In a product design workflow, generative AI is good for five things that have proven themselves in production: synthesizing user research, generating flow and content variants, prototyping in realistic code, documenting and checking a design system, and producing visuals on a tight budget. It replaces neither understanding the problem, nor making trade-offs, nor owning the outcome.
What I tested, and on what
Since 2023, I have used generative models on my own products (WeeFizz, Reformer Society) and on client work (Climeet, websites and landing pages). This field report separates what lasted from what ended up abandoned.
The five uses that hold up
1. User research synthesis
Interview transcripts, support tickets, customer reviews: the model groups, counts and quotes. A round of ten interviews gets synthesized in a day instead of a week. The condition: a prompt that forbids inventing quotes and asks the model to flag what the data does not support. And a full review, because the model smooths over contradictions.
2. Flow and content variants
Three alternative structures for an onboarding, five wordings for an error message, two ways to split a form. AI excels at producing alternatives quickly. The designer chooses, and justifies the choice. The gain is not the speed of producing the final choice, it is the breadth of exploration for the same budget.
3. Prototyping in code
This is the use that changed my work the most. A coding agent produces a realistic front-end prototype in a few hours: data, states, responsive layouts, interactions. I finish it by hand. The prototype gets tested with real users and handed to developers as a reference. The line between mockup and product has moved. See front-end development.
4. Design system: documentation and checks
Writing the documentation for forty components from the actual components, spotting naming inconsistencies between Figma and code, generating a component's missing states while respecting the tokens: repetitive tasks the model does well, under supervision. Two days instead of two weeks for the initial documentation.
5. Visual production on a budget
Campaign visuals, illustrations, format variations, demo visuals when there is no photo budget. On WeeFizz, twelve demo sites got their visuals through generation, with a locked prompt system that guarantees consistency. The condition: precise art direction upstream, and strict selection downstream.
The three uses that disappointed
- Generating final screens as images: pretty, inconsistent, impossible to build. Useful for a moodboard, not for a product.
- Replacing user testing with synthetic personas: the model confirms what you told it. It does not know your users.
- Automating prioritization decisions: the model produces a plausible matrix, but it knows nothing about internal politics, technical debt or the client threatening to leave.
The guardrails
- Precise context: design system, tokens, references, no-gos. Without context, the model produces the average of the web.
- Explicit quality criteria in every prompt, and a self-check requested from the model.
- Systematic human review of everything that comes out, especially what looks right.
- A usage policy: what you hand to AI and what you don't (personal data, decisions, commitments).
- Versioned prompts, treated like components: see Prompt engineering for designers.
What it changes for the profession
Production time goes down; decision time does not. The designer who thrives is the one who can frame, choose and stand behind the result. The one who only produced is now competing with a prompt. The rising skill is not knowing how to use AI: it is knowing what you want precisely enough to tell it. That is why I describe prompt engineering as a design practice, and why I offer it as an area of expertise.
One last principle
The machine amplifies intent. It never replaces it.
If the intent is clear, AI multiplies it. If it is vague, AI produces vagueness with confidence. The designer's work starts and ends there: knowing what you want, and why.

