From Vibe Coding in Design
Vibe Coding as a Design Practice
Who designs when AI writes the code?
Foundation / 4 min
After this reading, you can tell vibe coding from other ways of making with AI and keep the design decisions as your own.
In February 2025, the computer scientist Andrej Karpathy described a new way of making software: describe what you want to an AI, run what it builds, and ask for changes, without reading the code at all. He called it vibe coding. Within months the term was everywhere, and tools that turn a sentence into a working interface became ordinary.
For designers, this is a real change. An idea that once stayed a sketch until an engineer had time can now be clicked, tested and shown in an afternoon. But speed is not the same as design. A tool that will build anything you describe will also fill every gap you leave with its own defaults, and the result can look finished while answering none of the questions a designer is responsible for.
Three ways of making with AI
The words are used loosely, so it helps to separate three practices. They differ in who decides and how closely the result is examined.
| Practice | What happens | Where the judgment sits |
|---|---|---|
| Vibe coding | You describe the result, run it and ask for changes; the code is not read | In your description and in what you accept |
| AI-assisted design | AI is a tool inside a design process: exploring layouts, drafting copy, sketching variations | In the design process around it |
| AI-assisted engineering | AI writes the code, and people review its structure, correctness and security | In the review |
Nielsen Norman Group draws the line by asking who specified the result. When you describe what you want, the AI is responsible for building it well, but the design decisions belong to whoever made them. The more you leave unsaid, the more of those decisions the AI makes for you.
What it is good for
Vibe coding is at its best when the thing being made is meant to answer a question rather than to last. Martin Fowler describes it as suited to disposable software and quick prototypes, used by people who understand the risks.
| Good use | Why it works | Poor use |
|---|---|---|
| Testing an interaction idea | A flow you can click teaches more than one you can only imagine | Shipping that flow to real users unreviewed |
| Comparing directions | Several versions can be made and set side by side quickly | Choosing the first version because it already exists |
| Showing an idea to others | People respond to something they can use | Presenting generated work as entirely your own |
| Making a small tool for yourself | Only you depend on it, and you accept its flaws | Handling other people’s data or money |
Where authorship lives
A designed thing is the result of decisions, not of effort. When AI does the making, authorship moves to the decisions you make before, during and after it. These are the ones that should never be left to the defaults.
| Decision | The question to answer | What happens if you skip it |
|---|---|---|
| Purpose | Who is this for, and what must they be able to do? | A screen that looks plausible for anyone and suits no one |
| Content | What does it actually say and show? | Placeholder text and invented numbers |
| Structure | What matters most, and in what order? | A familiar template in place of a hierarchy |
| Behaviour | What happens when things go wrong, or when there is nothing to show? | Only the happy path works |
| Visual system | Which type, colour and spacing, and why? | The same look as everything else the tool has made |
| Judgment | Is it good enough, honest and usable by everyone? | Work accepted because it was fast |
A note for design aspirants
Entrance examinations do not test this. UCEED and CEED Part B, the NID DAT Mains and the NIFT Situation Test are answered by hand, in a hall or a studio, because they measure how you observe, draw, make and think. Time spent prompting is not a substitute for time spent drawing.
Portfolios are different, and so are schools’ expectations. If you use AI in work you show, say what it did and what you did, and read each school’s rules on original work. A portfolio is judged on your thinking; a project where the tool made the decisions shows very little of it.
Make with AI, keep the decisions
- Write down who the work is for and the one thing it must help them do, before you open a tool.
- Gather the real content—text, data, images—or mark clearly what is a placeholder.
- Describe structure, behaviour and constraints, not only appearance. Prompting Interfaces into Shape covers how.
- Generate more than one direction, and choose between them for reasons you can state.
- Review the result as you would anyone else’s work. De-slopping AI-Made Interfaces covers what to look for.
- Record what the AI did and what you decided, so the work can be explained and continued.
Vibe coding lowers the cost of making. It does not lower the standard for what is made. The designer’s part is the same as it has always been: to know what the work is for, and to be able to say why it is the way it is.
Takeaway