Marcus Andrey Study
Ler em português Library

AI for UX · Step 2 · 7 min read

Prototyping with AI

How designers use prompt-to-prototype tools in late 2026: what each kind of tool is good at, how to brief one, how to use AI prototypes in usability tests without falling into the fidelity trap, and where a prototype ends and production code begins.

AI tools can turn a description into a clickable app in minutes. That changes what you can test and when: interactions that used to wait for engineering can now be in front of users in the first week. It also brings a new risk, polish that hides wrong decisions. This note explains the tools, how to brief them, how to test with them, and when to stop.

What you'll be able to do after reading:

  • Choose a prompt-to-prototype tool for a given test or exploration
  • Brief it so the prototype tests what you need, not what the tool likes to build
  • Run a usability test with an AI prototype and avoid the "fidelity trap"
  • Explain the difference between a prototype and production code to your team

The tool details are as of November 2026. Prices and limits change often: check each tool's page before you plan around it.


1. What's new about AI prototypes

A traditional prototype fakes behaviour: you link frames, and a person imagines the rest. An AI prototype is real software: forms validate, lists filter, data changes, states appear. You describe it in words, and the tool writes the code behind it.

That makes it strong where static prototypes are weak:

  • Complex interactions with many states (filters, drag and drop, multi-step forms)
  • Realistic data that participants can actually use
  • Quick variations: three versions of a flow for comparison

NN/g recommends exactly this: use AI prototypes to test complex interactions earlier, plan the behaviours you want to test before prompting, and pilot the prototype before real sessions (NN/g, Sep 2026).


2. The tools, by what they're good at

Tool Best at How you pay (Nov 2026) Getting the work out
Figma Make Prototypes that start from your Figma designs and library Figma AI credits (see Figma AI) Publish to a link; "Copy to Design" pastes it back as layers on paid plans
Claude Design (in Claude) Prototypes, decks and one-pagers from a conversation, on your design system Included in every Claude plan, within usage limits PDF, PowerPoint, HTML and more; handoff to Claude Code
v0 (Vercel) Web interfaces in React, close to real front-end code Free tier with daily limits; paid plans with credits GitHub sync and deploy, even on the free tier
Lovable Full small apps with data and login Credits (free: a few per day) Two-way GitHub sync; "you own your code"
Bolt Full small apps in the browser Tokens (free monthly allowance) GitHub integration
Replit Apps with a back end, hosted Paid plans with credits Hosted app, code in the workspace
Google Stitch UI exploration on an infinite canvas, from text, images or voice Google Labs product Clickable prototypes; DESIGN.md import and export
Uizard Fast mock-ups from text or sketches Free tier with a few generations; paid plans Images, PDF, CSS/React handoff
UX Pilot Screens and wireframes from prompts Free start, paid plans Check the site

3. Briefing a prototype

A prototype tool needs the same brief as any AI tool (see Prompt patterns for designers), plus three things:

  1. What you're testing. "Can members find and book a free slot in under 20 seconds?" The tool can't guess the research question.
  2. The behaviours that must work. "Filtering by sport and time must work; payment can be fake." Everything else can be a placeholder.
  3. Realistic data. "Ten centres, five sports, slots from 7:00 to 22:00, some fully booked." Participants look closely at tables and numbers; fake-looking data breaks the session (NN/g).

Give it your design system if the tool supports one (Make, Claude Design and Stitch can import design systems), or at least screenshots and the main tokens. Without them, the prototype will look like the tool's default style, and participants will react to that.


4. Testing with AI prototypes

  1. 1Before

    Write the tasks, build only what they need, pilot it, check every number.

  2. 2During

    Say it is a prototype; note reactions to details you did not design.

  3. 3After

    Decide the structure yourself; polish never proves a pattern right.

Testing with an AI prototype, and avoiding the fidelity trap. Diagram: Marcus Andrey, CC BY 4.0.
A researcher points at a laptop screen showing a sign-up form while a participant types.
A usability test of a working prototype: the participant uses it for real, and the researcher watches where it helps and where it breaks. Photo: Samuel Mann, Wikimedia Commons, CC BY 2.0.

The fidelity trap A polished prototype feels finished, so people (participants and stakeholders) stop questioning it. NN/g warns that polish can hide wrong patterns, and reminds us that "AI helps you realize design ideas faster, but the design decisions remain yours" (NN/g). Decide the structure first; let AI do the building.


5. Prototype or product? Vibe coding vs production

"Vibe coding" means building software by describing it to AI, without reading the code closely. It's excellent for prototypes and personal tools. It's risky for products:

  • Code built this way tends to drift and decay as more prompts pile on; NN/g calls such systems "a house of cards, but… useful" (Vibe Architects, Jun 2026).
  • It may skip what production needs: security, accessibility, performance, error handling, tests.
  • Whatever you build, judge it by how it works for users, and test it with them (GenUI vs. Vibe Coding, NN/g).

See Vibe coding vs production code.

HighReal users and dataLow
Shadow production

Unreviewed code in real use. The boundary has quietly broken.

e.g. a demo given to ten users
Production code

Reviewed, tested and owned by engineering, then released.

e.g. the rebuilt feature, shipped
Prototype

Fast, labelled, mock data only. The right place for vibe coding.

e.g. a code prototype for user tests
Hardening

Engineering reviews and rebuilds what is worth keeping.

e.g. reusing the layout, rebuilding data
LowEngineering reviewHigh
Vibe coding belongs bottom left; the trap is moving straight up to real use without engineering review.

6. Vocabulary

Term Meaning
Prompt-to-prototype Tools that build a working app from a description
Vibe coding Building software by describing it to AI without reviewing the code. See Vibe coding vs production code
Credits / tokens How these tools meter usage; each generation costs some
Handoff / export How the work leaves the tool: a link, layers in Figma, code in GitHub
Fidelity trap When polish makes people accept a design without questioning it

7. Try it (no code needed)

  1. Choose one interaction from your product that's hard to show in static screens (a filter, a multi-step form).
  2. Write the research question and two tasks first.
  3. Brief any approved prompt-to-prototype tool with the question, the behaviours that must work and realistic data.
  4. Pilot it with a colleague for ten minutes. Note what broke, what looked invented, and what you'd fix before a real session.

Reflection questions: Which of your past tests would have been better with a working prototype? What would you never let a vibe-coded prototype do?


8. If you lead a team

  • Agree which tools are approved and what data they may receive.
  • Set the boundary: prototypes are for learning; production goes through engineering.
  • Add "pilot the prototype" and "check the data" to your research checklist.

Checklist

Learn more

Books, courses and videos to go deeper

Design specs for AI · Figma AI · Claude Design · Generating design with Claude Code · Vibe coding vs production code · AI tools landscape for designers

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close