapest design-system win available to a team without a designer.
Prompt 4: Design Tokens as JSON
Task: design system, handoff to code.
From these brand colors [#1A1A1A, #F5F5F5, #2D6CDF] generate a token set:
color, spacing (4px base), radius, typography scale.
Output as JSON in the W3C design tokens format, with semantic aliases
(bg.default, text.muted, accent.primary).
The W3C Design Tokens Community Group format is real and documented; the model will produce $value/$type objects you can feed into Style Dictionary or a Figma variables import. The catch: contrast is not checked — that's prompt 6.
Prompt 5: Accessibility Audit of a Screen Description
Task: accessibility review.
Here is my screen as text: [paste the section stack from prompt 2].
Check against WCAG 2.2 AA: contrast risks, focus order, target size,
form labels, error identification. For each issue: the criterion number,
what breaks, and the minimal fix.
WCAG 2.2 Success Criterion 2.5.8 sets a 24×24 CSS pixel minimum target size — a real, citable number the model should surface for icon buttons. Treat its output as a checklist, not a verdict: it cannot see your actual colors.
Prompt 6: Contrast Math You Can Verify
Task: accessibility, exact numbers.
Compute the WCAG contrast ratio between #777777 text and #FFFFFF background.
Show the relative luminance math for both, then the ratio.
State pass/fail for normal text (4.5:1) and large text (3:1).
This is arithmetic, not opinion — you can check it by hand. #777 on white lands around 4.48:1, just under the 4.5 threshold, which is exactly the kind of near-miss that ships to production unnoticed.
Prompt 7: UX Copy for Every State
Task: UX copywriting.
Write microcopy for a password reset form.
States: default, loading, success, error (wrong email), error (rate limited).
Tone: calm, no exclamation marks, max 60 characters per button label.
Give the placeholder, helper text, button, and error for each state.
The rate-limited state is the one teams forget. Naming all five states in the prompt forces the model to write copy you'd otherwise ship as a raw error string.
Prompt 8: Empty States and Edge Cases
Task: UX review.
For a dashboard with 0 items, 1 item, 500 items, and a failed fetch,
describe what each state shows: illustration, headline, body, primary action.
Empty states are where interfaces feel unfinished. A model that enumerates them for you closes that gap in one pass.
Prompt 9: Heuristic Review of a Flow
Task: design review.
Review this checkout flow against Nielsen's 10 usability heuristics.
Flow: cart → address → payment → confirm.
For each violated heuristic: the step, the problem, the fix, severity (1-3).
Nielsen's heuristics are a published, named framework — safe to reference. Ask for severity ranking so you fix the payment step before the footer link.
Prompt 10: Handoff Spec
Task: handoff to development.
Turn this component list into a handoff doc:
for each component — dimensions, states, spacing tokens used,
interaction notes, and the edge case the developer will hit first.
The last column is the useful one. "What breaks first" turns a spec into a conversation.
Prompt 11: Design System Documentation
Task: design system, onboarding.
Write a one-page README for this component library:
when to use each component, when NOT to use it, and one anti-pattern per component.
The "when not to use" section is what separates a real system from a folder of frames.
Prompt 12: Naming and Layer Hygiene Pass
Task: Figma cleanup.
Here are my layer names: [paste 30 names].
Normalize them to a convention: component/type/variant.
Group them, and flag duplicates that should be one component.
Unglamorous, but it's the difference between a file a new teammate can navigate and one they abandon.
What Actually Changes
The pattern across all twelve: the model is strongest where the work is structured text — flows, names, tokens, copy, checklists — and weakest where it needs to see pixels. Point it at the text layer of design and you get a design partner. Point it at "make it pretty" and you get filler.
Start with prompt 1 and 6 today: one plans the flow, the other catches a contrast failure you'd ship blind. If you're building the rest of the stack solo, the AI Solopreneur covers where this approach holds and where it cracks.