I love building prototypes with AI.
I can have an idea in the morning and use some version of it the same day.
Not click through it.
Use it.
There can be authentication.
Real persistence.
Search.
Keyboard shortcuts.
Settings.
A database.
Maybe even billing.
This still feels slightly absurd to me.
It has also created a new problem.
The prototype can feel complete long before the product is understood.
Implementation depth looks like product depth
I built a small tool recently and reached the point where it had a sidebar, authentication, populated data, settings, command-K, a decent dark mode, and several interactions that worked exactly as I imagined them.
Looking at it, I had the very strong feeling that the product was mostly there.
It wasn’t.
I had built a convincing answer to questions I had already asked.
The difficult questions had not appeared yet.
Who owns this object?
Can two people edit it?
What happens when they do?
What is recoverable after deletion?
Which preferences belong to the user?
Which belong to the workspace?
What happens when a background task fails?
Which changes deserve notification?
What happens when there are ten thousand objects instead of twelve?
There was nothing wrong with the prototype.
It simply had no reason to answer any of those questions yet.
And because everything around them looked so finished, their absence became harder to see.
Screenshots used to lie mostly about visual completeness
We know how to distrust a mockup.
The data is fake.
The interaction is implied.
The edge cases are missing.
It is a picture of an idea.
A working prototype feels different.
I can click it.
I can refresh the browser.
The data is still there.
The route exists.
I can sign out and back in.
That creates a much stronger sense of reality.
But functional completeness and product completeness are not the same thing.
A prototype can have a database and still be a sketch.
It can deploy and still be a question.
That distinction has become much more useful to me.
I try to break the illusion earlier now
Once something starts feeling finished, I deliberately stop polishing it.
I change the data.
I remove all the data.
I use it with the wrong permission.
I make the request fail.
I give it an object with an absurdly long name.
I ask what happens after the action everyone hopes nobody will need.
Delete.
Refund.
Cancel.
Recover.
Transfer.
Leave.
Those questions slow the prototype down.
That is usually good.
They turn a convincing demo into an actual product conversation.
Faster implementation does not mean faster certainty
This is the part I’m still learning.
AI has compressed the time between an idea and something that behaves like software.
It has not compressed the time required to understand what the software should do.
Sometimes it does the opposite.
A polished prototype can make me commit emotionally to an idea before I have earned that confidence.
So I’m trying to keep two separate questions in my head:
Does this work?
And:
Is this right?
The first question is getting dramatically cheaper to answer.
The second one may be becoming the work.
