Skip to content
APDD
See my library

Method

AI prototyping

Using AI tools that write code from instructions to build, quickly, a prototype that works and can be tested with people.

In Spanish
Prototipado con IA
Origin
Karpathy called describing what you want and letting AI write the code “vibe coding” (2025); first studies with UX professionals: Li et al. (2025)
Output
Working prototypes, generated from instructions

What it’s for

Before, a working prototype needed someone who could code, and many ideas were only tested as images. With AI, designers can test how an idea behaves, not just how it looks.

Key elements

  • Test the behavior

    The prototype really responds; you can tap, type and make mistakes.

  • Start from the end

    AI delivers something very finished from the start: designing means reviewing it, cutting and fixing.

  • Use and discard

    The code is for learning, not production; it often has errors and is hard to maintain.

  • The judgment is yours

    AI leans toward conventional solutions; the designer makes the design decisions.

An example

A filter to search a table with lots of data.

A mockup?
A static screen with the filter drawn in; the person has to imagine what would happen. They comment on the look.
A generated prototype
You describe the filter, AI builds it and the person uses it with real data. They test how it works.

What is new is not that it looks better, but that designers can test behavior without waiting for development.

How it’s done

  1. Describe the idea in words, examples or a sketch.

  2. Review what is generated, cut what is extra and fix what AI took for granted.

  3. Test it with people and discard the code once it has answered the question.

Applications in design

  • When to use it

    When you want to test an interaction with people, a mockup is not enough and there is no development time.

  • Avoid

    Shipping the generated code to production without anyone reviewing it.

  • Don’t fall for it

    That something works and looks finished does not mean it solves the right problem.

Authors and works

  1. 2025

    Andrej Karpathy

    Called it “vibe coding” to program by describing what you want and accepting the code AI generates without reviewing it line by line.

  2. 2025

    Jie Li et al.

    Interviewed 20 UX professionals: they prototype faster and depend less on developers, but the code is fragile and hard to maintain.

  3. 2025

    Tianyi Li, Tanay Maheshwari and Alex Voelker

    In a real case, generated prototypes that could be used drew more detailed feedback than sketches, though they pushed toward conventional designs.

Sources

Original sources

Last reviewed: