Fidelity has more than one dial.
A prototype can look like a napkin and still support a complete interaction. Choose the visual, interaction and content detail your question needs.
An argument you can actually click
Wireframes?
Wire not.
Keep the thinking.
Lose the unnecessary detour.
What if your first sketch could already work? Same questions. Same rough edges. A shorter path from “what if?” to “let’s find out.”
Low polish. High function. Zero pixel worship.
rough is a choice.
broken isn’t.pretty unfinished.
fully clickable.
broken isn’t.
EXHIBIT ALooks like
Looks like
a sketch.
That’s a real button, by the way.
Built to be questioned.
fully clickable.
Less ceremony.More things to try.Still plenty of thinking.
01Don’t skip the thinking.
The proposition
Don’t skip the thinking.
Question the extra artifact.
Research, information architecture, user goals and testing are not the problem. Treating a separate wireframe as a compulsory stop on every journey? That deserves a question mark.
AI-assisted tools can turn a deliberately rough idea into something people can use. The point is not to make it look finished. The point is to find out what happens.
Make every artifact earn its keep.
Research. Information architecture. User goals. Testing. All still matter. A separate deliverable earns a place by helping someone think, learn or decide.
Keep the alternatives alive.
One working idea is still one idea. Explore different structures before polishing a favourite. Three colour schemes are not three ideas.
Speed is not evidence. Neither is effort.
Fast work can answer the wrong question. So can expensive, painstaking work. Judge the question, the method and what you actually learned.
02One task. Three ways through.
Don’t take our word for it
One task. Three ways through.
All of them work.
Find the material for Wednesday’s prototyping discussion. Change the structure. Change the finish. Notice what actually helps.
the learning room / a made-up exampleWorking
learning roomPROTOTYPING WEEK
Illustrative content · authored variantsNo live AI in this demo
A demonstration, not an experiment.
Your run does not establish which design is best. Familiarity, order and exploration affect it. Real A/B claims need an appropriate hypothesis, allocation, sample and analysis. Same standard for every workflow.
03Yes, but…
A fair fight
Yes, but…
Good. Let’s hear it.
This argument should survive its strongest objections, not hide them in the footer.
No straw people.They’re terrible at UX.
01But drawing is part of thinking.
Absolutely. Sketching can expose structure, invite discussion and leave possibilities open. An executable sketch adds another way to think, especially when the uncertainty concerns behaviour.
When the objection wins
A shared sketch answers the question more clearly or cheaply than building an interaction.
02Working screens make unfinished decisions look settled.
They can. Keep the appearance rough, show alternatives and make open decisions visible. A working button should not quietly turn an assumption into a requirement.
When the objection wins
People read the prototype as a promise, and its apparent completeness prevents them from questioning the idea.
03AI will just give everyone the same obvious design.
It may offer familiar patterns. Ask for different approaches, make deliberate choices and challenge the defaults. The lab below the manifesto changes the structure, not just the colour.
When the objection wins
Nobody can explain the design decisions or change the behaviour. More generated screens will not fix that.
04This skips research and information architecture.
That would be a failure of the process. A prototype can help investigate a need; it cannot establish that need merely by existing. This applies however the prototype was made.
When the objection wins
The unresolved question is whether the problem matters, and a proposed interface teaches you little about that.
05Functional prototypes can become expensive throwaway work.
Yes. Count setup, iteration, debugging and maintenance, not just generation. Choose the smallest implementation that answers the question. Sometimes that is paper.
When the objection wins
Making the interaction work consumes more than the interaction is likely to teach you.
06We need a reviewable specification and an overview.
Keep them. An annotated map can show relationships that clicking through screens obscures. A versioned prototype can also document behaviour, but it is not automatically a complete specification.
When the objection wins
The separate artifact provides necessary traceability, shared understanding or an agreed delivery boundary that the prototype does not.
07A working demo is not a production service.
Agreed. A demo can test a journey without proving security, accessibility, reliability or operational readiness. Those need their own evidence. This site runs a tiny, bounded example, not a production-readiness test.
When the objection wins
The prototype is being used to claim those questions are solved, or a learning tool has become an unexamined delivery commitment.
08Isn't this still a wireframe, just in code?
Sometimes, yes. We are not policing the noun. The challenge is the compulsory extra stage: drawing a representation, wiring up another representation, then rebuilding it to find out how it behaves.
When the objection wins
Calling it a wireframe helps the team understand its purpose. Keep the name. Lose only the steps that add no value.
04An opinion.
The receipts
An opinion.
With its homework attached.
These sources inform the argument. None proves that this site, or every AI workflow, is better. Follow the evidence further than the headline.
Fidelity is multidimensional.
Visuals, interaction and content can have different fidelity. This informs the dials in our lab; it does not establish that one fidelity is always best.
Open source (new tab): Fidelity is multidimensional.Prototype in the medium.
Code prototypes can support realistic interaction testing. The guidance also distinguishes prototype code from production code. AI did not invent this idea.
Open source (new tab): Prototype in the medium.Explore more than one direction.
A study with 33 participants making web advertisements found benefits from parallel prototyping. Useful support for alternatives, not proof that AI-built websites always win.
Open source (new tab): Explore more than one direction.The critique belongs in the room.
An evaluation of AI-generated designs found contextual and design shortcomings. It examined a particular redesign through heuristic review, not every collaborative AI workflow.
Open source (new tab): The critique belongs in the room.Good experiments still need design.
Hypotheses, metrics, randomization and statistical power matter. Cheap variants do not remove those requirements. Expensive variants do not satisfy them either.
Open source (new tab): Good experiments still need design.This is not a vendor fight.
Figma Make also creates functional prototypes from prompts. The distinction is how we learn, not whose logo is on the tool.
Open source (new tab): This is not a vendor fight.Sources reviewed 22 September 2026. No model-council consensus is presented as independent evidence. No invented conversion lifts. No mandatory tool.
05Make a question
Your turn
Make a question
people can click.
Start with what you need to learn. Take this brief to Lovable, Figma Make, an AI coding partner, or whoever you build with.
A local template composer, not a live AI generator. Nothing you type here is sent to us.
One brief. Many possibilities.
Keep it rough.
Make it testable.
- Same question, different structures
- Real interactions, deliberate detail
- Errors and recovery included
- No invented evidence
Keep what helps. Question what doesn’t.
The sketch is not sacred.
The learning is.
Bring a case where a separate wireframe changed the outcome.
Bring a prototype that taught you something a picture could not.